A Novel Joint Adversarial Domain Adaptation Method for Rotary Machine Fault Diagnosis under Different Working...

Area of Science:

• Engineering
• Computer Science
• Machine Learning

Background:

• Real-world fault detection faces challenges due to domain shift caused by changing working conditions, equipment wear, and environmental factors.
• Existing deep learning algorithms exhibit poor performance when applied to domains different from their training data.
• This necessitates robust methods for fault diagnosis that can adapt to varying operational environments.

Purpose of the Study:

• To propose a novel fault diagnosis method, Joint Adversarial Domain Adaptation (JADA), designed to address the performance degradation of deep networks under different working conditions.
• To develop a domain adaptation technique that creates robust and invariant feature representations for fault detection.
• To enhance the discriminative power of learned features and prevent model collapse.

Main Methods:

• JADA employs a unified adversarial learning process to simultaneously align marginal and conditional distributions between source and target domains.
• The method constructs domain-invariant and category-discriminative feature representations.
• A center loss supervision signal is incorporated to penalize distances between deep features and their class centers, improving feature structure and preventing mode collapse.

Main Results:

• JADA was evaluated on twenty-four transfer fault diagnosis tasks across two experimental platforms.
• The proposed method demonstrated significant performance improvements compared to several popular domain adaptation techniques.
• Experimental results verified the effectiveness and robustness of JADA in handling substantial distribution differences.

Conclusions:

• Joint Adversarial Domain Adaptation (JADA) offers a significant advancement in fault detection under varying working conditions.
• The method's ability to align distributions and enhance feature discriminability leads to superior performance in transfer fault diagnosis tasks.
• JADA provides a robust solution for real-world fault detection systems requiring adaptability to environmental and operational changes.