Physics-Informed Hierarchical Multi-Source Domain Generalization for SOH Estimation of Lithium-ion Batteries

Accurate state-of-health (SOH) prediction of lithium-ion batteries under unseen operating conditions remains challenging because of both cross-condition domain discrepancy and battery-level heterogeneity within the same condition. To address this issue, this study proposes a physics-informed hierarchical multi-source domain generalization framework (PI-HMDG) for cross-condition SOH estimation. The framework introduces a condition-aware inter-domain weighting strategy to evaluate the transfer contribution of each source domain based on operating-condition similarity, and a reliability-aware intra-domain aggregation strategy to account for differences in battery degradation behavior and data quality within each source domain. A hierarchical target-domain excess-risk bound is further derived to provide theoretical support for the proposed hierarchical weighting design. Experiments on the NASA, TJU, and XJTU lithium-ion battery datasets show that the proposed method maintains consistent prediction accuracy across operating-condition shifts, low-temperature transfer, and cross-dataset transfer scenarios, while outperforming representative ablation variants and baseline methods. These results suggest that PI-HMDG provides a feasible approach for data-limited cross-condition battery health prediction.