Dynamic pseudo-label guided adversarial multi-scale graph convolutional network for cross-domain fault diagnosis

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Domain adaptation is crucial for mechanical fault diagnosis with limited labeled data.
• Existing methods often overlook feature cluster properties and lack robust pseudo-label screening.
• Cross-domain fault diagnosis faces challenges due to varying operating conditions.

Purpose of the Study:

• To propose a Dynamic Pseudo-Label Guided Adversarial Multi-Scale Graph Convolutional Network (DPAMGCN) for unsupervised cross-domain fault diagnosis.
• To enhance feature extraction and clustering properties for improved diagnostic accuracy.
• To develop a dynamic pseudo-label filtering strategy for better generalization.

Main Methods:

• Designed a network architecture combining a multi-scale CNN with an MRF-GCN for feature extraction.
• Constructed a multi-objective total-loss function integrating geometric loss for joint optimization.
• Implemented a dynamic threshold-based pseudo-label filtering strategy.

Main Results:

• DPAMGCN demonstrated outstanding cross-domain diagnostic performance on benchmark datasets.
• The proposed optimization strategy and pseudo-label screening mechanism significantly enhanced generalization.
• Effective feature clustering and spatial distribution were achieved.

Conclusions:

• The DPAMGCN model offers a robust solution for unsupervised cross-domain mechanical fault diagnosis.
• Integrating geometric loss and dynamic pseudo-label filtering improves model performance.
• The study highlights the importance of considering feature clustering in domain adaptation for fault diagnosis.