Regime-Aware Physics-Guided Early Warning of Lithium-Ion Battery Thermal Runaway Using Thermo-Mechanical Signals

This paper presents a regime-aware, physics-guided framework for lithium-ion battery thermal-runaway warning under controlled mechanical abuse. It fuses temperature, voltage, force, deformation, and state-of-charge measurements. A lightweight convolutional classifier identifies safe, warning, and danger regimes from mechanical signals, which condition a causal temporal convolutional model through feature-wise modulation, physics-biased attention, and regime-dependent gating. Across 30 leave-one-experiment-out tests covering three state-of-charge levels and two loading protocols, the method reports an F1 score of 0.89, 15.6 seconds of mean warning lead time, 0.92 detection success, and a 2.7% experiment-level false-alarm rate.