Cross-Battery Lithium-Ion Battery SOH Estimation Using Reference-Normalized Local Discharge-Time Features

Cross-battery state-of-health (SOH) estimation requires health indicators that remain informative when target-cell data are excluded from model training. We examine five fixed-voltage local discharge-time windows and three feature representations: raw time, initial-reference offset, and relative change. Four NASA cells (636 discharge records) and four CALCE CS2 cells (4011 records) are evaluated using battery-level protocols. The primary source-only nested leave-one-battery-out (LOBO) NASA experiment achieves a battery-macro RMSE of 0.01613 (between-cell SD 0.00349), with 3.7–3.5 V and 3.8–3.6 V each selected in two outer folds. In a separate retrospective fixed-window analysis at 3.7–3.5 V, relative normalization reduces NASA macro RMSE from 0.02104 to 0.01267, but the improvement on CALCE is marginal (0.02733 to 0.02719); offset subtraction yields 0.02684 on CALCE. Direct NASA-to-CALCE and CALCE-to-NASA cross-dataset transfers favor offset subtraction over relative scaling. These transfers do not involve retraining on target-dataset SOH labels but require an initial discharge-time reference from each target battery. Additional analyses examine reference-cycle exclusion, later references, deterministic reference perturbations, conventional single-feature baselines, and source-only support vector regression (SVR) tuning. These results support reference-assisted use under specific conditions rather than a universal advantage of relative normalization. The limited number of independent cells and dependence on a reliable target-cell reference constrain generalization.