Feature Selection and State of Health Estimation Based on Segmented Voltage Data of Lithium-Ion Batteries

Reliable estimation of battery state of health (SOH) is crucial for safe and efficient battery management; however, many data-driven methods rely on full charge/discharge cycles and high-dimensional features, leading to high computational cost in practice. A segmented-voltage feature selection and probabilistic regression framework is proposed for accurate yet efficient SOH prediction. Specifically, the voltage profile over a charge/discharge cycle is partitioned into segments, from which a 5 × 24 feature set is extracted for each cycle. Correlation screening, sequential feature selection (SFS), and LASSO are performed to derive compact feature subsets, which are used to train Gaussian process regression (GPR) models. The framework is validated on the public CALCE/UMD aging dataset (CS35) and an in-house high-temperature dataset (EVE INR18650/35 V) collected at 60 °C under three C-rate protocols. On CS35, the SFS-selected subset achieves performance comparable to the full feature set while reducing dimensionality (R 2 ≈ 0.992, RMSE ≈ 0.78%, and MAE ≈ 0.38%), indicating that informative degradation signatures can be captured from localized voltage segments. Experiments on the high-temperature dataset demonstrate robustness under elevated temperature and varying operating rates. Overall, combining segmented voltage features with probabilistic regression provides a compact framework for SOH estimation when reduced dimensionality and uncertainty awareness are desired.