Interaction-Aware Graph Learning for State-of-Health Estimation of Sodium-Ion Batteries

Abstract

Accurate state-of-health (SOH) estimation is critical to the reliable operation and large-scale deployment of sodium-ion batteries (SIBs). However, the complex and nonstationary degradation behavior of SIBs under different charging protocols pose a major challenge to achieving both high estimation accuracy and strong generalization. To address this issue, this study proposes a kernelized interaction graph learning framework (K-Ginter) for SIB SOH estimation. Specifically, 28 degradation-related features are extracted from charge–discharge signals to characterize battery aging from temporal, thermal, energy, voltage/current dynamic, and incremental-capacity perspectives. An interaction-aware feature condensation framework is then developed to identify a compact subset that preserves both the individual feature importance and cooperative interaction structure. On the basis of the selected features, a dynamic interaction graph is constructed using a position-aware variable-scale window and a degradation-aware Chebyshev metric to capture evolving cross-feature dependencies throughout the aging process. A kernelized interaction graph encoder is further developed to preserve interaction topology and learn robust nonlinear degradation representations from the resulting graphs. Experimental results on a self-constructed SIB data set show that K-Ginter consistently outperforms representative baseline methods, achieving root-mean-square errors below 0.8% in within-data set estimation and below 1.0% in generalization evaluation across charging protocols. In addition, Shapley additive explanations (SHAP) analysis reveal interpretable interaction patterns among key degradation features. These findings suggest that interaction-aware graph learning provides an effective and interpretable framework for SIB SOH estimation under investigated experimental conditions.