Hierarchical Multiple-Instance Learning for Vehicle-Level Battery Anomaly Detection from Charging Data Streams

Vehicle-level battery labels paired with many charging records create a weakly supervised rare-event problem. We evaluated a two-level multiple-instance framework on 191 vehicles (184 normal and seven source-coded abnormal) from the public EVBattery archive, using 38,200 charging snippets. Nine normal-reference anomaly detectors were combined with fixed or learned vehicle-level pooling under six repetitions of five-fold vehicle-disjoint internal cross-validation. All preprocessing and detector fitting used training vehicles only; score mapping used aggregation-training vehicles; decision thresholds used a separate 13-14-vehicle validation partition containing only one or two abnormal vehicles. PCA reconstruction with 95th-percentile pooling produced the highest numerical repeat-mean average precision (0.342) among 46 evaluated pipelines. A leave-one-repeat-out selection-sensitivity diagnostic selected this pipeline in five of six analyses and yielded mean held-repeat average precision 0.318 ± 0.078; because all repeats reuse the same 191 vehicles, this is not an independent test estimate. After averaging six out-of-fold probabilities per vehicle, average precision was 0.371 and AUROC 0.796. Threshold-based metrics are exploratory: the 30 validation-derived thresholds ranged from 0.267 to 0.999, and 17 equalled 0.500. No pairwise comparison survived Holm correction, uncertainty was wide, and calibration was poor. The results identify PCA plus upper-tail pooling only as the highest numerical internal benchmark in this cohort, not as an unbiased estimate of a selected final model, confirmed physical diagnosis, or deployment-ready probability estimation.