Machine learning-based predictive maintenance for industrial bearing fault diagnosis using vibration signal features and hybrid ensemble models
Abstract
This study presents a reproducible machine-learning workflow for industrial bearing fault diagnosis as a predictive-maintenance decision-support component. Drive-end vibration signals from the public Case Western Reserve University (CWRU) benchmark were segmented into non-overlapping 2048-sample windows. Twenty-eight time-domain, spectral, and envelope features were used to classify normal, inner-race, ball, and outer-race conditions. A hybrid soft-voting ensemble combined logistic regression, calibrated radial-basis-function support vector machines, random forest, extra trees, and gradient boosting. To reduce direct window leakage, the models were trained with 0, 1, and 2 hp files and tested on the held-out 3 hp files. The experiment contains 800 windows, with 600 for training and 200 for testing. The hybrid ensemble achieved 1.000 accuracy and macro F1-score under clean testing. A controlled noise experiment retained a 1.000 macro F1-score through 5 dB signal-to-noise ratio (SNR), but declined to 0.609 at 0 dB. The result demonstrates a transparent benchmark workflow while making clear that clean laboratory scores do not replace field validation.