Knowledge correlation graph-guided multi-source interaction domain adaptation network for rotating machinery fault...

Area of Science:

• Machine Learning
• Mechanical Engineering
• Artificial Intelligence

Background:

• Multi-source domain adaptation is challenging due to distribution discrepancies and expanded data categories.
• Adapting knowledge from labeled source domains to unlabeled target domains is crucial for realistic applications.
• Existing methods struggle with aligning multiple heterogeneous data sources for fault diagnosis.

Purpose of the Study:

• To develop a novel network for rotating machinery fault diagnosis using multi-source domain adaptation.
• To address challenges of distribution discrepancies and data category expansion in domain adaptation.
• To improve knowledge interaction and propagation across multiple labeled source domains and an unlabeled target domain.

Main Methods:

• Introduced a knowledge correlation graph-guided multi-source interaction domain adaptation network (KCGMIDAN).
• Utilized comprehensive feature representations (CFR) updated across epochs to promote knowledge interaction.
• Constructed a knowledge correlation graph (KCG) for inter-domain knowledge propagation and employed deep graph networks for sample recognition.

Main Results:

• KCGMIDAN effectively promotes knowledge interaction and propagation among various domains.
• Designed losses improved intra-class compactness and inter-class separation of features.
• Achieved superior recognition performance compared to existing domain adaptation methods in rotating machinery fault diagnosis.

Conclusions:

• The proposed KCGMIDAN framework offers a robust solution for fault diagnosis in rotating machinery.
• The knowledge correlation graph effectively bridges knowledge gaps between diverse data domains.
• This approach demonstrates significant potential for real-world applications requiring unsupervised domain adaptation.