A domain generalization network for imbalanced machinery fault diagnosis

Area of Science:

• Machine Learning
• Artificial Intelligence
• Industrial Fault Diagnosis

Background:

• Traditional imbalanced fault diagnosis (IFD) models struggle with domain shifts from varying working conditions.
• Domain Generalization (DG) offers improved generalization by learning domain-invariant features.
• Limited fault samples in real-world industrial settings pose a significant challenge for IFD.

Purpose of the Study:

• To enhance imbalanced fault diagnosis (IFD) performance under domain shifts.
• To develop a robust method for generalizing fault diagnosis to unseen conditions.
• To address the challenge of limited fault data in industrial applications.

Main Methods:

• Proposing a Domain Mixed-Enhanced Domain Generalization Network (DEMDGN).
• Utilizing mixup-based data augmentation for enhanced feature representation.
• Employing domain-based discrepancy metrics to align feature distributions across heterogeneous source domains.

Main Results:

• DEMDGN effectively creates domain-invariant features for robust fault diagnosis.
• The method successfully addresses both class imbalance and domain shift problems.
• Experiments on marine machinery and bearing datasets demonstrate superior diagnostic performance.

Conclusions:

• The proposed DEMDGN significantly improves fault diagnosis accuracy in the presence of domain shifts and data imbalance.
• DEMDGN offers a promising approach for real-world industrial fault diagnosis applications.
• The study highlights the effectiveness of combining data augmentation and domain alignment for robust generalization.