Accident prevention in electric vehicles through battery state-of-health estimation based on GRU-HSIC

The State of Health (SOH) of an Electric Vehicle (EV) battery has a direct impact on both vehicle performance and passenger safety. Accurate SOH estimation is critical for accident prevention, as estimation inaccuracies can enable undetected battery degradation to escalate into catastrophic operational failures. Traditional estimation methods that rely on manually extracted health features struggle with varying data distributions and sequence lengths across cycles, making the feature extraction process time-consuming and labor-intensive. Gated Recurrent Unit (GRU) network alleviates the need for manual feature engineering, but may capture redundant information in sequential data that affects estimation accuracy. To address this issue, we propose a new GRU network based on Hilbert-Schmidt Independence Criterion (GRU-HSIC). First, a zero masking network is used to transform battery data with varying lengths into sequences of the same length while retaining cycle-length information. Then, the HSIC bottleneck is extended to GRU, reducing input redundancy to improve model representation robustness. The proposed model demonstrates the lowest mean absolute error (MAE) of 0.004 and root mean square error (RMSE) of 0.006 on the CALCE CS dataset and the lowest MAE of 0.011 and RMSE of 0.014 on the CALCE CX2 dataset. In addition, experimental results on NASA RW dataset also demonstrate that the proposed method achieves higher accuracy than other recurrent methods.