Interpretable machine learning for battery health insights: A LIME and SHAP-based study on EIS-derived features

Abstract

This paper presents a comparative study of interpretable machine learning methods for lithium-ion battery state of health (SOH) estimation using features derived from electrochemical impedance spectroscopy (EIS) and distribution of relaxation times (DRT) analysis. Four DRT peak-area features capturing diffusion (A1), charge-transfer resistance (A2), solid-electrolyte interphase impedance (A3), and ohmic resistance (A4). These serve as inputs to five regression models: linear regression, support vector regression, k-nearest neighbors, random forest, and gradient boosting. All models achieve near-perfect predictive accuracy, demonstrating that these EIS-derived features reliably encode SOH information. To bridge the gap between high performance and transparency, we apply Local Interpretable Model-Agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) to quantify both local and global feature importance. Our interpretability analysis reveals a unanimous consensus: the SEI-related feature (A3) dominates SOH predictions, with the charge-transfer feature (A2) as a secondary contributor, while diffusion (A1) and ohmic (A4) features play lesser roles. Cross-model and cross-method agreement underscores the physical validity of these insights and paves the way for integrating transparent, trustworthy SOH estimators into safety-critical battery management systems.