Fault detection for Li-ion batteries of electric vehicles with feature-augmented attentional autoencoder.

The attainment of fault detection within battery packs is of paramount importance in ensuring the safe and reliable operation of electric vehicles. However, in the early stages of battery failure, its manifestations are often not obvious, making it difficult for conventional algorithms to detect them in time. Thus, this paper proposes a novel fault detection framework for battery packs to reduce detection time and eliminate false alarms. This framework employs a feature-augmented autoencoder that can effectively utilize the early-stage features for fault detection. The commonly used autoencoder is improved by applying a memory module to amplify the fault features between normal and abnormal cells and employing a similarity loss function to enhance the similarity of features for normal batteries. Then, the augmented features are input into the local outlier factor algorithm for battery fault detection. Real operational data from electric vehicles are used to train and test the autoencoder, covering the battery voltages, current, and temperature data. The results show that the proposed strategy can detect battery faults at most 10 hours earlier, and pinpoint the faulty battery without false alarms when compared to the existing methods.