Estimation of battery state of health and open circuit voltage at various depths of discharge based on deep learning and relaxation voltage

Data during the relaxation phase of batteries is readily obtainable, requiring no complex feature engineering, and is closely related to battery aging and OCV, which makes it a recent research hotspot. However, most existing methods primarily utilize relaxation voltage to estimate the SOH of batteries only after they have been fully charged, thereby limiting their application scenarios. Unlike methods that focus solely on relaxation voltage at high SOC, this study explores the influence of different depths of discharge on voltage response. First, relaxation voltages at various discharge SOCs are extracted to derive the voltage differences and their corresponding statistical characteristics. Subsequently, a multi-gate expert model that combines CNN and BiLSTM is established to achieve a joint estimation of battery SOH and OCV. Experimental results from a degradation dataset of 12 cells with 3 Ah capacity indicate that introducing the OCV estimation task enhances the accuracy of SOH estimation, with MAE for SOH and OCV reaching 1.51 % and 0.25 %, respectively. Finally, by fine-tuning a pre-trained model and utilizing only a small amount of target domain battery data, transfer learning for SOH and OCV is accomplished, yielding MAE values of 1.59 % and 0.46 %, respectively, which validates the model's strong generalization capability.