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

Area of Science:

• Electrical Engineering and Machine Learning
• Condition Monitoring and Predictive Maintenance

Background:

• Induction motor bearing faults are critical for industrial operations, impacting predictive maintenance and Industrial Internet of Things (IIoT) systems.
• Existing diagnostic models often falter under unseen operating conditions, necessitating robust evaluation protocols beyond simple data splits.

Purpose of the Study:

• To comparatively evaluate the robustness of different induction motor bearing fault diagnosis models.
• To assess models using vibration and phase-current signals under measurement-wise and operating-condition holdout scenarios.

Main Methods:

• Evaluated raw temporal 1D-CNN, STFT-based 2D-CNN, and vibration-current fusion models.
• Utilized a subset of the Paderborn bearing dataset with measurement-wise and operating-condition holdout validation.
• Conducted multi-seed validation and compared against classical baselines (e.g., XGBoost) with handcrafted features.

Main Results:

• The 1D-CNN Early Fusion model demonstrated superior robustness under operating-condition shifts, achieving a mean Macro-F1 of 0.8626.
• This lightweight model (0.282 MB) offers a favorable robustness-complexity trade-off.
• Classical baselines, like vibration-feature XGBoost, performed well (Macro-F1 of 0.8582) but lightweight CNN fusion showed better robustness.

Conclusions:

• Lightweight vibration-current early fusion models offer a promising solution for robust bearing fault diagnosis under operational variations.
• Generalization to completely unseen bearing identities remains a significant challenge, requiring further research.
• Future work should include embedded hardware validation and cross-domain studies for practical deployment.