Study On Lithium Ion Battery State Of Health Estimation Under Small Sample Condition

Lithium ion battery is one of the main energy storage devices.It is widely used in aerospace,intelligent equipment,new energy vehicles and other fields.Lithium ion battery will produce irreversible capacity attenuation in the use process.Its health will gradually deteriorate.If the battery reaches the end of life condition without timely replacement of the battery will produce a great potential safety hazard.It is of great safety significance to monitor the health status of the battery accurately.The chemical mechanism of lithium-ion batteries is very complex.Traditional electrochemical modeling methods involve complex chemical reactions.However,the data-driven modeling method can directly establish the health status estimation model based on the battery life degradation data.Because it doesn’t take into account the electrochemical mechanisms inside the cell.This method has become a hotspot in current research.Data-driven modeling requires sufficient training samples.However,the aging test period of lithium ion batteries is longer.It is difficult to obtain sufficient battery data in a short time.Due to the differences in operating conditions,capacity and electrode materials,the data distribution of different batteries is not the same.Models based on specific battery data do not apply to different battery objects.Therefore,it is of great application value to study how to realize a reliable battery health state estimation model through small sample learning.In order to solve this small sample problem in data driven modeling of lithium ion battery.This paper introduces the solution of knowledge transfer.Firstly,based on the differences in battery operating conditions and battery types,data fields were divided and different small sample data sets were constructed.The sample-based transfer learning method is used to transfer knowledge between two different data domains to enrich domain information in small sample data.In this process,this paper analyzes the data imbalance in machine learning and optimizes the processing strategy of sample weight.Then,aiming at the problem of limited knowledge transfer in a single domain,this paper expands the background of knowledge transfer in a single domain to multiple source domains.The multi-task learning mechanism of meta-learning is used to explore the knowledge transfer under the multi-domain background.Finally,domain measurement was carried out by the maximum mean difference algorithm to improve the model-agnostic Meta Learnings algorithm and improve the efficiency of knowledge transfer in multiple domains.This paper is based on the Python language and Pytorch machine learning framework for simulation experiments.The results show that the generalization performance of traditional machine learning model is poor due to the lack of sufficient domain information under small sample condition.Through knowledge transfer,the established model can reduce the information gap between the overall data and the samples and realize the knowledge fusion among different data fields.In the process of SOH prediction,the model can correctly track the capacity attenuation trend of the battery.The error results show that the prediction scheme based on knowledge transfer can significantly improve the estimation accuracy and generalization performance of the model.At the same time,the improved scheme presented in this paper has smaller estimation error than the original method.