Machine Learning-Based Thermal Runaway Prediction and Fault Diagnosis for Lithium-Ion Battery Packs in Electric Vehicles

Abstract

Thermal runaway in lithium-ion batteries presents a major safety challenge for electric vehicles because localized cell failure can develop into a pack-level hazardous event. Early-warning approaches must distinguish genuine failure precursors from normal changes caused by load variation, ambient conditions, aging, and measurement uncertainty. This article proposes a physics-guided, topology-aware machine-learning framework for thermal-runaway prediction, fault diagnosis, and fault localization in electric-vehicle battery packs. The framework combines electrical and thermal measurements with gas, pressure, mechanical, and acoustic signals when these modalities are available. An electrothermal observer estimates expected cell behavior, while a causal temporal model processes measurement histories and observer residuals. A graph-based representation incorporates relationships among neighboring cells to support cell-level localization and propagation-risk estimation. Separate prediction heads estimate thermal-runaway risk within predefined time horizons, classify observable fault categories, and identify affected cells or modules. A validation protocol is proposed using experiment-level separation, chemistry-aware evaluation, event-based detection metrics, probability calibration, and sensor-failure testing. Machine-learning outputs support, but do not replace, independent battery protection. The contribution is an integrated and testable research methodology; experimental superiority, warning lead time, and deployment readiness remain to be established through independent validation