Autoencoder with Multi-Head Attention for Voltage Anomaly Detection in Electric Vehicle Battery

With the rapid growth of electric vehicle (EV) adoption, ensuring the reliability and safety of their battery systems is of significant importance. Current anomaly detection methods, which rely primarily on temporal dependencies, often overlook spatial patterns, leading to inaccurate detection. This paper introduces a novel autoencoder that incorporates a multi-head attention mechanism for voltage anomaly detection in EV batteries. First, comprehensive datasets from six real EVs operating under diverse and dynamic conditions were collected. Second, to address the complexity and variability of the data, various pre-processing techniques were employed, and the optimal method was selected to ensure data quality and consistency. Third, an improved deep learning autoencoder model was developed, incorporating multiple multi-head attention mechanisms to capture intricate patterns and temporal dependencies within the battery data. The effectiveness of the proposed method is verified by the tests on the comprehensive dataset of real electric vehicles operating under dynamic conditions.