A Gated Recurrent Generative Transfer Learning Network for Fault Diagnostics Considering Imbalanced Data and Variable...

Area of Science:

• Machine Learning
• Signal Processing
• Engineering

Background:

• Intelligent fault diagnosis often faces challenges with imbalanced data and varying working conditions.
• Existing generative adversarial networks (GANs) for data synthesis overlook critical fault-sensitive temporal features.

Purpose of the Study:

• To propose a novel gated recurrent generative transfer learning network (GRGTLN) for enhanced intelligent fault diagnosis.
• To address the loss of temporal information in synthetic data generated by current methods.

Main Methods:

• A smooth conditional matrix-based gated recurrent generator was developed to extend imbalanced datasets, focusing on fault-sensitive features.
• Wasserstein distance (WD) was incorporated to improve data generation and transfer learning performance.
• An iterative 'generation-transfer' co-training strategy was employed for continuous model training and parameter optimization.

Main Results:

• The proposed GRGTLN effectively generates high-quality synthetic data that retains crucial temporal information.
• The method demonstrated satisfactory cross-domain diagnosis accuracy in comprehensive case studies.
• Attention to fault-sensitive features in the generated sequences was adaptively increased.

Conclusions:

• GRGTLN offers a significant advancement in intelligent fault diagnosis, particularly for imbalanced datasets and diverse operating conditions.
• The network's ability to preserve temporal information leads to more reliable and accurate fault diagnosis.
• This approach enhances the transfer learning capability for fault diagnosis models.