Enhanced state of health estimation for lithium-ion batteries using statistical feature extraction and feature selection

Abstract

State of health (SOH) estimation remains a critical research focus in battery management systems, where feature extraction and selection play crucial roles in improving estimation accuracy. This study examines the effectiveness of statistical feature extraction from constant current-constant voltage (CC-CV) charging curves, focusing on voltage and current characteristics. A total of 50 health indicators (HIs) are derived, including several novel features introduced for the first time. Five fundamental machine learning models-including backpropagation neural networks (BPNN), regression trees (RT), and three types of linear regression (Ridge, Lasso, and Elastic Net)-are trained on these features, with hyperparameter optimization conducted via random search. The best-performing model, RT, is further refined through seven feature selection techniques. Experimental results demonstrate that selecting only the top five features using sequential feature selection (backward) (SFS_backward) and recursive feature elimination (RFE) significantly enhances performance. Compared to using all features, SFS_backward and RFE reduce root mean square error (RMSE) by 12.8 % and 12.5 %, respectively, while mean absolute error (MAE) decreases by 13.1 % and 15.2 %. The proposed methodology also outperforms conventional and deep learning approaches, achieving up to a 127.0 % reduction in RMSE and 113.3 % in MAE. These findings underscore the potential of statistical feature engineering and selection to enhance SOH estimation accuracy while reducing model complexity.