A time-frequency collaborative cross-device bearing fault diagnosis model based on supervised transfer learning with...

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning models excel at bearing fault diagnosis under ideal conditions.
• Domain shift in industrial settings significantly degrades model generalization, especially in cross-device tasks.
• Existing transfer learning models struggle with diverse feature sensitivities and limited target data, risking overfitting.

Purpose of the Study:

• To develop a robust cross-device bearing fault diagnosis model overcoming limitations of existing methods.
• To enhance model generalization and stability despite data distribution differences and limited target samples.
• To reduce overfitting and training complexity in transfer learning for bearing fault diagnosis.

Main Methods:

• A time-frequency collaborative modeling paradigm to extract multi-view complementary information from time-domain impact and frequency-domain resonance.
• Introduction of a low-rank adaptation mechanism for lightweight fine-tuning of pretrained models.
• Validation through cross-device diagnosis scenarios using three real-world cases.

Main Results:

• The proposed model effectively captures fault patterns and device-specific characteristics.
• Lightweight fine-tuning preserves source domain knowledge while adapting to target devices, reducing overfitting.
• Experimental results demonstrate superior diagnostic accuracy and generalization compared to mainstream models.

Conclusions:

• The time-frequency collaborative model offers a practical and effective solution for cross-device bearing fault diagnosis.
• The low-rank adaptation mechanism enhances model stability and reduces training complexity.
• The findings highlight the model's potential for real-world industrial applications with limited data.