Robust partial-window state-of-health estimation for lithium-ion batteries using chemistry-aware multi-health indicators and missing-aware ensembles

Accurate battery state-of-health (SoH) estimation is a central requirement for reliable battery-management systems, yet practical deployment is constrained by partial charging, chemistry-dependent voltage behaviour, and intermittent availability of health indicators. The central limitation addressed in this work is the dependence of single-HI incremental-capacity-analysis-based SoH estimators on one dominant voltage-window peak, which may be chemistry-dependent or unavailable during partial charging. This paper extends single-peak incremental-capacity SoH estimation to a partial-window multi-health-indicator framework. Incremental-capacity, differential-voltage, time, temperature, and partial-charge-throughput descriptors are extracted from chemistry-specific voltage windows and evaluated under leakage-safe leave-one-cell-out validation. To support partial-window operation, a Top- K Missing-Aware Window-Gated Multi-Model Ensemble (TopK-MAW-GME) is introduced. Across three public degradation datasets, the best fixed-window multi-HI models reduce RMSE from 0.0175 to 0.0104 on Dataset 1, from 0.0085 to 0.0043 on Dataset 2, and from 0.0117 to 0.0023 on Dataset 3 relative to the best controlled single-HI baselines. In BMS-like chronological online replay, TopK-MAW-GME-K5 achieves the lowest RMSE on Datasets 1 and 3, whereas TopK-MAW-GME-K3 gives the best accuracy-latency-memory trade-off on Dataset 2. Structured missing-window and feature-removal stress tests show two complementary behaviours: TopK-MAW-GME-K3/K5 preserve substantially greater measurement-update capability when the selected best voltage window is unavailable, with K5 maintaining full update coverage across all three datasets, while feature-removal cases identify chemistry-dependent critical HI families and expose sensitivity to missing feature groups. These results indicate that multi-HI partial-charging representations improve SoH estimation accuracy, while top- K missing-aware expert fusion provides a practical robustness mechanism for partial-window BMS operation.