Welding technology is essential in industrial manufacturing, but traditional methods for seam recognition, real-time positioning, and defect detection rely on manual expertise, resulting in low efficiency and accuracy. This research focuses on vision sensing technologies for pre-welding, in-welding, and post-welding stages, proposing corresponding solutions and verifying their feasibility in improving welding quality and efficiency. Firstly, to address the need for weld seam recognition before welding, an algorithm based on area structured light sensors was developed for weld seam position recognition. By obtaining 3D point cloud from the workpiece surface and employing point cloud preprocessing and segmentation algorithms, an automatic method for recognizing weld seam types and calculating their positions was proposed. This method effectively distinguishes various weld seam types, including I-type butt welds, V-type butt welds, simple fillet welds, and complex fillet welds, and accurately calculating weld positions and start and end point coordinates. Secondly, a real-time recognition sensor for welding seam positions during the welding process based on line laser was designed. By optimizing the laser stripe images, the accuracy and continuity of weld seam point cloud data were improved. Real-time recognition algorithms suitable for various weld seam types, including I-type butt welds, V-type butt welds, and fillet welds, were developed. This sensor can calculate welding seam coordinates in real time, significantly improving welding accuracy and stability. Finally, to address the limitations of traditional post-welding surface defect detection and dimension evaluation, this research proposed a method combining deep learning with 3D point cloud processing. To solve the issue of high reflectivity of weld surfaces, the structured light encoding of the area structured light sensor was optimized, and algorithms were employed to supplement the missing point cloud data. By optimizing the YOLOv7 model structure and employing data augmentation methods, weld surface defects were efficiently identified. Additionally, defect dimension was accurately calculated using point cloud data from the defect areas. This study systematically addresses critical challenges in weld seam recognition, tracking, and surface defect detection, improving efficiency and accuracy. The proposed methods provide valuable insights and establish a robust foundation for future advancements in intelligent welding systems.
| Date of Award | 14 Jul 2025 |
|---|
| Original language | English |
|---|
| Supervisor | Wenhan Zeng (Main Supervisor) |
|---|