Transfer Learning Based Fault Diagnosis with Missing Data Due to Multi-Rate Sampling.

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Deep learning excels at feature extraction for fault diagnosis using multi-sensor data.
• Disparate sensor sampling rates create incomplete datasets, limiting deep neural network (DNN) input.
• Sufficient data is crucial for efficient deep learning-based fault diagnosis.

Purpose of the Study:

• To propose a fault diagnosis framework for handling missing data in multi-sensor systems.
• To leverage transfer learning to utilize structurally incomplete samples effectively.
• To improve fault diagnosis accuracy by extracting additional fault features.

Main Methods:

• Developed a fault diagnosis framework based on transfer learning for missing data.
• Designed transfer learning mechanisms to extract supplementary fault features.
• Applied the framework to multi-rate sampled data from a rotating machinery vibration experimental platform.

Main Results:

• The proposed framework effectively utilizes structurally incomplete samples.
• Transfer learning enhances the extraction of useful fault features.
• Improved accuracy in fault diagnosis compared to methods using only complete samples.

Conclusions:

• The developed framework enables online and offline fault diagnosis for multi-rate sampling data.
• The method demonstrates efficiency and improved accuracy in fault diagnosis.
• This approach addresses the challenge of limited complete data in deep learning for fault diagnosis.