Cell-Wise Fault Warning for Lithium-Ion Battery Packs Using Local Outlier Factor and Validity-Gated Extreme-Value Thresholds

Abstract

Early fault warning in lithium-ion battery packs is challenged by weak labels and changing operating conditions. This study combines cell-wise Local Outlier Factor (LOF) scoring with validity-gated Streaming Peaks-Over-Threshold (SPOT). Five descriptors are used to summarize each seven-sample median-centered voltage trajectory, and the maximum cell LOF is logarithmically transformed into a pack score. A generalized Pareto distribution (GPD) supplies the threshold when calibration diagnostics support the fitted tail; otherwise, an empirical quantile is used. Across nine real-world vehicles and six AI-generated CH-BatteryGen cases, 12 streams retain GPD-SPOT and three use the empirical fallback. At a target risk parameter q = 0.01, aggregate false-alarm rates are 1.85% for pre-event baselines and 1.64% for independent normal vehicles. Seven recorded field events have an alarm within the preceding 24 h, with a median lead time of 17.39 h. However, car10 provides only 0.053 h of notice, and window-level recall is 0.029 under broad ±24 h event labels. The supplied faulty cell ranks first in all three labeled lithium iron phosphate cases. The workflow provides statistical fault warnings and suspicious-cell rankings without fault-labeled model training.