Domain Perturbation With Uncertainty for Bearing Fault Diagnosis Under Unseen Conditions

Area of Science:

• Machine Learning
• Artificial Intelligence
• Engineering

Background:

• Domain adaptation (DA) excels in cross-domain fault diagnosis but requires target data.
• Existing domain generalization methods for fault diagnosis have limited performance improvement.
• Current methods often fail to capture fault information beyond source domain distributions.

Purpose of the Study:

• To develop advanced domain generalization techniques for fault diagnosis.
• To address limitations of existing methods when target domain data is inaccessible.
• To enhance model generalization performance for unseen fault conditions.

Main Methods:

• Proposed multiplicative noise Gaussian perturbation strategy to simulate domain shift uncertainty.
• Introduced additive noise linear fusion strategy to ensure authenticity of generated feature styles.
• Combined feature statistics with random convex weights for reliability and diversity.

Main Results:

• The proposed strategies effectively capture fault information beyond source domain distributions.
• Generated diversified feature styles improve the network's ability to learn fault information.
• Experimental results on public and real datasets demonstrate significant effectiveness.

Conclusions:

• The novel approach enhances fault diagnosis generalization performance without target domain data.
• The combination of perturbation and fusion strategies offers a robust solution for cross-domain fault diagnosis.
• This work advances the field of domain generalization for industrial fault diagnosis applications.