Degradation-informed feature-graph selective state-space network for state-of-health estimation of lithium-ion batteries under fast-charging conditions

Accurate state-of-health (SOH) estimation of lithium-ion batteries under fast-charging conditions is important for safe operation and lifetime management in battery management systems. High-rate charging currents, transitions between charging stages, and nonuniform sampling can make charging-derived health features more sensitive to noise, outliers, and cell-to-cell variability, thereby increasing the difficulty of SOH estimation. This study proposes a Degradation-Informed Feature-Graph Selective State-Space Network (DIFG-SSM) for this task. Cycle-level health features are first extracted from the charging process and subjected to correlation-based screening, redundancy reduction, and robust preprocessing. The retained features are represented as graph nodes, and statistical correlations, feature-category priors, and task-adaptive relationships are combined to form a degradation-informed hybrid health-feature graph. A multi-scale Lite-SSM encoder then models short-term disturbances, stage-wise changes, and long-term fading trends across cycles. The output stage combines trend-residual prediction, bias calibration, and a degradation-consistency objective to improve trajectory stability while retaining degradation-related local variation. Experiments are conducted on one private fast-charging dataset, one private conventional-charging dataset, and one public fast-charging dataset. The evaluation covers recent baselines, supplementary cell-level tests, single-module and joint ablations, trajectory-consistency analysis, and feature-relation interpretation. On the main fast-charging dataset, DIFG-SSM provides lower prediction errors and more stable trajectories than the evaluated baselines. When the conventional-charging dataset and the public multi-stage fast-charging dataset are trained and evaluated separately, low within-dataset cell-level prediction errors are also obtained.