An interpretable adaptive framework for cross-domain state of health prediction in lithium-ion batteries

The data domain shift encountered across diverse operating conditions and disparate material systems constitutes a critical bottleneck, substantially hindering the generalization capability of state-of-health (SOH) prediction models for lithium-ion batteries (LIBs). Furthermore, existing data-driven transfer learning methods exhibit pronounced “black-box” characteristics during the cross-domain adaptation phase, resulting in a profound lack of interpretability regarding their decision-making mechanisms. To circumvent these limitations, this study proposes an efficient and interpretable transfer learning framework tailored for cross-domain scenarios. By incorporating SHapley Additive exPlanations (SHAP) attribution analysis, the framework effectively alleviates the opacity inherent in model transfer. Specifically, the high consistency observed in the feature importance distribution under multi-rate conditions for identical materials demonstrates that the model successfully mitigates the interference of dynamic operations, stably anchoring the common underlying degradation laws. Conversely, across distinct material systems, the significant reconstruction of feature weights intuitively reflects the model capability to perceive and adapt to shifts in dominant aging modes. Ultimately, this research offers a novel perspective on bridging the gap between data-driven “black-box” models and the underlying electrochemical aging mechanisms.