Feature engineering-driven multi-scale voltage anomaly detection for Lithium-ion batteries in real-world electric vehicles

Battery fault diagnosis and thermal runaway warnings hold significant implications for the safety of electric vehicles. However, developing a reliable battery fault detection method that encompasses voltage anomaly patterns remains challenging due to the concealment and uncertainty of anomalies under complex profiles. A framework for detecting battery multi-scale voltage anomalies using feature engineering is proposed. Hundreds of dimensionless indicators (DIs) are constructed by assigning parameters to the unified mathematical model of DI. The Laplacian Score and Sequential Forward Selection are then successively implemented on all DIs to analyze their sensitivity and redundancy to anomalies. The optimal DI set, which includes two DI central moments for short-timescale discharging anomalies and a waveform factor for short-timescale charging anomalies and long-timescale anomalies, is screened out as the most sensitive features. The adaptive clustering algorithm is applied to the standardized differential DI set for anomaly isolation, and the standardized differential voltage is calculated to identify voltage anomaly patterns. The proposed method is designed to distill the optimal feature set for multi-scale battery voltage anomalies and amplify minor abnormal fluctuations of battery voltage signals into conspicuous outliers. The robustness has been corroborated using data from 197 electric vehicles, yielding an accuracy of 96.95 %. The method offers precise delineation of the patterns and moments of voltage anomalies and can provide early warnings for battery systems in real-world electric vehicles.