基于深度学习和图像识别的RC框架节点表面损伤检测评估研究-《建筑科学与工程学报》

文章信息/Info

Title:: Study on surface damage detection of RC frame joints based on deep learning and image recognition

作者:: 杨铄，冷予冰，许清风; 1. 天津城建大学土木工程学院,天津 300384； 2. 天津城建大学天津市土木建筑结构防护与加固重点实验室，天津 300384； 3. 上海市建筑科学研究院有限公司上海市工程结构安全重点实验室，上海 200032

Author(s):: YANG Shuo, LENG Yubing, XU Qingfeng; 1. School of Civil Engineering, Tianjin Chengjian University, Tianjin 300384, China; 2. Tianjin Key Laboratory of Civil Structure Protection and Reinforcement, Tianjin Chengjian University, Tianjin 300384, China; 3. Shanghai Key Laboratory of Engineering Structure Safety, Shanghai Research Institute of Building Sci. Co., Ltd, Shanghai 200032, China

关键词:: 钢筋混凝土框架节点; 深度学习; 掩码区域卷积神经网络; 表面损伤检测

Keywords:: reinforced concrete frame joint; deep learning; Mask RCNN; surface damage detection

分类号:: -

DOI:: -

文献标志码:: A

摘要:: 为提高钢筋混凝土（RC）结构表面损伤检测效率，实现RC构件表面损伤定位、精细化量测和评估的一体化，针对RC框架结构表面损伤检测场景，提出了基于深度学习和图像识别的损伤检测评估方法。基于掩码区域卷积神经网络（Mask RCNN）模型和图像处理建立RC框架构件表面损伤检测模型；采用以往RC构件试验中的损伤和破坏照片，经裁剪和数据增强后建立包含裂缝、剥落、压碎、露筋、钢筋屈曲和钢筋断裂6种类型损伤的RC构件表面损伤数据集用以训练模型；基于三维重建所生成的构件正立面全景图以及损伤检测模型，编写可同步进行RC构件表面损伤定位和损伤几何参数分析的检测程序，并集成RC构件损伤等级划分和破坏形态判断标准，用以量化评估和判断RC构件损伤程度和破坏形态；设计制作了两个足尺RC梁柱节点试件并进行了试验，使用智能手机和小型无人机对破坏的节点试件梁进行拍照，检测构件表面损伤，以验证上述损伤检测方法的有效性。结果表明：基于智能手机拍照的检测结果好于基于无人机数字变焦拍照的检测结果；因训练样本较少，钢筋屈曲和钢筋断裂的检测精度较差；当镜头距构件表面200 mm左右时，裂缝宽度测量精度可达0.1 mm，若裂缝实测宽度大于0.5 mm，基于上述方法的裂缝宽度检测精度较高；使用检测程序分析得到的构件损伤等级（程度）和破坏形态与观测结果一致；基于深度学习和图像识别的损伤检测方法可为受损或震后RC框架构件表面损伤定量化快速检测和评估提供支撑。

Abstract:: In order to improve the efficiency of surface damage detection of reinforced concrete (RC) structures and to realize the integration of RC component surface damage localization, fine measurement and evaluation, a damage detection and assessment method based on the deep learning and image recognition was proposed for the scenario of detection of surface damage for RC frame structures. The surface damage detection model of RC frame members was established based on the mask region convolutional neural network (Mask RCNN) model and image processing. Based on the images of damage and destruction in previous RC component tests, the surface damage dataset of RC components including six types of damage, i.e. crack, spalling, crushing, rebar exposure, rebar buckling and rebar fracture, was established after cropping and data augmentation to train the model. Based on the 3D reconstruction and the damage detection model, an inspection program, which could simultaneously locate the surface damage of RC components and analyze the damage geometric parameters, was made and integrated the criteria for classifying the damage level and judging the damage pattern of RC components to quantitatively assess the damage degrees and judge the damage patterns for RC components. Two fullscale RC beamcolumn joint specimens were designed, fabricated, and used to conduct an experimental test. The damaged joint specimens were photographed with a smart phone and a small unmanned aerial vehicle (UAV) to detect the surface damage of the beams for validation of the above damage detection method. The results show that the detection results based on the smart phone are better than those based on UAV digital zoom shooting. The accuracy of the rebar buckling and fracture is poor due to the fewer training samples. When the lens is about 200 mm from the surface of the members, the measurement accuracy of crack width can reach 0.1 mm. If the measured crack width is greater than 0.5 mm, the crack width detection accuracy of the above method is higher. The damage grades (degrees) and failure patterns of components analyzed by the detection program are consistent with the observation. The damage detection method based on the deep learning and image recognition can support quantitative surface damage rapid detection and assessment for damaged or postearthquake RC frame components.

《建筑科学与工程学报》[ISSN:1673-2049/CN:61-1442/TU]

文章信息/Info

参考文献/References:

相似文献/References:

备注/Memo

常用功能

导航/Navigate

工具/Tools

统计/Statistics