Thermal fault detection in battery systems using principal component analysis with adaptive thresholding

Lithium-ion cells pose serious safety risks when they enter a state with highly exothermic reactions known as thermal runaway. Because elevated temperature is the ultimate trigger for this failure mode, reliable and timely detection of abnormal cell temperature is critical. Early detection enables, user warning, fast emergency response, and provides the basis for effective active prevention strategies. In this work, we present an unsupervised data-driven approach that detects thermal faults by monitoring inter-cell voltage deviations. We apply principal component analysis (PCA) to capture systematic changes in voltage homogeneity that occur when a cell within a battery module heats abnormally. By systematically analyzing the effects of thermal stress on voltage homogeneity under varying operating conditions, we reveal requirements for a reliable detection method. Leveraging these insights, we introduce an adaptive thresholding mechanism. This novel approach significantly boosts the sensitivity to faults for a wide range of operating conditions while maintaining detection robustness. We validate the method through extensive experiments in which we externally heat a single cell within a module with the power of 1 W. Compared to both conventional linear PCA and nonlinear kernel PCA with a constant threshold, linear PCA with adaptive thresholding achieves a significantly better balance between sensitivity and robustness across the full range of test conditions.