Partial transfer learning in machinery cross-domain fault diagnostics using class-weighted adversarial networks

Area of Science:

• Machine learning
• Artificial intelligence
• Mechanical engineering

Background:

• Transfer learning (TL) is increasingly used for machinery fault diagnosis.
• Existing TL methods often assume identical label spaces, which is unrealistic in industrial settings.
• Real-world scenarios frequently involve target domains with a subset of source domain labels.

Purpose of the Study:

• To address the challenge of partial transfer learning in machinery fault diagnosis.
• To develop a deep learning-based domain adaptation method for scenarios with limited target domain labels.
• To enable knowledge transfer from comprehensive source domains to target domains with fewer machine conditions.

Main Methods:

• A class-weighted adversarial neural network was proposed for deep learning-based domain adaptation.
• The method encourages positive knowledge transfer from shared classes between domains.
• It effectively mitigates the impact of source domain outliers irrelevant to the target domain.

Main Results:

• Experimental validation was conducted on two rotating machinery datasets.
• The proposed method demonstrated promising performance in partial transfer learning scenarios.
• Positive transfer of diagnostic knowledge was achieved despite label space differences.

Conclusions:

• The developed class-weighted adversarial neural network is effective for partial transfer learning in machinery fault diagnosis.
• The method offers a viable solution for industrial applications where label spaces differ.
• This approach advances the capabilities of intelligent fault diagnosis systems.