Mechanism-Informed Fault Diagnosis of Battery Thermal Runaway Using CRS-DenseNet with Risk Accumulation Features

The widespread transition to electric mobility has made battery safety an urgent engineering challenge. Conventional diagnostics relying on instantaneous battery management system (BMS) telemetry often overlook long-term degradation cues. To address this, we present a mechanismguided diagnostic architecture that utilizes real-world EV fleet data, pivoting to a risk accumulation paradigm rather than relying solely on real-time snapshots. By merging a populationlevel Comprehensive Risk Score (CRS) with a DenseNet neural architecture, we establish a dual-validation framework for robust feature selection and non-linear mapping. Evaluated against real-world data, the proposed model yields a flawless 100% diagnostic accuracy. Furthermore, utilizing SHapley Additive exPlanations (SHAP) for model interpretation successfully decouples feature contributions, pinpointing prolonged exposure to high state-of-charge (SOC) paired with mid-to-high C-rate charging as the primary catalysts for failure, thereby guiding targeted BMS calibration and optimization.