Toward early battery anomaly detection: A multimodal dataset and deep learning framework for lithium-ion batteries

Lithium-ion batteries (LIBs) power electric vehicles and grid-scale energy storage systems, yet their high energy density makes them prone to severe safety incidents, including thermal runaway and fire hazards. Traditional threshold-based battery management systems mainly rely on fixed safety limits and may miss early abnormal patterns that appear before a fault becomes severe. This paper presents a deep learning-based spatiotemporal anomaly detection framework that fuses multivariate time-series electrical measurements with thermal imaging to identify early-stage battery anomalies. We introduce a novel multimodal dataset acquired from a pouch cell under diverse cycling profiles, encompassing both normal operation and four categories of induced fault conditions. Building on this dataset, we propose an unsupervised multimodal autoencoder and systematically compare three fusion strategies early, bottleneck, and late fusion to investigate how the stage of modality integration influences detection performance. The late fusion architecture, which processes time-series data through a fully-connected encoder and thermal images through a convolutional neural network while preserving modality-specific representations, achieves the best overall performance: 91.85% accuracy, 77.23% precision, 87.50% recall, 82.04% F1-score, and 95.78% AUC. Case study analysis further demonstrates that late fusion effectively suppresses false positives from sensor malfunctions and detects localized thermal hotspots invisible to contact sensors.