Adaptive shapley-embedded neural network ensemble for accurate state of health estimation using electrochemical impedance spectroscopy

State-of-health (SOH) estimation of lithium-ion batteries using electrochemical impedance spectroscopy (EIS) has emerged as a promising approach due to its sensitivity to internal degradation. Recent advancements have incorporated Shapley Additive Explanations (SHAP) to improve interpretability by quantifying the contributions of EIS measurements at different frequencies, thereby facilitating frequency selection. However, two key challenges remain. First, the most informative EIS frequencies vary substantially across temperatures and cells, limiting model generalizability. Secondly, SHAP values are typically employed only for post hoc analysis, lacking a direct mechanism for integration into model decision-making without complex feature engineering. To address these challenges, this study proposes an adaptive and generalizable ensemble framework based on Shapley-embedded single-layer neural networks (SHAP-SLNNs) for accurate and robust SOH estimation under varying temperature conditions. First, SHAP values computed from EIS data are embedded as the initial connection weights between the input and hidden layers of SHAP-SLNNs, effectively infusing domain knowledge directly into the model architecture. An ensemble of SHAP-SLNNs, each trained on different batteries, is then constructed to capture diverse degradation behaviors. Finally, a convex optimization-based adaptive weighting strategy is introduced to dynamically integrate the SHAP-SLNNs, enabling strong generalization across temperatures and battery conditions. Despite being trained using data from individual temperature settings, the proposed framework demonstrates strong generalization capability, consistently achieving high accuracy across a wide range of operating conditions. Moreover, comparative experiments demonstrate that the proposed method achieves superior SOH estimation, with an average root mean square error (RMSE) of 1.04 % and mean absolute error (MAE) of 0.75 % across three temperatures, while effectively balancing computational efficiency and accuracy compared with existing machine learning, deep learning, and transfer learning approaches. To the best of our knowledge, this is the first work to directly embed SHAP values into the architecture of neural networks in the SOH-EIS field, offering a novel and interpretable perspective to improve both accuracy and generalizability.