Numerical Model Driving Multi-Domain Information Transfer Method for Bearing Fault Diagnosis.

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Rolling bearings are critical components in machinery, but their fault diagnosis is challenging due to complex operating conditions and a scarcity of fault samples.
• Existing methods often struggle with varying working conditions and insufficient fault data, hindering reliable fault diagnosis.

Purpose of the Study:

• To propose a numerical model-driven cross-domain fault diagnosis method for rolling bearings operating under variable working conditions with limited fault samples.
• To enhance the robustness and accuracy of fault diagnosis by leveraging simulation data and advanced machine learning techniques.

Main Methods:

• Construction of simulation datasets encompassing multiple fault types under variable working conditions to address incomplete fault samples.
• Expansion of simulation datasets using generative adversarial networks (GANs) to ensure sufficient data for model training.
• Application of cross-Domain Nuisance Attribute Projection (cDNAP) to derive a projection matrix that mitigates feature distribution discrepancies between measured and simulated data.

Main Results:

• The proposed cross-domain fault diagnosis method achieved a diagnostic accuracy of up to 99% in experiments involving variable working conditions.
• The method demonstrated superior performance compared to established domain adversarial neural networks, including DANN, DSAN, and DAAN.

Conclusions:

• The developed numerical model-driven cross-domain fault diagnosis approach effectively handles the challenges of varying working conditions and scarce fault samples in rolling bearing diagnostics.
• The cDNAP technique proves instrumental in aligning feature distributions across domains, leading to significant improvements in diagnostic accuracy.