YOLO-Weld: A Modified YOLOv5-Based Weld Feature Detection Network for Extreme Weld Noise.

Area of Science:

• Robotics and Automation
• Computer Vision
• Manufacturing Technology

Background:

• Accurate weld feature point detection is crucial for welding trajectory planning and tracking.
• Existing methods struggle with performance degradation in high-noise welding environments.
• Conventional convolutional neural network (CNN) approaches face limitations under extreme noise.

Purpose of the Study:

• To develop an advanced feature point detection network for robust weld localization in challenging, high-noise industrial settings.
• To improve the speed, accuracy, and robustness of weld feature point detection compared to existing methodologies.
• To address the performance bottlenecks of current methods in extreme welding noise conditions.

Main Methods:

• Proposed YOLO-Weld network based on an improved You Only Look Once version 5 (YOLOv5).
• Incorporated the reparameterized convolutional neural network (RepVGG) module for optimized network structure and enhanced detection speed.
• Utilized a normalization-based attention module (NAM) to improve feature point perception and a lightweight decoupled head (RD-Head) for classification and regression accuracy.
• Developed a welding noise generation method to increase model robustness in extreme noise environments.

Main Results:

• The YOLO-Weld model demonstrated superior performance compared to two-stage detection methods and conventional CNN approaches on a custom dataset.
• Achieved an average feature point detection error of 2.100 pixels in images and 0.114 mm in the world coordinate system.
• The model meets real-time welding requirements while accurately detecting feature points in high-noise environments.

Conclusions:

• The proposed YOLO-Weld network effectively addresses the limitations of existing methods in noisy welding conditions.
• The integration of RepVGG, NAM, and RD-Head significantly enhances detection speed and accuracy.
• The model provides a robust and accurate solution for feature point detection in practical welding tasks, meeting stringent accuracy demands.