A unified ensemble learning framework for electric motor fault diagnosis, thermal state prediction, and remaining useful life estimation

Electric motors are critical components of modern industrial and transportation systems, making accurate fault diagnosis and predictive maintenance essential for operational reliability. This study proposes a unified machine learning framework for motor fault classification, temperature prediction, and remaining useful life (RUL) estimation using three publicly available electric motor datasets. The proposed pipeline integrates sliding-window segmentation, multi-domain feature extraction, consensus-based feature selection, and physically motivated descriptors derived from electrical and thermal measurements. Five machine learning algorithms, namely Random Forest, Support Vector Machine (SVM), XGBoost, LightGBM, and Long Short-Term Memory (LSTM), were systematically evaluated. For fault diagnosis on the Ottawa UOEMD-VAFCVS dataset, evaluated with a file-level train–test split in which all segments of a recording are kept in the same partition, LightGBM achieved the highest performance on the held-out recordings (Accuracy = 1.0000, F1-score = 1.0000, ROC-AUC = 1.0000). Sliding-window segmentation increased the number of trainable observations from 128 to 2,432 and improved the macro F1-score from 0.569 to above 0.994 across all models. For thermal prediction on the PMSM dataset, LightGBM obtained R² values of 0.993, 0.991, 0.989, and 0.987 for rotor, stator winding, stator tooth, and stator yoke temperatures, respectively, with mean absolute errors below 1.4°C. In an exploratory experiment on the synthetically generated EVIoT dataset, which contains no genuine run-to-failure trajectories, XGBoost and LightGBM reached R² = 0.933 on a 30-step smoothed RUL target; this value largely reflects the smoothing of the target and is reported only as a demonstration of the pipeline, not as validated prognostic capability. A Friedman test over N repeated runs indicated significant differences among the five classifiers (χ² = 11.556, p = 0.021, Kendall’s W = W). Within the evaluated configurations, feature-engineered gradient-boosting models matched or exceeded the LSTM baseline at low computational cost; a systematic comparison with end-to-end deep architecture is left for future work.