Related Experiment Video
Updated: Jun 3, 2025
00:05
Crack Monitoring in Resonance Fatigue Testing of Welded Specimens Using Digital Image Correlation
Published on: September 29, 2019
8.2K
Artificial-Weld-Crack Detection Network, YOLOv6-NW, Based on Target Recognition Technology.
Yiming Wang1, Lunhua Shang2, Bin Li1
1Luoyang Institute of Science and Technology, Luoyang 471023, China.
Materials (Basel, Switzerland)
|January 8, 2025
Summary
This study introduces a novel method for preparing artificial weld cracks to improve datasets for defect detection. A lightweight YOLOv6-NW model achieves high accuracy with reduced parameters, enhancing weld crack identification.
Area of Science:
- Materials Science
- Computer Vision
- Artificial Intelligence
Background:
- Scarce datasets and low accuracy hinder effective weld crack detection.
- Industrial applications require efficient and resource-conscious defect identification models.
Purpose of the Study:
- To address weld crack detection challenges by augmenting datasets and developing a lightweight, accurate detection model.
- To improve the sample size and richness for weld crack defect detection models.
- To design an efficient weld crack detection network suitable for industrial constraints.
Main Methods:
- Developed an artificial weld crack preparation method using dissimilar metal particle doping.
- Applied data augmentation techniques including random cropping, scaling, and Mosaic.
- Designed and optimized the YOLOv6-N model for width and depth compression, creating YOLOv6-NW.
Main Results:
- YOLOv6-NW demonstrated superior performance and smaller model size compared to YOLOv5.
- YOLOv6-NW achieved comparable accuracy and recall to YOLOv6-N with only 16% of its parameters.
- Maintained high crack detection precision (above 0.9) under low-resolution and low-illumination conditions.
Conclusions:
- The proposed artificial weld crack preparation and data augmentation effectively support defect detection models.
- YOLOv6-NW offers an efficient and accurate solution for weld crack detection, balancing performance and resource usage.
- The model shows robustness in challenging industrial environments with varying image quality.
