Early Warning of Power Battery Faults Based on Multi-Scale Time Window Features and External Safety Characteristics

To enable the early detection of thermal runaway risks in power batteries, this paper proposes a fault early warning method based on multi-scale time window features and the XGBoost classification model. Utilizing vehicle-measurable external signals - such as voltage, temperature, current, state of charge (SOC), and driving mileage - this method constructs feature variables including the statistical characteristics of cell voltage and battery pack temperature, as well as the complexity of the total voltage. Furthermore, these features are aggregated across three time scales (short-term, medium-term, and long-term) to simultaneously capture abrupt signal variations, dynamic state evolution, and long-term heat accumulation trends. At the modeling level, XGBoost is employed to learn these multi-scale features, with its hyperparameters fine-tuned through 5-fold cross-validation and grid search. The experimental results show that on the test set, the proposed approach achieves an AUC of 0.971, a precision of 0.75, a recall of 0.60, and an accuracy of 0.91. Its overall performance is markedly superior to that of baseline models, including logistic regression, support vector machines, decision trees, and random forests.