Comprehensive feature extraction for battery health prognostics: Identifying predictive indicators of state of health

Accurate estimation of battery state of health (SOH) is crucial for ensuring the safety, reliability, and operational efficiency of energy storage systems in electric vehicles, consumer electronics, and grid applications. Traditional approaches often rely on a limited set of handcrafted features derived from electrochemical analyses, such as incremental capacity, differential voltage, and constant-current/constant-voltage (CC-CV) phases, which restrict their predictive power and generalizability. This study introduces a comprehensive machine learning pipeline for large-scale feature engineering and SOH modeling using only standard sensor data: current, voltage, temperature, and time. Using a public dataset, we generate over 40,000 features across seven domain-informed groups that capture both charge and discharge dynamics. Feature relevance is assessed through univariate analyses (Spearman correlation, Predictive Power Score, and single-feature models) and multivariate modeling within a unified selection pipeline. Prediction targets include remaining useful life (RUL) and future discharge capacity at 10, 50, 100, and 250 cycles ahead. In total, we develop 40 final LightGBM (Light Gradient Boosting Machine) models, spanning the complete feature space and individual feature groups. Model optimization employs a hybrid selection strategy combining SHAP (SHapley Additive exPlanations)-based importance ranking, forward feature selection, and recovery techniques using BorutaShap and minimum redundancy maximum relevance (MRMR). Across all models, 773 unique features are retained, forming a compact yet highly informative subset. The best RUL models achieve a mean absolute percentage error of approximately 10 %, while capacity-forecasting errors remain below 1 % across all prediction horizons. Notably, sliding-window descriptors are frequently retained by the multistage selection pipeline and recurrently appear among the top SHAP contributors in the final models, suggesting that short-term temporal aggregation provides complementary information to single-cycle descriptors. These findings demonstrate that broad and systematic feature exploration, integrated with robust univariate-multivariate selection and interpretable modeling, substantially improves SOH prediction accuracy and generalizability. The proposed framework is scalable and adaptable for data-driven SOH estimation, offering a strong basis for advancing battery diagnostics and prognostics.