Thermo-vibrational fusion for early-stage ball bearing fault diagnosis: a multi-modal machine learning framework

Rolling bearing failures account for 40–50% of rotating machinery breakdowns, but the complementary diagnostic potential of thermal sensing remains underutilized. This study presents a thermo-vibrational fusion framework for early-stage ball bearing fault diagnosis using the KAIST run-to-failure dataset (128 h, NSK 6205 bearing). A 40-dimensional feature vector comprising FFT-based spectral features at the bearing defect frequencies (BPFO, BPFI, BSF), sub-band energy ratios, spectral entropy, and six time-domain statistical descriptors across all four sensor channels is extracted. Three classifiers – SVM, Random Forest, and DNN – are evaluated under five-fold stratified cross-validation. The DNN achieves 95.7% accuracy (AUC = 0.983), RF 94.1% (AUC = 0.975), and SVM 91.3% (AUC = 0.961). Feature importance analysis confirms vibration kurtosis (14.2%) and bearing temperature RMS (5.8%) as the most discriminative features, validating the multi-modal approach. On this dataset, acquired under constant speed and load, thermal features preceded an unambiguous vibration RMS escalation by 15–25 h during incipient degradation, improving early-stage precision by 4–6% points over vibration-only classification. The novelty of the work lies in providing the first quantitative characterisation of this thermal-to-vibration diagnostic lead time on a publicly available, fully synchronised run-to-failure dataset, benchmarked across three classifier families under an identical leakage-free protocol with physically grounded feature attribution; the transfer of these findings to variable-speed and fluctuating-load environments is discussed as a limitation of the present study.