A Pseudo-Labeling Multi-Screening-Based Semi-Supervised Learning Method for Few-Shot Fault Diagnosis.

Area of Science:

• Machine Learning
• Mechanical Engineering
• Signal Processing

Background:

• Few-shot fault diagnosis is crucial in industry where labeled data is scarce.
• Existing semi-supervised learning (SSL) methods struggle with low-quality labeled samples, degrading performance.
• Prototypical network methods often overlook individual sample contributions to class prototypes.

Purpose of the Study:

• To propose a novel SSL method for few-shot bearing fault diagnosis.
• To address the performance degradation caused by low-quality samples in industrial datasets.
• To enhance the generalization ability and accuracy of bearing fault diagnosis models.

Main Methods:

• A pseudo-labeling multi-screening strategy is employed for accurate pseudo-label selection.
• An AdaBoost-based weighted technique is used to create robust class prototypes from clustered samples.
• Squeeze and excitation blocks are utilized for effective feature extraction, enhancing relevant information.

Main Results:

• The proposed method demonstrates superior performance in few-shot bearing fault diagnosis.
• It effectively mitigates the negative impact of low-quality labeled samples.
• Experimental validation on three bearing datasets confirms its effectiveness against state-of-the-art methods.

Conclusions:

• The developed SSL method offers a robust solution for few-shot bearing fault diagnosis in industrial settings.
• Pseudo-label screening and adaptive prototype weighting significantly improve model generalization and accuracy.
• The approach provides a promising direction for handling noisy datasets in machine diagnostics.