A New Universal Domain Adaptive Method for Diagnosing Unknown Bearing Faults

Area of Science:

• Machine Learning
• Artificial Intelligence
• Mechanical Engineering

Background:

• Domain adaptation in transfer learning typically assumes shared label spaces, limiting real-world applications.
• Existing models face network failure when encountering unknown fault types in target domains.

Purpose of the Study:

• To propose a universal domain method that overcomes limitations of shared label spaces in transfer learning.
• To improve fault diagnosis accuracy for mechanical systems, particularly with unknown fault types.

Main Methods:

• Developed a framework that extracts fault features and forms feature centers for each fault type.
• Incorporated three optimization functions to mitigate negative transfer issues with unknown fault types.
• Validated the framework's performance on multiple datasets under variable speed conditions.

Main Results:

• The proposed universal domain method demonstrated superior fault diagnosis performance compared to existing transfer methods.
• Effectively reduced network failures caused by unknown fault types in the target domain.
• Showcased performance advantages for variable speed fault diagnosis.

Conclusions:

• The novel framework successfully breaks the assumption of shared label spaces in domain adaptation.
• The method offers a robust solution for mechanical fault diagnosis, especially in scenarios with unknown fault types and variable operating conditions.