An Attention EfficientNet-Based Strategy for Bearing Fault Diagnosis under Strong Noise

Area of Science:

• Artificial Intelligence
• Machine Learning
• Industrial Fault Diagnosis

Background:

• Data-driven fault diagnosis methods are gaining traction with AI advancements.
• Industrial environments present inevitable noise, degrading traditional intelligent diagnosis performance.
• Existing methods struggle to meet accuracy requirements in complex, noisy conditions.

Purpose of the Study:

• To propose a novel intelligent diagnosis framework to address performance limitations in noisy industrial settings.
• To enhance the accuracy and stability of fault diagnosis using EfficientNet and an attention mechanism.
• To provide insights into the model's high performance through visualization and analysis.

Main Methods:

• Developed an intelligent diagnosis framework incorporating EfficientNet for optimal performance with limited resources.
• Integrated an attention mechanism to improve the relationship between fault features and modes, boosting accuracy in noise.
• Employed model visualization techniques (weights and features) for performance analysis.

Main Results:

• The proposed framework demonstrated superior accuracy and stability compared to benchmark methods.
• Achieved a diagnostic accuracy of 86.24% under actual working conditions.
• Visualization confirmed the model's ability to effectively identify fault-related features.

Conclusions:

• The developed intelligent diagnosis framework significantly improves fault diagnosis accuracy and stability in noisy industrial environments.
• The integration of EfficientNet and attention mechanisms offers an effective solution for data-driven diagnostics.
• The method provides a robust and explainable approach for real-world industrial applications.