Rotating Machinery Fault Diagnosis Method by Combining Time-Frequency Domain Features and CNN Knowledge Transfer

Area of Science:

• Engineering
• Computer Science

Background:

• Rotating machinery generates vast operational data, often with insufficient labeled samples for effective fault diagnosis.
• Traditional fault diagnosis methods struggle with limited labeled data, hindering accurate analysis of complex machinery issues.

Purpose of the Study:

• To propose a novel fault diagnosis method for rotating machinery utilizing deep learning and knowledge transfer.
• To address the challenge of diagnosing faults in the presence of limited labeled data within large operational datasets.

Main Methods:

• Representing operational data as 2D images capturing time and frequency-domain characteristics.
• Transferring a trained source domain model to a shallow model for small-sample training in the target domain.
• Utilizing the shallow model to label a large volume of unlabeled data, creating an augmented dataset.
• Training a deep Convolutional Neural Network (CNN) fault diagnosis model using both original and pseudo-labeled data for knowledge migration.

Main Results:

• The proposed FFCNN-SVM shallow model tagger method demonstrated a significant improvement in fault diagnosis accuracy compared to other transfer learning approaches.
• The developed deep CNN model successfully performed online fault diagnosis for rotating machinery.
• Knowledge transfer from an expert system to the deep CNN architecture was effectively realized.

Conclusions:

• The proposed knowledge transfer method offers a viable solution for fault diagnosis in rotating machinery with limited labeled samples.
• This approach provides innovative strategies for future fault diagnosis research, particularly in scenarios with scarce labeled data.
• The study highlights the potential of deep learning and transfer learning for enhancing machinery health monitoring.