State-of-health estimation of lithium-ion batteries from random local charging fragments: A reliability-aware local-global fusion framework

Accurate battery health estimation from incomplete and irregular charging records remains challenging because local fragments differ not only in length but also in diagnostic quality. This study interprets random local charging fragments as reliability-varying partial health observations and develops RLCF-Net, which fuses local Q V response, masked voltage-location context, and scalar operating conditions. A training-only multi-evidence procedure selects a coverage-stable mother voltage interval, while category-level empirical reliability is defined as the inverse calibration RMSE for each span-region category and used for prediction-level aggregation. The framework is evaluated on two laboratory datasets comprising 121 NCA/NCM cells and a field dataset from 20 commercial EVs. On the main NCA dataset, validation-derived category reliability transferred strongly to held-out cells (Spearman 0.936; Pearson 0.983). Random single-fragment estimation achieved a cycle-level RMSE of 0.0082, whereas cross-fitted selection identified the 0.15 V + 0.20 V Val-K2 protocol with an RMSE of 0.003918. Across all nine K2 protocols, reliability weighting reduced RMSE in every case and remained statistically significant in seven protocols after Holm correction. Entity-level analyses further showed that Val-K2 improved all 14 held-out NCA cells relative to K1 (paired Wilcoxon p = 0.000122), RLCF-Net outperformed the strongest unified baseline in 11 of 14 cells ( p = 0.0419), and NCM reliability weighting reduced physical-cell macro RMSE from 0.002807 to 0.002258 ( p = 0.0322). The independently retrained NCM pipeline achieved a pooled Val-K2 RMSE of 0.002613. In the EV one-event evaluation, reliability-weighted K2 achieved RMSEs of 1.590 and 1.541 pp., with availabilities of 65.15% and 43.49%, respectively. Strict event-disjoint label-independence and common-all-protocol EV audits preserved the directional K2 benefit, although exact inference was limited by the four held-out vehicles. Single-thread CPU latency was 5.105 ms for K1 and 10.058 ms for K2. These results support reliability-aware aggregation for incomplete charging observations while identifying operating-condition-dependent reliability, limited fleet size, label-free adaptation, and automotive-MCU validation as remaining boundaries.