When Fusion Hurts: Diagnosing Base-Detector Admissibility in Hybrid Machine-Learning Anomaly Detection for Electric-Vehicle Battery CAN Telemetry

Abstract

Fixed-threshold battery management systems can miss faults that appear across several signals at once, which has encouraged the use of hybrid machine-learning detectors for electric-vehicle (EV) batteries. This study examines when combining such detectors actually improves anomaly detection. Using a public EV Battery Charging Dataset of 1900 one-second observations, four controlled fault types were injected: voltage imbalance, thermal instability, resistance growth, and sensor malfunction. The record was segmented into 125 overlapping windows and represented by 36 statistical and temporal features, which principal component analysis reduced to seven components. Isolation Forest, Local Outlier Factor (LOF), and an autoencoder were evaluated individually and in majority-vote and weighted-score ensembles. Six more recent machine-learning and deep-learning detectors were evaluated on the same windows for comparison. A Support Vector Machine was excluded from fusion because its outputs could not be aligned with the ground-truth windows. LOF gave the best discrimination (ROC-AUC 0.852), whereas the weighted ensemble reached 0.805 and majority voting 0.623. None of the recent detectors outperformed LOF, and no difference was statistically significant. Fusion therefore did not improve on the strongest single detector. Detector admissibility, complete score alignment, and leakage-free validation should be verified before ensemble gains are claimed for battery monitoring.