Changepoint detection as a light data-driven approach to battery state-of-health prediction

Abstract

Accurate prediction of battery state of health (SOH) remains challenging because degradation processes are highly sensitive to cell chemistry, manufacturing variability, and operating conditions, while available field data are often limited. Generalized and data-efficient modeling approaches are therefore required for reliable battery health assessment across different applications. Here, we report a data-driven feature extraction framework based on changepoint detection (CPD) to identify statistically meaningful transitions in battery aging data. The approach is applied to both capacity-check and regular aging cycles of LiNixMnyCozO2|graphite cells. The extracted features are used to train an extreme-gradient-boosting regressor, enabling accurate SOH estimation with root-mean-square errors of 0.013 and 0.023 for capacity-check and aging-cycle datasets, respectively. The features show strong correlation with lithium loss and active-material degradation, demonstrating that CPD provides a physics-aware and computationally efficient pathway for battery health prognosis.