Adversarial learning for robust battery thermal runaway prognostic of electric vehicles

Abstract

The thermal runaway risk in batteries is a critical issue affecting their safety. Although considerable research is dedicated towards the thermal runaway prognostic of batteries, a common challenge remains that how to reduce the false positive rate to enable practical real world application. To address this problem, this paper proposes a robust thermal runaway prognostic method for Li-ion batteries using a generative adversarial network. Specifically, the model utilizes the charging curves of batteries as features during training to reduce the randomness of sparse data and enhance the robustness of the model. A resampling algorithm is proposed to realize the standardized conversion of the charging curve to retain the complete charging information. By employing a one-class generative adversarial network, the method generates a detection model that provides a reference normal curve for the original battery charging process, detecting potential anomalies. Real accident data of electric vehicles that experienced thermal runaway and subsequent fires is used to verify the performance of the model. Experimental results show that the proposed method can identify all abnormal cells with a false positive rate of 1.75% before thermal runaway occurs, reducing 7.54% to 31.18% false positive rate compared to existing methods.