Method for Diagnosing Bearing Faults in Electromechanical Equipment Based on Improved Prototypical Networks

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Data Science

Background:

• Electromechanical equipment health monitoring faces challenges with complex data, limited datasets, and long-tailed distributions, hindering existing fault diagnosis methods.
• Traditional approaches often overlook crucial characteristics of health monitoring data, leading to suboptimal performance in identifying bearing faults.

Purpose of the Study:

• To propose an improved prototypical network, Weight Prototypical Networks (WPorNet), for enhanced fault diagnosis of rolling bearings in electromechanical systems.
• To address the limitations of small-sample classification and long-tailed data distributions prevalent in health monitoring datasets.

Main Methods:

• Developed Weight Prototypical Networks (WPorNet) by enhancing prototypical networks to account for varying support sample distribution influences, calculated via Kullback-Leibler divergence.
• Utilized the Gramian Angular Field (GAF) algorithm to convert one-dimensional time-series data into two-dimensional vibration images, optimizing the performance of 2D Convolutional Neural Networks (CNNs).

Main Results:

• The WPorNet model demonstrated significant improvements in classification accuracy across various few-shot learning scenarios (e.g., 2-way 10-shot, 4-way 20-shot) on MAFAULDA and CWRU bearing datasets.
• The method effectively addressed data scarcity and long-tailed issues, enhancing sample-data dependency, feature extraction, and overall classification accuracy compared to standard prototypical networks.
• Achieved performance increases of up to 12.02% in few-shot classification tasks.

Conclusions:

• The proposed WPorNet model offers a robust solution for rolling bearing fault diagnosis, particularly in scenarios with limited and imbalanced health monitoring data.
• The enhanced network exhibits superior sample classification accuracy and stronger anti-interference capabilities compared to traditional small-sample classification models.
• This approach advances the field of condition monitoring and fault diagnosis for electromechanical equipment.