Weld Feature Point Recognition Method for Large-Diameter Penstocks Based on Improved YOLOv8
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Aiming at the low segmentation accuracy of weld laser stripes caused by arc noise and spatter interference in actual welding environments, as well as the problems of dense smalltarget defects, weak feature information, and high positioning difficulty in welds of large-diameter penstocks, this paper proposes a multi-dimensional improvement of the YOLOv8 model and presents a weld feature point recognition and positioning method. Firstly, an image acquisition system adapted to complex industrial on-site conditions is designed, which effectively suppresses environmental interference through the collaborative design of anti-arc lenses, adjustable light sources, and high-speed cameras. Secondly, the traditional Shuffle Attention (SA) mechanism is abandoned, and the Coordinate Attention (CA) mechanism is introduced to embed into the model's backbone network and detection head pathway, enhancing the model's ability to capture and recall features of small defects such as micro pores and cracks. Thirdly, through the bidirectional connection and learnable weighted fusion mechanism of BiFPN, more efficient and accurate cross-scale feature interaction is realized. Finally, to address the insufficient accuracy of small-target detection, the Slicing Aided Hyper Inference (SAHI) framework is integrated to improve the detection accuracy of dense small targets through dynamic slicing and weighted fusion strategies. Experimental results show that the mean Average Precision (mAP@0.5) of the improved model on the self-built large-diameter penstock weld dataset reaches 92.8%, which is 5.6% higher than that of the original YOLOv8 model, and the inference speed is maintained at 32 FPS, meeting the requirements of industrial real-time detection and providing reliable technical support for the autonomous operation of welding robots.
Fields of Study & Subjects
Industrial Vision Systems and Defect Detection
Advanced Neural Network Applications
Infrastructure Maintenance and Monitoring
Publication details
- Journal / source
- International Conference on Smart City and Information System (ICSCIS)
- Pages
- 74–79
- Publication date
- 30 January 2026
- Language
- EN