Physics-informed unified network for robust and interpretable SOH estimation of lithium-ion batteries

Accurate and reliable estimation of lithium-ion battery state of health (SOH) is essential for battery management systems (BMSs), directly affecting safety, lifetime prediction, and energy efficiency. Although data-driven models have achieved promising accuracy, their practical deployment is often limited by insufficient physical consistency, weak cross-chemistry generalization, and poor interpretability. In this work, we propose a physics-informed unified network (PIUN) for single-step SOH estimation, which integrates physically consistent constraints into a compact gated backbone termed the Update–Candidate Branch Network (UCBNet). By jointly enforcing data fidelity and degradation-aware physical regularization, PIUN captures both static health indicators and dynamic aging behaviors within a unified framework. The PIUN is evaluated on four public battery datasets (XJTU, TJU, MIT, and HUST). Across 11 evaluation subsets, PIUN reduces MAPE and RMSE by an average of 11.5% and 7.4%, respectively, compared with the strongest baseline. Under noise-augmented training, the average improvement further increases to approximately 16%, demonstrating enhanced robustness to measurement perturbations. Cross-dataset experiments additionally confirm superior zero-shot and few-shot generalization capability. To improve interpretability, SHAP analysis is conducted and reveals strong consistency between learned feature contributions and electrochemical degradation mechanisms. Guided by SHAP-derived importance within interpretable physical dimensions, a compact subset of 7 representative features is constructed. Experimental validation demonstrates performance comparable to the full 17-feature model, thereby experimentally confirming the reliability of the SHAP-based interpretations while enabling lightweight deployment. Overall, PIUN provides a physically consistent, interpretable, and generalizable SOH estimation framework with strong practical potential for real-world BMS applications.