Reweighted Regularized Prototypical Network for Few-Shot Fault Diagnosis

Area of Science:

• Machine Learning
• Artificial Intelligence
• Industrial Automation

Background:

• Few-shot fault diagnosis (FSFD) is challenging due to limited faulty samples.
• Existing metric-based meta-learning methods often neglect intraclass and interclass distribution information.

Purpose of the Study:

• To develop an improved FSFD approach leveraging distribution information.
• To enhance diagnostic performance in low-data scenarios.

Main Methods:

• Proposed a novel reweighted regularized prototypical network.
• Introduced an intraclass reweighting strategy to stabilize fault prototype estimation.
• Developed a balance-enforcing regularization (BER) to address class imbalance and improve discrimination.

Main Results:

• The proposed method effectively reduces intraclass differences and enlarges interclass differences.
• Achieved improved metric space and diagnostic performance in few-shot learning.
• Demonstrated superior performance compared to state-of-the-art methods on benchmark and real-world datasets.

Conclusions:

• The novel FSFD approach significantly enhances diagnostic accuracy with limited data.
• The reweighted regularized prototypical network offers a robust solution for industrial fault diagnosis.
• Effective handling of data scarcity and class imbalance is crucial for FSFD success.