Feature Selection Framework Based on Local Data for State of Health Estimation of Lithium-Ion Batteries

Abstract

Lithium-ion batteries (LIB) are essential for electronic devices, electric vehicles, and renewable energy storage, necessitating efficient battery management systems (BMS) for optimal performance and longevity. Accurate state of health (SOH) estimation is crucial for BMS efficiency, while existing machine learning based SOH estimation methods usually focus on global health state features while overlook localized health state features. In this study, we have developed a feature selection framework to obtain a streamlined local feature set within an optimal voltage range, and achieve accurate SOH estimation based on both multivariate and univariate model. We divided datasets into fixed voltage ranges and constructed a comprehensive health feature library for each range. A general feature selection framework has been built to identify optimal voltage range, within which further estimation is performed and results in a minimum RMSE of 0.14%. The physical significance of each feature is analyzed, and the minimum RMSE of the univariate model is 0.5%. The method is validated on two public datasets with different chemical properties and charge-discharge protocols. Using univariate feature from local data for accurate SOH estimation provides new insights into battery health feature selection and enhances the interpretability of the estimation.

Keywords: Machine Learning, Li-ion battery, state of health, Feature selection