基于卷积神经网络的钢结构螺栓松动缺陷识别  

Identification of Bolt Loosening in Steel Structure Based on Convolutional Neural Network

作　　者：李永刚 贾王龙 杨欣悦 林义景 游新 LI Yonggang;JIA Wanglong;YANG Xinyue;LIN Yijing;YOU Xin(MCC22 Group Corporation Ltd,Tangshan 064000,China;College of Civil Engineering,Tongji University,Shanghai 200092,China;College of Civil and Hydraulic Engineering,Qinghai University,Xining 810016,China)

机构地区：[1]中国二十二冶集团有限公司,河北唐山064000 [2]同济大学土木工程学院,上海200092 [3]青海大学土木水利学院,青海西宁810016

出　　处：《土木工程与管理学报》2024年第5期9-17,25,共10页Journal of Civil Engineering and Management

基　　金：国家自然科学基金(51978513,52078359);科技援青合作专项(2024-QY-202)。

摘　　要：本文实现了一种基于YOLO v3目标检测算法的钢结构螺栓快速定位与松动检测方法。参考传统人工扭矩法的标记方法,利用在三维建模软件中构建的合成数据集训练YOLO v3网络模型,训练后的模型在测试集中的表现为:单张图片平均检测耗时0.024 s,平均精确率95.2%。进一步考虑算法在真实试件图片上的表现,检测松动角度的阈值以及各环境因素的影响,进行试验。结果表明算法识别螺栓松动角的阈值为2°,且在不同拍摄距离、透视角度、光照强度、室外背景、镜头模糊度、标记清晰度条件下,算法均取得了符合工程应用条件的检测结果,能够为螺栓松动实时检测提供理论和实践应用基础。A machine learning detection method based on YOLO v3 was developed to quickly locate and detect the loosening of steel structure bolts using pictures.The method combines the traditional idea of manual torque method with the target detection algorithm of deep learning,and trains YOLO v3 network model using synthetic datasets collected in 3D graphical modeling software.The trained model has an average detection time of 0.024 seconds per image and the average accuracy of 95.2%.Considering the algorithm’s performance on images of real test piece,the threshold of bolt loosening detection and the influence of various environmental factors on the detection,experiments were carried out.The results show that the threshold for algorithm recognition of bolt loosening angles is 2 degrees,and the algorithm has achieved detection results that meet engineering application conditions under different conditions such as shooting distances,perspective angles,light intensity,outdoor backgrounds,lens blurriness,and marker legibility.This provides a theoretical and practical foundation for real-time bolt loosening detection on mobile terminals.

关 键 词：螺栓松动检测 人工扭矩法 深度学习 机器视觉 

分 类 号：TU391[建筑科学—结构工程] TP391.41[自动化与计算机技术—计算机应用技术]