Bearing-DETR: A Lightweight Deep Learning Model for Bearing Defect Detection Based on RT-DETR

Area of Science:

• Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Accurate bearing defect detection is crucial for industrial safety and operational efficiency.
• Existing methods may lack the necessary precision or computational efficiency for real-world deployment.
• The need for lightweight, high-performance models for industrial monitoring is growing.

Purpose of the Study:

• To introduce Bearing-DETR, a novel deep learning model for enhanced bearing defect detection.
• To optimize the Real-Time Detection Transformer (RT-DETR) architecture for improved performance and efficiency.
• To validate the model's effectiveness on real-world industrial data.

Main Methods:

• Utilized the Real-Time Detection Transformer (RT-DETR) architecture.
• Incorporated Dysample Dynamic Upsampling, Efficient Model Optimization (EMO) with Meta-Mobile Blocks (MMB), and Deformable Large Kernel Attention (D-LKA).
• Trained and validated the model on a dataset from a chemical plant environment.

Main Results:

• Bearing-DETR achieved a mean average precision (mAP) of 94.3% (IoU = 0.5) and 57.5% (IoU = 0.5-0.95).
• The model demonstrated significant improvements over the standard RT-DETR.
• Reduced computational demands with 8.2 G floating-point operations (FLOPs) and 3.2 M parameters.

Conclusions:

• Bearing-DETR offers a highly accurate and efficient solution for bearing defect detection.
• The lightweight design makes it suitable for low-resource industrial devices.
• Potential to revolutionize maintenance strategies, quality control, and sustainability in manufacturing.