Multi-Source Impedance and Discharge Feature Learning for Cross-Battery State-of-Health Estimation of Lithium-Ion Batteries

Abstract

Accurate state-of-health (SOH) estimation is essential for the reliable operation of lithium-ion batteries. However, predicting SOH for previously unseen batteries remains challenging because degradation behavior varies among cells. This study proposes a multi-source feature-learning framework that combines electrochemical impedance spectroscopy (EIS), discharge-profile, and aging-related information for cross-battery SOH estimation. EIS and discharge data from 34 batteries in the National Aeronautics and Space Administration (NASA) battery aging dataset are processed to construct 1830 matched impedance–SOH samples. A total of 101 features are extracted from raw and rectified impedance spectra, resampled impedance points, NASA-provided impedance parameters, discharge profiles, and cycle-related information. Random Forest (RF), Extra Trees (ET), Gradient Boosting (GB), Histogram Gradient Boosting (HGB), and Extreme Gradient Boosting (XGBoost) models are evaluated using random sample splitting and strict battery-wise validation. Under random validation, GB achieves the best performance, with a coefficient of determination (R2) of 0.9788. Under strict battery-wise validation, ET achieves a mean absolute error (MAE) of 4.9131 percentage points, a root mean square error (RMSE) of 7.3664 percentage points, and an R2 of 0.7855. The performance difference between the two validation strategies demonstrates the importance of battery-grouped evaluation when assessing generalization to unseen cells. Overall, the results indicate that combining impedance- and discharge-derived information provides a promising basis for cross-battery SOH estimation, although further leakage-controlled validation across broader operating conditions is required.