Towards general and efficient fault diagnosis: A novel framework for multi-fault cross-domain diagnosis of lithium-ion batteries in real-world scenarios

Compared with laboratory settings, lithium-ion battery systems in electric vehicles (EVs) lack comprehensive, high-frequency sensor data as a result of cost constraints. Consequently, current research on battery fault diagnosis primarily focuses on binary detection, distinguishing between faulty and normal data. To diagnose various faults and offer valuable insights to upstream manufacturers, this study initially reviews the typical manifestations of battery failures in real-world EVs and elucidates their potential causes. Subsequently, we proposed a general and efficient multi-fault diagnosis framework applicable to cross-vehicle scenarios. The framework introduces a feature transformation method to obtain consistent diagnostic feature maps across different types of EVs while integrating Squeeze-and-Excitation (SE) modules for dynamic feature extraction. Further, a novel transfer metric, maximum mean square discrepancy (MMSD), is combined with domain adversarial learning to achieve cross-domain feature alignment. Additionally, a classifier with a dynamic margin is employed to enhance the diagnostic performance in imbalanced sample conditions. The proposed framework's effectiveness is validated in three cross-vehicle fault diagnosis scenarios. The results show that the proposed framework outperformed other advanced models in diagnostic accuracy, providing a fresh perspective for tackling this challenging scenario.