首页|钢结构工程高强度螺栓微小松动视觉检测方法研究

钢结构工程高强度螺栓微小松动视觉检测方法研究

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
基于计算机视觉的螺栓松动检测方法中,一类是通过对连接节点图像进行透视矫正并分析螺栓轮廓边缘直线角度的变化以及时发现微小的松动缺陷.然而,目前这类方法在图像矫正和螺栓轮廓边缘检测的可靠性方面还有较大的改进空间.为此,针对钢结构螺栓连接节点,提出了基于深度学习的螺栓松动视觉检测方法.首先,基于YOLOv5检测螺栓目标.针对螺栓感兴趣区域(RoI),基于鲁棒性较高的Attention U-Net提取螺栓轮廓边缘直线.为了提高螺栓目标检测精度,目标检测模型应设定较低置信度阈值以保证目标无漏检,再通过螺栓RoI提取的边缘直线数量筛除伪螺栓.使用透视变换法对节点图像进行矫正,变换所需的参考点是根据螺栓检测框之间空间移动后的相交关系 自动定位的.最后,根据矫正图像中的螺栓轮廓边缘直线计算螺栓角度,通过检测螺栓和基准螺栓之间的角度差异判断松动情况.研究结果表明:螺栓目标检测的AP值为0.97;螺栓轮廓边缘检测的准确率、召回率和F1的均值分别为0.846、0.807与0.825,且在多种复杂背景干扰下具有较高的鲁棒性;伪螺栓筛除法可过滤掉99.82％的伪螺栓目标;提出的图像矫正法适用于常见的多种螺栓排列形式的连接节点;当松动判别阈值仅为2.8°时,螺栓松动检测的准确率可达99.7％.该方法在大型钢结构螺栓连接节点自动化运维方面具有较好的应用前景.
Visual Detection Method of Small Loosening of High-strength Bolts in Steel Structure Engineering
A method used in computer vision-based bolt loosening detection involves correcting the perspective of connection node images and analyzing changes in the straight-line angles of the bolt contour edges.However,this method still has significant room for improvement in the reliability of image correction and bolt contour edge detection.Therefore,a visual detection method based on deep learning is proposed for detecting bolt loosening in the bolted nodes of steel structures.First,the bolts were detected using YOLOv5.Next,for the bolt region of interest(RoI),the contour edge straight lines of the bolts were extracted using Attention U-Net with high robustness.To improve the bolt object detection accuracy,the object detection model should set a lower confidence threshold to ensure that there are no misdetected objects and then screen out false bolts based on the number of edge straight lines extracted from the bolt RoI.The node image was corrected using the perspective transformation method.The reference points required for the transformation were automatically positioned according to the intersection relationship between the bolt detection boxes after spatial movement.Finally,the bolt angle was calculated based on the straight lines at the edge of the bolt contour in the corrected image,and loosening was determined by detecting the angle difference between the detection and reference bolts.The results show that:the AP value of bolt object detection is 0.97;the mean values of accuracy,recall,and F1 of bolt contour edge detection are 0.846,0.807,and 0.825,respectively,with high robustness under a variety of complex background interferences;the false bolt screening method can filter out 99.82％of the false bolt objects;the proposed image correction method applies to connection nodes with a variety of common bolt arrangements;when the loosening discrimination threshold is only 2.8°,the accuracy of bolt loosening detection is up to 99.7％.This method has good application prospects for the automated operation and maintenance of bolted nodes in large-scale steel structures.

bridge engineeringsmall loosened bolt detectioncomputer vision and deep learningconnection nodeobject detectioncontour edge straight line detectionperspective correction

姚志东、陈志华、刘红波、卢佳祁

展开 >

天津大学建筑工程学院,天津 300072

中冶建筑研究总院(深圳)有限公司,广东深圳 518055

桥梁工程 螺栓微小松动检测 计算机视觉与深度学习 连接节点 目标检测 轮廓边缘直线检测 透视矫正

国家自然科学基金深圳市科技计划项目

52278202JSGG20201102173802006

2024

10.19721/j.cnki.1001-7372.2024.02.011

中国公路学报
中国公路学会

中国公路学报

CSTPCD北大核心
影响因子：1.607
ISSN：1001-7372
年,卷(期)：2024.37(2)
  • 1
  • 31