Transfer learning for bearing performance degradation assessment based on deep hierarchical features

Area of Science:

• Mechanical Engineering
• Artificial Intelligence

Background:

• Bearing performance degradation exhibits scattered life cycle distributions due to design, production, and operating conditions.
• Generalizing performance degradation assessment models is challenging, and acquiring labeled data for new conditions is costly and time-consuming.

Purpose of the Study:

• To propose a novel bearing degradation assessment model leveraging transfer learning and deep hierarchical feature extraction.
• To transform degradation assessment into a classification task for distinct degradation patterns (normal, slight fault, fault development, damage).

Main Methods:

• A hierarchical network with random weights extracts local sub-band spectral characteristics via alternating convolution and pooling layers without supervised fine-tuning.
• Joint Geometrical and Statistical Alignment (JGSA) creates a shared feature space for knowledge transfer.
• The model assesses bearing degradation patterns under varying operating conditions.

Main Results:

• The proposed method successfully reduces feature distribution divergence between different degradation processes.
• Experimental results validate the model's effectiveness in bearing fault severity and degradation process assessment.
• The approach enables effective performance degradation assessment across diverse operating conditions.

Conclusions:

• The developed transfer learning and deep hierarchical feature extraction model offers a robust solution for bearing performance degradation assessment.
• This method addresses the challenges of scattered degradation data and the cost of acquiring labeled data for new conditions.
• The model demonstrates significant potential for improving the reliability and efficiency of bearing health monitoring systems.