A hierarchical multi-domain adaptation method for generalizable cross-domain state of health estimation of lithium-ion batteries

Accurate state of health (SOH) estimation of lithium-ion batteries is essential for ensuring safe and reliable operation of battery-powered vehicles. However, cross-domain SOH estimation remains challenging due to significant domain discrepancies and the scarcity of labeled data. To address the challenge, this paper proposes a generalizable cross-domain SOH estimation method. First, a feature-based domain discrepancy quantification method is proposed integrating the construction of physically interpretable health features and maximum mean discrepancy quantification, which evaluates the feasibility for hierarchical multi-domain adaptation. Further, a hierarchical multi-domain adaptation method is proposed by organizing the adaptation process into a structured two-stage paradigm. It performs supervised multi-source alignment to reduce intra-source discrepancies and enrich the feature space, followed by adaptive adversarial adaptation to learn domain-invariant representations across target domains. Specifically, an adaptive adversarial network based on gradient reversal layer and a two-dimensional convolutional neural network is designed to enable hierarchical multi-domain adaptation for generalized SOH estimation. Extensive validation on three datasets comprising 114 batteries demonstrates the method's superior performance, achieving average RMSE and MAE below 0.02 and R² exceeding 90%. These results outperform existing benchmarks and confirm the method's strong robustness and applicability for practical cross-domain SOH estimation.