Transformer-based conditional generative transfer learning network for cross domain fault diagnosis under limited...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Current Generative Adversarial Networks (GANs) struggle to balance data quality and diversity, especially with limited data and varying conditions.
• This limitation hinders effective fault diagnosis, particularly in cross-domain scenarios.

Purpose of the Study:

• To propose a novel Transformer-based conditional GAN transfer learning network (TCGAN) to address the limitations of existing GANs.
• To enhance both the quality and diversity of generated data for improved fault diagnosis.

Main Methods:

• Developed a Transformer-based conditional GAN (TCGAN) generative network incorporating sample label information.
• Established a transfer learning network based on TCGAN.
• Implemented a collaborative 'generation-transfer' training strategy using expectation maximization for parallel parameter updating.

Main Results:

• The proposed TCGAN method demonstrated superior data generation quality compared to TTS-GAN and CorGAN.
• Effectiveness validated on bearing datasets from CWRU and KUST-SY.
• Achieved higher quality generated data, outperforming existing methods.

Conclusions:

• The TCGAN approach offers a significant improvement in generating high-quality and diverse data for fault diagnosis.
• Provides a novel and effective solution for enhancing cross-domain fault diagnosis performance.
• Addresses the critical need for robust data generation in challenging industrial environments.