State of health estimation for lithium-ion batteries based on optimal feature subset algorithm

Currently, most of data-driven lithium-ion battery capacity estimation studies use feature screening by analyzing the level of linear association between features and capacity, leading to the neglect of the hidden non-linear relationship between features and capacity. To overcome this limitation, this study constructs a feature selection evaluation function COMAN, which comprehensively describes the dependency between features and capacity originating in linear and nonlinear aspects. Based on it, the paper further presents an Optimal Feature Subset Algorithm (OFSA), which selects the optimal subset of features that are significantly associated with capacity and have low redundancy with the features. First, 26 features are extracted grounded in various parameters, including voltage, current, temperature, and time. Next, the proposed OFSA is applied to these features to select the optimal feature combination. Finally, CatBoost model was employed to build the connections between the selected features and capacity, with automated hyperparameter tuning performed using the Optuna library. Comparing with existing machine learning algorithms, the root mean square error (RMSE) and mean absolute error (MAE) of the method do not exceed 0.5 %, and the R2 value exceeds 0.9887. The findings from experiments indicate that the proposed method enhances the robustness of battery's SOH.