A GradCAM++-Driven State-of-Health Estimation Method for Lithium-Ion Battery Based on Relaxation Voltage With Physical Interpretability

To ensure the safe and reliable application of lithium-ion batteries, accurate estimation of the state of health (SOH) is essential. However, most existing data-driven methods rely on charge/discharge data and lack interpretability due to their "black-box" nature. To address these challenges, this paper pro-poses an interpretable-algorithm-driven SOH estimation method based on relaxation voltage. To enhance the interpretability of the approach, we integrated the GradCAM++ algorithm into the con-volutional neural network (CNN). This integration not only im-proved the model’s estimation accuracy, with an average reduc-tion in root mean square error (RMSE) of 16.34%, but also gener-ated important class activation maps. Subsequently, these maps were correlated with a battery equivalent circuit model for aging analysis. The results reveal how the CNN focuses on either diffu-sion depolarization or activation depolarization across different battery lifecycle stages to perform SOH estimation. For the batter-ies exhibiting different aging trends, the model emphasizes distinct depolarization processes. Comprehensive validation experiments confirm that the proposed method achieves accurate SOH estima-tion with an average RMSE of 1.70% using only the first 6 minutes of relaxation voltage data, while providing insights for guiding SOH estimation strategies in practical applications.