Multi-grouping-compatible frame-level unsupervised fault diagnosis and localization for electric vehicle battery packs in realistic conditions

Implementing real-world safety monitoring and fault diagnosis for lithium-ion batteries in electric vehicles is crucial. In this work, we design a fault diagnosis method based on an Attention-Gated Recurrent Unit (GRU)-Variational Autoencoder (VAE)-StatFusion neural network. This method not only identifies faults in batteries used in real-world applications but also meets the requirements for frame-level diagnosis, compatibility with different pack groupings, and fault cell localization. We constructed 15-dimensional online features of electrical and thermal characteristics to map battery safety. By combining the probabilistic distribution of the network's latent variables and unsupervised reconstruction loss, we design a comprehensive diagnostic index that can be output in real-time. Additionally, we quantify the contribution of each cell to the fault at abnormal moments, enabling online fault localization. Through fault cases such as electrolyte leakage, connection anomalies, excessive aging, and internal short circuits, the algorithm demonstrates effective fault diagnosis and localization for typical safety issues. Furthermore, we test the algorithm's Receiver Operating Characteristic (ROC) performance within 500 realistic vehicles. Under the same data conditions, compared to previous diagnostic networks, our proposed method showed a 42.1%–58.3% improvement in true positive rate (TPR) within the false positive rate (FPR) range of [0, 0.3]. Overall, this paper achieves a more accurate and practical battery fault diagnosis method under more refined application requirements, promoting the safety of battery applications.