Multi-Axle Reference and Temporal-Consistency Deep SVDD for EMU Traction Motor Bearing Anomaly Detection Using Field...

Area of Science:

• Engineering
• Mechanical Engineering
• Condition Monitoring

Background:

• Field vibration monitoring of EMU traction motor bearings faces challenges like weak labels, operating condition fluctuations, and limited abnormal samples.
• Learning normal boundaries from single bearing positions can be unstable, leading to unreliable alarms from isolated score spikes.

Purpose of the Study:

• To propose a robust framework for field anomaly detection in EMU traction motor bearings under weak-label conditions.
• To enhance the reliability of anomaly detection by addressing limitations of single-point monitoring and unstable normal boundaries.

Main Methods:

• Developed a Multi-Axle Reference and Temporal-Consistency-enhanced Deep SVDD (MA-TC-Deep SVDD) framework.
• Constructed a compact 10-dimensional time-frequency representation and utilized multi-axle stable samples for one-class normal-boundary learning.
• Incorporated feature recalibration, temporal-consistency regularization, and causal smoothing for improved robustness.

Main Results:

• The MA-TC-Deep SVDD framework demonstrated superior performance compared to baseline methods, achieving an AUC of 0.909.
• Field data validation showed the framework could identify transitional and persistent state deviations in bearing positions.
• Achieved high Precision (0.887), Recall (0.802), and F1-score (0.843) under weak-label evaluation.

Conclusions:

• The proposed MA-TC-Deep SVDD framework effectively provides field anomaly warnings and severity interpretation for EMU traction motor bearings under weak-label monitoring.
• The method enhances robustness against field disturbances and offers improved anomaly detection performance.
• Results indicate the framework's utility for early warning but not as a replacement for detailed fault diagnosis or remaining useful life prediction.