Cross-Dataset Evaluation of Multisensor Self-Supervised Anomaly Detection for Early Thermal-Runaway Warning in Electric Vehicle Battery Cells

Abstract

Thermal runaway can develop quickly and create severe hazards in electric vehicle battery systems. Early warning is difficult because public fault records are scarce and sensor sets differ across experiments. This study evaluates a multisensor self-supervised anomaly detector using three public cell-level datasets. Healthy fast-charge records supplied unlabeled normal data from six lithium iron phosphate cells. Five target records covered dynamic operation, local heating, uniform heating, and two prismatic NMC abuse tests. A masked denoising autoencoder learned 42 thermal, electrical, pressure, and availability features. We compared it with a conventional autoencoder, Isolation Forest, PCA, one-class SVM, and temperature features. Evaluation used physical-cell splits, leakage-safe prefix calibration, real-time persistence, and block-bootstrap confidence intervals. The masked model detected all three runaway events. Corrected warning leads were 43.77, 34.52, and 1.25 minutes. However, its pooled AUROC was 0.744, and its AUPRC was 0.492. The temperature score performed better, reaching 0.924 AUROC and 0.812 AUPRC. The masked model's negative-window false-positive rate was 0.546. It also produced one persistent warning during the healthy dynamic-load record. Masked training did not improve the conventional autoencoder significantly. The paired AUROC difference was -0.001, with p = .294. These findings show that event detection alone can hide severe calibration problems. Simple thermal features remained more portable under strong domain shift. Multisensor self-supervision still offers useful representations, but it needs better domain adaptation and event labeling. The study provides a reproducible benchmark for safer future validation.