基于深度学习的X射线燃料棒端塞缺陷自动检测方法研究  

Detection Method of X-ray Fuel Rod End Plug Defect Based on Deep Learning

作　　者：张小刚 俞东宝[1] 汤慧[1] 朱永利 ZHANG Xiaogang;YU Dongbao;TANG Hui;ZHU Yongli(China North Nuclear Fuel Co.,Ltd.,Baotou 014035,China)

机构地区：[1]中核北方核燃料元件有限公司,内蒙古包头014035

出　　处：《原子能科学技术》2024年第8期1767-1776,共10页Atomic Energy Science and Technology

摘　　要：为了提高深度学习在X射线燃料棒端塞缺陷检测中的准确性,实现更高精度的无损检测,本文基于YOLOX的目标检测模型,针对该场景下目标缺陷尺寸极小的特点,对网络结构和损失函数进行了相应的改进,并在工业数据集上进行了验证。结果表明,该算法方案在保持较高识别速度的同时,识别精度获得了明显的提升,达到生产检测要求。该研究方法为今后燃料棒端塞焊缝X射线数字检测图像的高精度自动分析评价打下了坚实的基础。Amidst the global expansion of nuclear power generation,ensuring the integrity of nuclear fuel rods is crucial for the safe operation of nuclear power plants.As a vital component of fuel rods,the detection of defects in the end plugs is a key step in ensuring nuclear safety.Traditional manual detection methods are not only time-consuming and inefficient but also susceptible to subjective influences.To address these issues,this study proposed an automatic detection method for defects in fuel rod end plugs based on deep learning X-ray imaging,aiming to enhance the accuracy and efficiency of detection.The research began by collecting a large number of X-ray images of fuel rod end plugs and preprocessing these images,including single-rod segmentation and extraction of effective evaluation areas,to optimize image quality.Subsequently,an improved YOLOX model was adopted as the core detection algorithm,with adjustments made to the network structure and loss function to address the characteristics of small target defects.The introduction of a coordinate attention module enables the model to more accurately locate and identify tiny defects.Additionally,the CIoU loss function was employed in place of the traditional IoU loss function to improve the model’s localization precision for small targets.During the model training phase,data augmentation techniques such as Mosaic,Copy and Paste,and Mixup were implemented to enhance the model’s adaptability to new scenarios.The experimental results demonstrate that the improved model excels in the task of end plug defect detection,with significant enhancements in detection accuracy and speed compared to traditional methods and unimproved deep learning models.Tests on industrial datasets show a notable increase in the model’s mean average precision(mAP)while maintaining a fast detection speed,meeting the requirements of actual production.The model also performs well in detecting various types of defects,including accurate identification of porosity,swelling,incomplete welding,t

关 键 词：燃料棒 焊缝探伤 缺陷检测 深度学习 X射线 

分 类 号：TL99[核科学技术—核技术及应用] TP181[自动化与计算机技术—控制理论与控制工程]