Matching Pursuit Network: An Interpretable Sparse Time-Frequency Representation Method Toward Mechanical Fault...

Area of Science:

• Mechanical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Rotatory machinery operates in challenging environments with noise and variable conditions.
• Time-frequency representation is crucial for analyzing nonstationary signals and identifying transient fault features.
• Existing methods struggle to extract robust fault features from machinery under variable speeds.

Purpose of the Study:

• To propose a novel sparse time-frequency representation (STFR) method for mechanical fault diagnosis.
• To develop a deep network capable of automatically learning discriminative features from time-frequency data.
• To enable robust and interpretable fault feature extraction for machinery operating under variable speeds.

Main Methods:

• A matching pursuit network (MPNet) utilizing interpretable matching pursuit (MP) units was constructed.
• A deep network structure was designed for signal decomposition and automatic feature learning.
• An optimization criterion with a structural similarity metric was employed for end-to-end model training.

Main Results:

• The MPNet successfully extracted robust and interpretable time-frequency features.
• The proposed method demonstrated superior performance compared to state-of-the-art time-frequency representation techniques.
• Model training and testing were validated using simulated and experimental gearbox fault data.

Conclusions:

• The MPNet offers a powerful solution for mechanical fault diagnosis in complex operating conditions.
• The method effectively addresses the challenge of extracting fault features from variable-speed machinery.
• MPNet provides a significant advancement in robust and interpretable time-frequency feature extraction for rotating machinery.