Metric Learning-Guided Semi-Supervised Path-Interaction Fault Diagnosis Method for Extremely Limited Labeled Samples...

Area of Science:

• Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Intelligent fault diagnosis faces challenges due to limited labeled data and varying operational conditions.
• Existing methods struggle with adaptability across diverse working environments.

Purpose of the Study:

• To propose a novel network, the triplet-guided path-interaction ladder network (Tri-CLAN), for robust intelligent fault diagnosis.
• To develop a method for learning distribution-invariant representations from limited labeled and abundant unlabeled data.

Main Methods:

• Utilized an encoder-decoder structure with path interaction and simplified CNN architecture.
• Incorporated an element additive combination activation function for network efficiency.
• Introduced metric learning, specifically triplet loss, to the feature space for improved sample discrimination.

Main Results:

• The Tri-CLAN network demonstrated effective utilization of unlabeled data with fewer parameters.
• Metric learning enabled the mining of hard samples and learning of working condition-independent representations.
• Experimental results validated the generalization and applicability of the proposed Tri-CLAN model.

Conclusions:

• Tri-CLAN offers a promising solution for intelligent fault diagnosis under data scarcity and variable conditions.
• The combination of network architecture and metric learning significantly enhances diagnostic performance.
• The approach facilitates the development of more adaptable and reliable industrial diagnostic systems.