Acoustic Fingerprint Identification of Lithium-Ion Battery Thermal Runaway and Application in Fire Investigation

Identifying the specific battery type that caused a fire is a critical part of forensic investigation, yet it becomes difficult when severe heat destroys physical evidence. To provide a non-contact forensic solution, this study proposes an interpretable identification framework based on eXtreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP). Thermal runaway signals from prismatic, pouch, and cylindrical cells were recorded using various devices under realistic conditions, including environmental noise and wall obstructions. By integrating time-domain statistics, frequency-domain metrics, and Mel-frequency Cepstral Coefficients (MFCCs), we characterized the unique acoustic signatures of each battery format. The results show that battery structure determines the sound. Cylindrical cells produce high-frequency impulsive sounds due to valve jetting, while pouch cells produce low-frequency tearing sounds. Using these physical features, the optimized XGBoost classifier achieved an overall accuracy of 96.39% and an F1-score of 96.21% on a mixed dataset. This model outperformed Support Vector Machine and Random Forest classifiers. It proved robust even when processing low-quality audio from surveillance equipment. Additionally, SHAP analysis interprets the decision logic, identifying Peak Frequency and specific cepstral textures as the most important features. In conclusion, acoustic fingerprinting serves as independent, quantifiable evidence, enabling accurate ignition source identification even when traditional physical analysis is impossible.