盾构隧道结构病害的图像深度学习识别与演化预测

The development of metro system in China is gradually transformed from the "construction-oriented" to "both construction and operation oriented". The leakages and structural cracks of shield tunnel are becoming a great threat to the operational safety and durability of metros. The “Law of Fives” in preventive maintenance for concrete durability has put much emphasis on the importance and necessity of early identification of structural defects. Aim to ensure the safety and support the rational maintenance for shield tunnel lining structures, this project will carry out the research on the precise and fast image recognition and evolution prediction of structural defects of shield tunnel linings in terms of leakage and cracks. Firstly, by collecting the typical data of the metro tunnel and the indoor experimental analysis, the characteristics of the spatial distribution of the defects and the image of the defects are revealed and the method of image quality evaluation is presented. By combining the method of full convolution network with the conditional random field in deep learning, the precise and fast defect image recognition method is proposed with considering the severe environmental factors in tunnel. The full-scaled load test on concrete segments of shield tunnel and the 3D numerical simulations are carried out to build the evolution model of structural defects and subsequently predict the development of defects. On the basis of recognition and prediction models, the corresponding algorithms will be programmed and compiled into a newly built software. The study can identify the leakage and crack defects of shield tunnel structure with high accuracy and efficiency, greatly reduce the input and misjudgment of on-site inspection, and scientifically serve the maintenance of metro tunnel structure, and provide technical guarantee for operational safety of metro system.

我国地铁建设正由“建设为主”向“建养并重”转变，盾构隧道结构渗漏水和裂缝病害已成为威胁地铁运营安全与耐久性的重要风险源，结构养护中的“五倍定律”更加强调了早期识别诊断病害的重要性和必要性。项目以保障地铁盾构隧道结构安全和科学养护为目标，开展精准快速病害图像识别方法和病害演化预测的研究。首先，通过收集已有地铁隧道典型病害数据和室内实验分析，揭示病害空间分布和病害图像的特征规律，提出图像质量评价方法；再利用深度学习中全卷积网络和条件随机场相结合方式，针对隧道内部的干扰环境，开展病害图像的精准和快速识别方法研究；在室内足尺管片实验基础上，结合数值模拟分析，构建多病害演化模型，预测病害的发展；基于建立的识别方法和演化模型，研发盾构隧道病害识别和演化预测软件。研究可高精准高效率地识别盾构隧道结构的渗漏水和裂缝病害，大大减小现场巡查的投入和误判，科学服务于地铁隧道结构养护，为地铁运营安全提供技术保障。

本项目以保障地铁盾构隧道结构安全和科学养护为目标，针对渗漏水和裂缝高频率的病害，实现了基于深度学习的全卷积网络的隧道结构病害图像识别方法，高精度、高效率地识别盾构隧道复杂环境下病害定量图像信息；研究了多病害对隧道结构安全的影响，预测病害演化发展规律；研发了地铁盾构隧道结构病害识别与演化预测软件系统，实现精准快速识别病害和对病害的演化进行预测。具体内容及相关成果如下：.（1）建立了复杂隧道环境下隧道结构病害规律，并分析了图像影响因素。总结了分析渗漏水与裂缝病害的分布特点和规律，揭示了病害图像特征及光谱特征，研究了不同环境因素对病害图像的影响，提出病害图像清晰度评价方法。.（2）建立了基于全卷积网络的盾构隧道病害图像识别高效算法。 构建了多个数据集组成的隧道结构病害的样本集，提出了基于全卷积网络的图像精准识别方法。改进了端对端的网络结构和分割训练，实现了快速定量化获取渗漏水和裂缝病害信息。.（3）提出了盾构隧道结构多病害演化预测方法。利用管片和纵缝接头的足尺试验，研究了盾构隧道结构典型破坏模式。建立了盾构隧道管片和接头的多尺度模型，揭示了盾构隧道衬砌结构裂缝的演化机理。分析了渗漏水不同位置和状态演化对既有裂缝的演化影响，并以上海某事故盾构隧道为工程背景，验证了建立的裂损病害演化模型。.（4）隧道病害的图像识别与演化预测软件研发与现场检验。研发了自动、快速、精准的渗漏水和裂缝等病害的图像识别与演化预测软件，实现了线阵相机图像的采集、显示、存储记录以及图像文件回放等功能。在上海地铁12号线开展工程应用，并得到了较好的现场验证。.通过本项目研究，现已发表论文22篇，其中中文期刊论文9篇，英文期刊论文12篇；申请发明专利2项，申请软著1项；组织召开机器学习系列国际学术研讨会1次，成立1个韧性基础设施国际联合研究中心，在国际上做了相关研究的5次主旨报告。培养8名博士硕士研究生。项目总体超额完成了预期要求。