Battery Fault Detection and Early Warning for Electric Vehicles: A Deep Learning-Powered End-to-End Solution
Jinwen Li, Yunhong Che, Yusheng Zheng, Kai Zhang, Xiaosong Hu
Abstract
Accurate battery fault detection and early warning are important to the efficient and safe operation of electric vehicles (EVs). Artificial intelligence-based methods make it possible to monitor faults in a timely and low-cost manner. To this end, this paper introduces a deep learning-powered end-to-end framework for fault detection and early warning of lithium-ion batteries in real-world EVs. The framework leverages an unsupervised architecture that relies solely on cell voltage for fault detection and early warning. It supports real-time detection and simultaneous monitoring of all battery cells. Firstly, the fault detection model is established using a Transformer with two-phase adversarial training, facilitating automatic feature extraction in an unsupervised manner. Subsequently, the threshold is dynamically and adaptively determined based on extreme value theory. Finally, the battery’s safety risk is quantified by weighing voltage and temperature information. Field data from 20 faulty EVs validate the effectiveness of the proposed framework. Results indicate the method is robust to data fidelity, requiring only 4 days of historical fault-free data for model training, achieving accurate fault detection with an F1 score of 0.969 and an FPR of 0.095. Additionally, our model can provide several hours and seconds of early warning for evolutionary and sudden faults, respectively.