Research on Bearing Surface Scratch Detection Based on Improved YOLOV5

Area of Science:

• Mechanical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Manual inspection of bearings is time-consuming and cannot meet industrial demands.
• Automated scratch detection is crucial for ensuring machinery reliability and high pass rates.
• Existing methods lack the speed and accuracy required for real-time industrial applications.

Purpose of the Study:

• To develop an advanced AI model for accurate and efficient detection of bearing surface scratches.
• To improve upon existing object detection networks for industrial defect identification.
• To establish a machine vision system for automated bearing surface inspection.

Main Methods:

• An improved YOLOV5 network, termed YOLOV5-CDG, was developed.
• YOLOV5-CDG integrates Coordinate Attention (CA) mechanism, Deformable Convolutional Networks (DCNs), and GhostNet.
• A machine vision system and a custom dataset were utilized for training and validation.

Main Results:

• The YOLOV5-CDG model achieved a 97% Average Precision (AP) for scratch detection, surpassing YOLOV5 by 3.4%.
• The model demonstrated 99.46% accuracy in classifying defective and qualified products.
• Detection times were 263.4 ms on CPU and 12.2 ms on GPU per image.

Conclusions:

• The proposed YOLOV5-CDG model offers superior speed and accuracy for bearing surface scratch detection.
• The developed system meets industrial requirements for automated inspection of bearing defects.
• This research contributes to enhancing quality control in machinery and equipment manufacturing.