XGBoost-Powered Predictive Analytics for Early Identification of Thermal Runaway in Lithium-Ion Batteries

Abstract

Lithium-ion batteries are pivotal in powering modern technology, from electric vehicles to portable electronics. However, their safety is challenged by the risk of thermal runaway, a critical failure mode leading to catastrophic consequences such as fires and explosions. This study presents a machine learning framework for the early detection of thermal runaway events using sensor data from over 210 open-source battery tests. The framework utilizes voltage, temperature, and force measurements from experimental mechanical indentation tests, with force data providing additional predictive value beyond standard BMS sensors. Key features such as the rate of temperature change and voltage change were engineered from raw time-series data. An XGBoost classifier was trained to detect critical patterns up to 20 s in advance, with lead-time shifting applied to simulate real-time warnings. Critical conditions were operationally defined as temperature exceeding 80 °C or voltage dropping below 3.0 V. The model achieved an F1-score of 0.98 on a test set of 734k data points from 42 independent mechanical indentation battery tests (natural class distribution: 45% critical, 55% normal). SHAP analysis revealed that low voltage (below 3.0 V) and rapid temperature rise (above 80 °C/s) were the most influential features. The system identified patterns 5–10 s before threshold crossing, with a mean detection of 8.3 s. This research demonstrates the potential for machine learning-enhanced battery safety, providing a foundation for future advancements in the field.