Rolling Bearing Fault Diagnosis Based on Multi-Source Domain Joint Structure Preservation Transfer with Autoencoder.

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Rolling bearing fault diagnosis is crucial for industrial machinery.
• Existing domain adaptation methods struggle with inaccurate data embeddings due to one-way embedding without backward validation.
• This limitation leads to suboptimal diagnostic performance in real-world conditions.

Purpose of the Study:

• To propose a novel rolling bearing fault diagnosis method that overcomes limitations of existing domain adaptation techniques.
• To enhance the accuracy of data embeddings and improve overall diagnostic performance.
• To ensure robustness of the fault diagnosis system across different operating conditions.

Main Methods:

• A multi-source domain joint structure preservation transfer with autoencoder (MJSPTA) method is proposed.
• Similar source domains are screened using inter-domain metrics.
• Data is projected into a shared subspace using different projection matrices, with reconstruction minimizing accuracy. Graph embedding theory preserves local manifold structure, and label propagation with voting determines fault type.

Main Results:

• The MJSPTA method effectively reduces domain differences through distribution matching and sample weighting.
• Preservation of local manifold structure using graph embedding theory enhances adaptation.
• The proposed method demonstrates effectiveness and robustness in diagnostic tests.

Conclusions:

• The MJSPTA method offers a significant advancement in rolling bearing fault diagnosis.
• The approach provides accurate data embeddings and robust diagnostic performance.
• This technique is well-suited for addressing domain shift challenges in fault diagnosis.