Source-Free Domain Adaptation Framework for Rotary Machine Fault Diagnosis.

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Intelligent fault diagnosis for rotary machinery faces performance degradation due to domain shifts.
• Existing methods often lack robustness when applied to new, unseen operational environments.

Purpose of the Study:

• To develop a robust fault diagnosis framework that effectively addresses domain shifts in rotary machinery.
• To enhance the adaptability and performance of fault diagnosis systems in practical, real-world scenarios.

Main Methods:

• Implemented an order-frequency-based preprocessing method to normalize rotational variations.
• Utilized a U-Net variational autoencoder (U-NetVAE) for adaptation via reconstruction learning.
• Employed a test-time training (TTT) strategy for unsupervised target domain adaptation.

Main Results:

• The proposed framework significantly outperformed conventional machine learning and deep learning models in F1-score and recall across domains.
• Achieved an F1-score of 0.47 and recall of 0.51 in the target domain under challenging conditions.
• Ablation studies validated the effectiveness of each component in improving adaptation performance.

Conclusions:

• The developed framework demonstrates superior robustness and adaptability for intelligent fault diagnosis under domain shifts.
• Combining mechanical priors, self-supervised learning, and lightweight adaptation strategies is effective for practical fault diagnosis.
• The approach offers a promising solution for reliable fault detection in diverse operational environments.