A Bearing Fault Diagnosis Method Based on Weighted Differential Time-Frequency Features and a Dual-Branch Interactive...

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Existing bearing fault diagnosis methods struggle with single input features, limiting performance under variable conditions and few-shot learning.
• Insufficient interaction and fusion between local fault impulse features and contextual relationships hinder diagnostic accuracy.

Purpose of the Study:

• To propose an advanced bearing fault diagnosis method addressing limitations of current techniques.
• To enhance the characterization of fault information by integrating local and contextual features.

Main Methods:

• Developed a novel weighted differential time-frequency features (WDF) input by enhancing transient information and extracting frequency-domain features.
• Designed a dual-branch network incorporating a multi-scale wide-kernel deep convolutional neural network (MS-WDCNN) for local features and a Swin Transformer for contextual relationships.
• Introduced a Feature Interaction Module (FIM) for bidirectional fusion of features from both branches.

Main Results:

• Achieved high average accuracies of 99.45% (CWRU) and 96.12% (HUST) under variable operating conditions.
• Demonstrated strong performance in few-shot learning, reaching 89.61% (CWRU) and 74.73% (HUST) accuracy with only five training samples per class.
• Validated the method's effectiveness on public bearing datasets (CWRU and HUST).

Conclusions:

• The proposed WDF and dual-branch interactive fusion network effectively diagnoses bearing faults under challenging conditions.
• The method shows significant promise for real-world applications requiring robust fault diagnosis with limited data.
• Integration of local impulse features and contextual relationships via FIM enhances diagnostic discriminability.