From surrogate task to multiscale external-reference consistency learning: Adaptation-free cross-topology early warning of thermal runaway in electric vehicle batteries

Early warning of thermal runaway (TR) in electric vehicle batteries is critical for vehicle safety and fleet deployment. Deep models trained only on normal samples usually depend on reconstruction or prediction residuals and insufficiently encode the hierarchical constraints from local electrical neighborhoods, regional thermal environments, and pack-level operating boundaries. This study proposes a multiscale external-reference consistency learning framework for thermal runaway precursor warning without fault labels. It uses routine battery management system data. Voltage, temperature, current, SOC, battery-pack topology, and group-structure priors are organized as a heterogeneous graph. A shared temporal encoder and type-aware gated graph aggregation learn the expected consistency between each target cell and its multiscale external references. During inference, direct mismatch and intra-group structural mismatch evidence are fused into a unified risk score monitored by a globally fixed threshold and a consecutive-window rule. The model was trained and calibrated using only normal 95-cell vehicles, and then deployed without target topology adaptation. In the primary test set covering 95-cell vehicles and an unseen 96-cell topology, the method detected 10 of 11 TR vehicles, with an average lead time of 18 days for detected cases and one false alarm among 20 normal vehicles. In the supplementary 92-cell and 72-cell ternary Li-ion cohorts, both TR vehicles were detected, and no sustained warnings occurred among ten normal vehicles. The 198-cell LFP result is reported as a case-level observation under a larger shift in pack scale and chemistry