Efficient battery fault monitoring in electric vehicles: Advancing from detection to quantification

Effective monitoring of battery faults is crucial to prevent and mitigate the hazards associated with thermal runaway incidents in electric vehicles (EVs). This paper presents a novel framework for comprehensive fault monitoring, encompassing detection, identification, and quantification. Initially, a hybrid neural network combining long short-term memory and Bayesian neural network is proposed to estimate cell voltage and its uncertainty distribution. The estimated uncertainties are utilized as adaptive thresholds for fault detection. Subsequently, statistical analysis is employed to distinguish sudden thermal runaway from data anomaly, self-discharge, and internal short circuits (ISC). Two sensitive features are extracted, and unsupervised learning is applied for the identification of other battery faults and data anomalies. Finally, a new fault quantification strategy incorporating dynamic bounds based on uncertainties is proposed. The effectiveness and advancements of the proposed framework are verified using real-world data from three EVs with 334 battery cells. Verification results approve the accurate detection, identification, and quantification capabilities of the proposed method for early faults. Notably, compared to the alarm time provided by the existing battery management system, the proposed method exhibits significantly earlier warning times for self-discharge, ISC, and sudden thermal runaway, with durations of 155.10 h, 586.87 h, and 3 s, respectively.