An explainable artificial intelligence-based feature engineering strategy for lightweight battery health estimation

Accurate and efficient Lithium-ion (Li-ion) battery State-of-Health (SOH) estimation is essential for reliable energy storage management. Current data-driven approaches often prioritize model complexity over systematic and explainable feature engineering, which may limit their robustness and transparency in practical applications. We propose an integrated framework that combines systematic feature engineering with explainability analysis to identify robust feature engineering strategies for lightweight data-driven SOH estimation using multiple machine learning models and SHapley Additive exPlanations (SHAP). Our results and findings show that accurate battery SOH estimation can be achieved using short-time data by adopting explainable feature engineering strategies. Hybrid-feature fusion that integrates raw measurements with derived features and their smoothed counterparts consistently outperforms single-feature groups across both charging and discharging phases. External validation on an independent dataset confirms the robustness and generalization capability of the identified feature combinations, yielding stable and low error SOH estimation across multiple battery cells. SHAP-based analysis reveals clear phase-dependent feature relevance, with voltage-related features dominating during charging, while derived features capturing dynamic behavior are more critical during discharging. Overall, our results and findings indicate that carefully designed and explainable feature engineering strategies are more important than model complexity for short-time SOH estimation, providing a practical and lightweight solution for fast battery health diagnostics.