Rolling Bearing Fault Diagnosis Model Based on Multi-Scale Depthwise Separable Convolutional Neural Network...

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Rolling bearing diagnostics face challenges due to complex operating conditions and scarce labeled data, hindering fault feature extraction and accuracy.
• Existing methods struggle with variable environments and few-shot learning scenarios, limiting their real-world applicability.

Purpose of the Study:

• To develop a robust fault diagnosis model for rolling bearings that addresses data limitations and operational variability.
• To improve diagnostic accuracy and generalization capabilities in challenging industrial settings.

Main Methods:

• Raw bearing signals are converted to Gramian angular difference field (GADF) images to preserve temporal and phase information.
• A novel model architecture combining a spatial attention (SA) mechanism with a multi-scale depthwise separable convolution module is proposed.
• The model is pre-trained and transferred to new environments for few-shot fault diagnosis.

Main Results:

• The proposed model demonstrates superior fault recognition performance on bearing datasets and industrial field data.
• Experimental results validate the model's effectiveness in diverse working conditions, small-sample scenarios, and real industrial environments.
• The spatial attention mechanism effectively focuses on discriminative features and suppresses noise.

Conclusions:

• The integrated spatial attention and multi-scale convolution model offers outstanding fault recognition and generalization capabilities for rolling bearings.
• The approach significantly enhances diagnostic accuracy in few-shot learning and variable operating conditions, proving its industrial relevance.