IoT-Driven Robust Bearing Fault Diagnosis for Induction Motors Under Operating-Condition Shift.

Reliable bearing fault diagnosis in induction motors is essential for predictive maintenance and Industrial Internet of Things (IIoT) applications. However, diagnostic models that perform well under random or measurement-wise data splits may fail when deployed under unseen operating conditions. This study presents a robustness-oriented comparative evaluation of induction motor bearing fault diagnosis models using vibration and phase-current signals from a controlled medium subset of the Paderborn bearing dataset. Raw temporal 1D-CNN models, STFT-based 2D-CNN representations, and vibration-current fusion strategies were evaluated under measurement-wise and operating-condition holdout protocols. Under measurement-wise validation, the 1D-CNN Early Fusion model achieved a Macro-F1 score of 0.9251. Under the stricter operating-condition holdout setting, the same model achieved the highest robustness among the evaluated CNN models. Multi-seed validation confirmed its stability, with a mean Macro-F1 of 0.8626, a worst-case Macro-F1 of 0.7159, and a robustness score of 0.7850. The selected model remained lightweight, requiring 73,891 trainable parameters and an estimated model size of 0.282 MB. Additional revision experiments were conducted to address bearing-identity sharing and classical baseline comparisons. In the bearing-code-disjoint validation test, both raw temporal models showed reduced performance, and early fusion did not significantly outperform vibration-only learning. The 1D-CNN Vibration model achieved a mean Macro-F1 of 0.5616, while the 1D-CNN Early Fusion model achieved 0.5485; the paired Wilcoxon test was not significant (