An Incipient Multi-fault Diagnosis Method Based on Deep Learning for Large-Scale Lithium-ion Battery Packs

The application of effective fault diagnosis methods is a key measure to enhance the comprehensive performance of lithium-ion battery (LIB) systems. However, it is challenging for conventional threshold diagnosis methods to detect minor faults in the early stages. Therefore, an incipient multi-fault diagnosis method based on long short-term memory (LSTM) network is proposed. An interleaved current-voltage sequence is set as the input of the LSTM, while the differential voltage value is used as the output. The short-circuit and connection faults can be detected effectively by calculating a residual between the actual sample value and the predicted value generated from the standard reference model, which is trained by LSTM and capable of representing the volt-ampere relationship (VAR) of a healthy battery. The experiment results demonstrate that the proposed method can not only detect the type, location, occurrence moment, and duration of faults but also perform a superior accuracy across the whole state of charge (SOC) range. The lower limit of the detectable fault residual is as low as 0.01 V. In summary, the proposed incipient multi-fault diagnosis method has excellent generalizability and accuracy and is feasible and valuable in large-scale lithium-ion battery packs.