Adaptive Label Refinement Network for Domain Generalization in Compound Fault Diagnosis

Area of Science:

• Machine Learning
• Artificial Intelligence
• Industrial Fault Diagnosis

Background:

• Domain generalization (DG) is crucial for real-world fault diagnosis but challenged by compound faults and limited multi-source data.
• Existing DG methods often require extensive data, which is impractical for industrial settings due to cost and operational constraints.
• Hard labels and label smoothing inadequately represent complex fault relationships, hindering cross-domain robustness.

Purpose of the Study:

• To develop a novel adaptive label refinement network (ALRN) for effective domain generalization in compound fault diagnosis.
• To enable robust model training using imperfect labels under source-scarce conditions (one or two source domains).
• To create richer, more robust soft labels that capture inter-class semantic similarities.

Main Methods:

• An adaptive label refinement network (ALRN) was designed, leveraging a convolutional neural network (CNN) for initial predictions.
• Iterative label refinement using sample-wise cross-entropy loss as an adaptive weighting factor to compute weighted averages of predictions.
• A label refinement stability coefficient, based on max-min Kullback-Leibler (KL) divergence ratio, was proposed to assess label quality and determine iteration termination.

Main Results:

• ALRN achieved accuracy gains exceeding 22% on unseen operating conditions compared to a conventional CNN baseline.
• The proposed method demonstrates superior performance with only one or two source domains for training.
• The refined soft labels effectively encode semantic similarities between fault classes, enhancing diagnostic accuracy.

Conclusions:

• The adaptive label refinement network (ALRN) provides a novel and practical solution for cross-domain compound fault diagnosis with imperfect supervision.
• ALRN significantly enhances cross-domain diagnostic performance, particularly under source-scarce conditions.
• The method offers a robust approach to learning with imperfect labels, improving model generalization capabilities.