A Fast Graph Construction-Driven Rotating Machine Fault Diagnosis Method Using Edge Predictor.

Area of Science:

• Machine Learning
• Mechanical Engineering
• Data Science

Background:

• Graph-based methods excel at extracting relational information for machine fault diagnosis.
• The computational intensity of K-nearest neighbor graph (KNNG) construction hinders practical application.
• Existing methods face challenges with scalability due to heavy computational demands.

Purpose of the Study:

• To propose a fast graph construction-driven rotating machine fault diagnosis method.
• To reduce the computational burden associated with traditional KNNG construction.
• To enhance the efficiency of graph-based fault diagnosis techniques.

Main Methods:

• A novel edge predictor is introduced, pretrained for edge connection prediction.
• The edge predictor learns to generate a distance matrix from an initial KNNG (IKNNG).
• Direct input of samples to the edge predictor bypasses traditional distance matrix calculations for rapid KNNG construction.

Main Results:

• The proposed method achieves comparable diagnostic performance to existing graph data-driven approaches.
• Significant reduction in computational load compared to traditional KNNG construction.
• Experimental validation confirms the efficiency and effectiveness of the fast graph construction technique.

Conclusions:

• The edge predictor facilitates efficient KNNG construction for rotating machine fault diagnosis.
• The method offers a computationally efficient alternative without compromising diagnostic accuracy.
• This approach addresses the scalability limitations of traditional graph-based fault diagnosis.