Rolling Bearing Fault Monitoring for Sparse Time-Frequency Representation and Feature Detection Strategy.

Area of Science:

• Engineering
• Machine Learning
• Signal Processing

Background:

• Deep learning has advanced rotating machinery fault diagnosis.
• Traditional methods struggle with fault feature extraction.
• Object detection theory is underutilized in fault diagnosis.

Purpose of the Study:

• To develop an advanced fault diagnosis method for rotating machinery.
• To integrate sparse time-frequency analysis with object detection for improved feature extraction.
• To enhance the interpretability and accuracy of fault diagnosis models.

Main Methods:

• Utilized sparse short-term Fourier transform (SSTFT) with sparse constraints for high-resolution time-frequency representation (TFR).
• Employed proximal gradient descent (PGD) for efficient model optimization.
• Developed a fault diagnosis model using a region-based convolutional neural network (RCNN) to extract fault-specific TFR regions.

Main Results:

• Achieved high-quality TFR with improved resolution and no cross-term interference.
• Successfully extracted multiple fault-feature-characterizing regions from TFR.
• Demonstrated effective multicategory rolling bearing fault identification using simulation and experimental data.
• Outperformed existing fault diagnosis methods in validation tests.

Conclusions:

• The proposed SSTFT and RCNN-based method significantly enhances fault diagnosis in rotating machinery.
• This approach improves fault feature extraction and model interpretability.
• The method offers a more effective solution for identifying bearing faults compared to traditional techniques.