An Imbalanced Fault Diagnosis Method Based on Multi-Sensor Selection and Graph Attention Mechanism

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Data Science

Background:

• Bearing diagnostic errors are common due to imbalanced datasets (normal vs. fault data).
• Existing methods struggle with the significant data imbalance in bearing fault detection.
• Accurate fault diagnosis is crucial for preventing machinery failure and ensuring operational safety.

Purpose of the Study:

• To propose a novel method for effective bearing fault diagnosis under imbalanced dataset conditions.
• To enhance the accuracy and stability of fault diagnosis by addressing data imbalance.
• To introduce a robust feature extraction and selection mechanism for multi-sensor bearing data.

Main Methods:

• A graph attention convolutional neural network (SCGAT) was developed for feature extraction from multi-sensor data.
• Sensor sensitivity analysis was employed to filter and select relevant sensor information.
• Sensor correlation analysis was used to merge highly correlated sensor data, reducing redundancy.
• The integrated features were then fed into a classifier for final fault diagnosis.

Main Results:

• The SCGAT method demonstrated effective bearing fault diagnosis capabilities even with imbalanced datasets.
• Experimental validation on a power transmission simulation platform confirmed the method's performance.
• The proposed SCGAT model achieved higher diagnostic accuracy and superior stability compared to existing models.
• The sensitivity and correlation analysis modules successfully refined sensor data for improved diagnosis.

Conclusions:

• The SCGAT method offers a robust solution for bearing fault diagnosis in imbalanced data scenarios.
• The integration of sensitivity and correlation analysis enhances feature representation for accurate fault detection.
• This approach provides a significant advancement in condition monitoring and predictive maintenance for rotating machinery.
• The SCGAT model shows promise for real-world applications requiring reliable fault diagnosis under challenging data conditions.