Bearing State Recognition Method Based on Transfer Learning Under Different Working Conditions.

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Bearing state recognition faces challenges with data reusability, accuracy, and generalization under variable conditions.
• Traditional transfer learning requires expert knowledge for cross-condition feature extraction, limiting practical application.
• Existing methods struggle to adapt models effectively to diverse operational environments.

Purpose of the Study:

• To propose an improved transfer learning method, multi-core balanced distribution adaptation (MBDA), for robust bearing state recognition.
• To enable end-to-end state recognition by eliminating the need for pre-defined cross-condition features.
• To enhance the accuracy and generalization capability of bearing monitoring models.

Main Methods:

• Developed a weighted mixed kernel function to map data from different working conditions into a unified feature space.
• Employed the A-Distance algorithm to estimate distribution and kernel function balance factors for efficient adaptation.
• Integrated a stacked autoencoder (SAE) neural network for feature self-learning and final state classification.

Main Results:

• MBDA effectively reduces distribution differences between datasets from various working conditions.
• The proposed method significantly improves transfer learning performance compared to existing algorithms.
• Accurate identification of bearing states was achieved even under challenging, variable working conditions.

Conclusions:

• MBDA offers a simplified and efficient approach to bearing state recognition in practical applications.
• The method overcomes limitations of traditional transfer learning by not requiring prior expert knowledge.
• MBDA demonstrates superior performance in enhancing model accuracy and generalization for bearing diagnostics.