Fault Diagnosis Method of Special Vehicle Bearing Based on Multi-Scale Feature Fusion and Transfer Adversarial...

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Rolling bearings are critical components in machinery, and their failure can lead to significant downtime and safety hazards.
• Traditional fault diagnosis methods struggle with feature extraction, accuracy, and overfitting, especially under complex operating conditions.
• The need for robust fault diagnosis methods that can generalize across different datasets and operating environments is crucial.

Purpose of the Study:

• To develop an advanced rolling bearing fault diagnosis method addressing limitations in feature extraction and cross-domain generalization.
• To enhance diagnostic accuracy and anti-noise capability through multi-scale feature fusion and attention mechanisms.
• To enable effective fault diagnosis across diverse operating conditions, equipment types, and virtual-real migrations using transfer learning.

Main Methods:

• A multi-scale convolutional fusion layer was designed for effective fault feature extraction from vibration signals at multiple time scales.
• A feature encoding fusion module utilizing the multi-head attention mechanism was employed for enhanced feature fusion and contextual modeling.
• Domain adaptation (DA) cross-domain feature adversarial learning was implemented to achieve domain-invariant feature extraction, minimizing data distribution discrepancies.

Main Results:

• The proposed method demonstrated superior performance in cross-domain and variable load environments.
• Experimental validation using public and specialized vehicle bearing datasets showed exceptional diagnostic accuracy.
• The method achieved an average migration fault diagnosis accuracy rate of up to 98.65% in multiple cross-domain transfer learning tasks.

Conclusions:

• The developed method significantly improves data feature extraction capabilities for rolling bearing fault diagnosis.
• The approach offers a robust and accurate solution for fault diagnosis, particularly in cross-domain and complex operational scenarios.
• This work advances the field of intelligent fault diagnosis by enabling effective transfer learning for rolling bearings.