Numerical Simulation Data-Aided Domain-Adaptive Generalization Method for Fault Diagnosis.

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Mechanical fault diagnosis faces challenges with cross-domain distribution shifts under varying operating conditions.
• Existing domain-adaptive methods require target data, limiting real-time applications.
• Generalizing fault features from source to unseen target domains is crucial for machinery fault detection.

Purpose of the Study:

• To develop a domain generalization method for mechanical fault diagnosis that overcomes the need for target data.
• To enhance the generalization capability of fault diagnosis models for out-of-distribution data.
• To improve the accuracy and robustness of fault diagnosis systems in diverse operational environments.

Main Methods:

• A finite element model (FEM) generated numerical simulation data as an auxiliary domain.
• Integrated auxiliary domain data with real-world measurement data to create a multi-source domain.
• Employed adversarial training on the multi-source domain to learn domain-invariant features.

Main Results:

• The proposed method demonstrated superior generalization performance compared to baseline methods.
• Achieved an average accuracy improvement of 2.83% for bearing fault diagnosis.
• Achieved an average accuracy improvement of 8.9% for gear fault diagnosis.

Conclusions:

• The developed domain generalization technique effectively addresses cross-domain distribution offsets in mechanical fault diagnosis.
• Integrating simulated and real-world data with adversarial training enhances model generalization for unseen conditions.
• The method offers a viable strategy for real-time, robust machinery fault detection.