Cycle-consistent Adversarial Adaptation Network and its application to machine fault diagnosis.

Area of Science:

• Machine learning
• Artificial intelligence
• Industrial engineering

Background:

• Deep learning models achieve success in machine fault diagnosis, driven by industrial big data and intelligent manufacturing.
• Current models face challenges due to domain discrepancy, limiting their performance on different datasets.
• Adversarial domain adaptation methods show promise but often fail to guarantee sufficient feature similarity.

Purpose of the Study:

• To develop a novel approach for effective machine fault diagnosis that overcomes domain discrepancy.
• To enhance the transferability and reliability of deep learning models in industrial settings.
• To ensure domain-invariant and class-separate feature learning for improved diagnostic accuracy.

Main Methods:

• Introduction of a Cycle-consistent Adversarial Adaptation Network (CAAN) for machinery fault diagnosis.
• Utilizing an adversarial game between a feature extractor and a domain discriminator for transferable feature learning.
• Implementing feature translators and discriminators with a cycle-consistent generative adversarial constraint to ensure domain-invariant and class-separate features.

Main Results:

• The proposed CAAN effectively addresses the domain discrepancy issue in machine fault diagnosis.
• Experiments on three diverse datasets demonstrate the superiority of CAAN over existing methods.
• The network ensures more reliable domain-invariant and class-separate feature characteristics.

Conclusions:

• CAAN offers a more effective solution for machinery fault diagnosis in the presence of domain shifts.
• The cycle-consistent adversarial adaptation approach enhances the robustness and generalizability of diagnostic models.
• This method holds significant potential for improving intelligent manufacturing and industrial big data applications.