A deep neural network for multi-fault diagnosis of battery packs based on an incremental voltage measurement topology

Fault diagnosis of batteries has been a crucial task to ensure safe operation of energy storage systems. In battery packs, a comprehensive diagnosis strategy for multi-fault is always a challenge. Especially, soft short circuit (SSC) faults with insignificant voltage drops also make the task of diagnosing battery pack faults more challenging. This paper presents a multi-fault diagnosis method for battery packs using deep neural networks and an autoencoder (AE) framework. The method employs the residual difference between the original input features and the features reconstructed by the AE as a fault indicator, determining if the residual exceeds a threshold to produce the final diagnosis. We introduce an incremental voltage measurement topology and construct a high-dimensional voltage signature matrix as the model input. Multi-fault experiments demonstrate the method's accuracy in diagnosing various faults. Comparative studies with other common methods show significant advantages in external short circuit (ESC) fault diagnosis, achieving a recall rate of 96.29 % and an F1 score of 93.74 %. The method excels particularly in the rapid and accurate diagnosis of soft short circuit (SSC) faults, with an F1 score of 99.13 %. A model ablation study confirms the superior performance of the proposed method in battery pack fault diagnosis.