Feature decoupling integrated domain generalization network for bearing fault diagnosis under unknown operating...

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Real-world engineering faces challenges with varying operating conditions, causing data distribution shifts between training and actual fault data.
• These distribution shifts often lead to the failure of deep learning diagnostic models in mechanical equipment.
• Developing models that generalize diagnostic knowledge from known (source) domains to unseen (target) domains is crucial.

Purpose of the Study:

• To propose a domain generalization network, the Feature Decoupled Integrated Domain Generalization Network (FDIDG), for diagnosing bearing faults under unknown operating conditions.
• To address the limitations of deep learning models when faced with distributional differences between training and real-world fault data.
• To enhance the robustness and accuracy of diagnostic systems in dynamic industrial environments.

Main Methods:

• A 'feature decoupling' algorithm was developed to extract generalized fault feature representations from multiple source domains.
• This algorithm shrinks data distributions across source domains, generalizing fault features to reduce their coupling with operating conditions.
• A multi-expert integration strategy and domain-private features were employed to improve diagnostic accuracy and mitigate the impact of edge samples.

Main Results:

• The proposed FDIDG network demonstrated excellent generalization capabilities in cross-domain experiments.
• Experiments were conducted on both public and private datasets, validating the model's performance under diverse conditions.
• The feature decoupling and multi-expert integration effectively improved diagnostic accuracy for unknown operating conditions.

Conclusions:

• The FDIDG network offers a robust solution for bearing fault diagnosis in scenarios with unknown operating conditions.
• The method successfully extracts generalized diagnostic knowledge, overcoming data distribution discrepancies.
• FDIDG shows significant promise for improving the reliability of AI-driven diagnostics in complex engineering applications.