Detection and Localization of Early Internal Short Circuits in Battery Packs Using Hybrid Machine Learning

Detection and localization of early internal short circuits (ISCs) in battery packs are critical for mitigating safety risks, including thermal runaway (TR). To address this issue, this study proposes an efficient early ISC detection and localization method based on a hybrid machine learning framework. Specifically, leveraging the excellent performance of deep learning methods in high-level feature learning, a lightweight 1D convolutional auto-encoder (1D-CAE) is designed and employed as an online ISC detector, eliminating the need for training on collected anomalous battery data. Furthermore, considering the slow and time-varying characteristics of early battery ISC, a median filter (MF) is employed to denoise the reconstruction error (RE) in real time, effectively reducing false alarms caused by complex measurement noise. The necessity and effectiveness of the MF in mitigating outliers and handling the strong nonlinearity of batteries are validated through comparisons with mainstream filtering methods. Finally, due to its computational efficiency and strong anomaly detection capability, an Isolation Forest (iForest) is adopted to locate the ISC battery cell. The effectiveness of the proposed method is validated using both simulated laboratory data and real-world operational data collected from two electric vehicles. Experimental results demonstrate that the proposed method achieves accurate and robust diagnostic performance for detecting early-stage ISC faults in battery packs.