DSMDTN: A Data-Selective Multiscale Dual Transfer Network for Fault Diagnosis of Key Components in Rotating...

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Traditional deep learning models struggle with bearing fault diagnosis under varying working conditions and noise.
• Reliable fault diagnosis is crucial for rotating machinery maintenance and operational safety.

Purpose of the Study:

• To propose a data-selective multiscale dual transfer network (DSMDTN) for enhanced rolling bearing fault diagnosis.
• To improve diagnostic performance across diverse operational conditions and noisy environments.
• To enable robust transfer learning for bearing fault classification.

Main Methods:

• A data selector (DS) module identifies high-quality source samples using math, entropy, and anomaly scores.
• A multiscale concatenation U-Net (MCU-Net) extracts multiscale, domain-invariant features using gated convolutional blocks.
• A dual classifier (DC) module minimizes distribution discrepancy and classification loss using separate source and target classifiers.

Main Results:

• DSMDTN demonstrated superior accuracy and transfer capability compared to state-of-the-art models.
• Experiments were validated on the CWRU and PT datasets, showing robust performance under various transfer tasks.
• The model effectively handled different noise levels, enhancing diagnostic reliability.

Conclusions:

• The proposed DSMDTN significantly enhances rolling bearing fault diagnosis accuracy and transferability.
• The integration of data selection, multiscale feature extraction, and dual classification provides a robust solution.
• DSMDTN offers a promising approach for intelligent fault diagnosis in challenging industrial environments.