摘要
大多数隧道在建成初期都存在钢筋不连续、内部脱空和混凝土浇筑不密实等衬砌内部缺陷,严重影响隧道结构的耐久性和稳定性。为解决隧道衬砌内部缺陷常规检测手段的主观性强、准确性差、效率低等问题,基于YOLOv7目标检测算法,提出一种改进的YOLOv7隧道衬砌内部缺陷检测算法,将主干特征提取网络输出的最深层特征层输入视觉显示中心(EVC),使其更加关注层内的细节信息,同时将边界框回归误差替换成基于最小点距离的新型损失函数(MPDIoU)。使用有限差分时域方法获取模拟雷达图像,与真实雷达图像一起构成隧道衬砌缺陷雷达图像数据集,进而将改进YOLOv7算法与YOLOv8、YOLOv7、YOLOv5、SSD和Faster RCNN共5种算法进行模型对比实验,在实际隧道质量检测中验证提出方法的优越性和有效性。模型对比实验中对于钢筋、内部脱空以及衬砌不密实这3种缺陷的识别,改进YOLOv7算法的F 1分数分别为94.51%、84.53%和97.66%,平均精度分别为97.13%、83.78%和98.30%,并且相较于其他5种模型在3种缺陷信号的均值F 1值和均值平均精度上均有一定提升;实例验证中,脱空信号和注浆不密实信号的识别准确率分别为70%和75%。结果表明改进的YOLOv7隧道衬砌内部缺陷检测算法具有优越的综合检测性能和泛化能力,有效满足了隧道二衬缺陷的检测需求。
Most tunnels have internal lining defects such as discontinuous reinforcement,internal voids and incomplete concrete pouring at the early stage of construction completion,which seriously affect the durability and stability of the tunnel structure.In order to solve the problems of strong subjectivity,poor accuracy and low efficiency of the conventional means of detecting internal defects oftunnel lining,based on the YOLOv7 target detection algorithm,an improved YOLOv7 tunnel lining internal defect detection algorithm was proposed,which input the deepest feature layer output from the backbone feature extraction network into an explicit visual center(EVC)to pay more attention to the detailed information within the layer while replacing the bounding box regression error with a novel loss function based on minimum point distance(MPDIoU).A finite-difference time-domain method was used to obtain simulated radar images,together with the real radar images collected in the tunnel,to constitute a radar image dataset of tunnel lining defects.Furthermore,experiments were conducted to compare the improved YOLOv7 algorithm with YOLOv8,YOLOv7,YOLOv5,SSD,and Faster RCNN,to verify the superiority and effectiveness of the proposed method in real tunnel quality inspection.The experimental results show that for the recognition of three types of defect signals,namely rebar signals,voiding signals,and lining incompactness signals,the improved YOLOv7 algorithm achieves F 1 scores of 94.51%,84.53%,and 97.66%,with average precision scores of 97.13%,83.78%,and 98.30%,respectively.Compared to the other five models,the improved YOLOv7 algorithm demonstrates certain improvements in the mean F 1 scores and mean average precision across the three types of defect signals.In the instance verification,the accuracy rates for voiding signals and grouting incompactness signals are 70%and 75%,respectively.The results indicate that the improved YOLOv7 algorithm for detecting internal defects in tunnel lining structures possesses superior overall detection performance and generalization capabilities,effectively meeting the detection requirements for tunnel secondary lining defects.
作者
周中
周诗荣
李世帅
鲁四平
ZHOU Zhong;ZHOU Shirong;LI Shishuai;LU Siping(School of Civil Engineering,Central South University,Changsha 410075,China)
出处
《铁道学报》
北大核心
2025年第9期201-211,共11页
Journal of the China Railway Society
基金
国家自然科学基金(52478426)
湖南省自然科学基金(2024JJ5428)
长沙理工大学公路养护技术国家工程研究中心开放基金(kfj220101)。
关键词
隧道工程
衬砌内部缺陷
目标检测
深度学习
tunneling engineering
internal defects in tunnel lining
target detection
deep learning
作者简介
第一作者:周中(1978-),男,河南驻马店人,教授,博士。E-mail:369144091@qq.com;通信作者:鲁四平(1973-),男,湖南长沙人,讲师,博士。E-mail:siping1018@126.com。
