外观病害检测和服役性能评估是确保既有钢筋混凝土桥梁安全运营的重要途径。本论文关注基于裂缝指标的桥梁智能检测评定，聚焦“测”和“评”这两方面问题，以“高效而准确的裂缝检测”与“定量的服役性能评估”为总体目标。从性能评估定量化、裂缝检测准确化和图像获取自动化三方面进行探索，取得的主要研究成果如下：（1）提出了采用点云分割与逆投影提取图像ROI的桥梁裂缝识别方法。建立了典型RC桥梁的三维点云数据集，用于训练基于深度学习框架的大规模桥梁三维点云语义分割网络RandLA-BridgeNet。根据三维点云至二维图像的逆向投影原理编写了图像ROI提取程序，可将无人机桥检图像处理成仅含ROI的图像。采用基于深度卷积神经网络的网格分类-方框检测混合模型对提取ROI后的图像进行裂缝识别，良好解决了背景误识别问题。开展真实公路桥梁的无人机检测试验，验证了该方法的有效性。（2）建立了基于裂缝指标、可服务于RC受弯构件定量化服役性能评估的纤维梁模型修正方法。采用典型的RC梁板算例，对该方法进行了验证。（3）开展了既有RC空心板梁桥的试验研究，应用和验证提出的裂缝识别方法以及模型修正和服役性能评估方法。以某即将拆除重建的RC空心板梁桥为对象，开展了外观病害检测、荷载横向分布试验、原桥位和试验室内的破坏试验以及截面构造和材料性能测定试验。试验揭示了空心板梁桥的荷载横向分布性能，以及空心板梁在单轴单调和滞回荷载下的全过程受力性能。（4）开展了既有RC空心板梁桥的模型修正和服役性能评估。提出了考虑铰缝和主梁损伤的改进的铰接板梁法，根据荷载横向分布性能试验的结果推算出各板梁和铰缝的实际刚度。结合试验数据，评估了空心板梁的刚度和承载力，以及国内外典型规范的计算方法适用性。此外，对空心板梁进行纤维梁模型修正，反演材料性能参数并准确复现梁的全过程力学响应。（5）组建了桥梁表面裂缝全检任务下的无人机路径规划算法。通过一个桥梁概念模型的路径规划算例，验证了这套方法的可行性。本论文获得国家自然科学基金项目（52192662，51978376）的资助。
Appearance defect detection and service performance evaluation are important ways to ensure the safe operation of existing reinforced concrete (RC) bridges. This dissertation focuses on the intelligent bridge inspection and evaluation based on crack indicator, concentrating on the two aspects of "measurement" and "evaluation", taking efficient and accurate crack detection and quantitative service performance evaluation as the overall goal. Investigation efforts are made from the aspects of quantitative performance evaluation, accurate crack detection and automatic image acquisition, and the main contributions are as follows:(1) A bridge crack detection method based on image region of interest (ROI) extraction using point cloud segmentation and inverse projection is proposed. A 3D point cloud dataset of typical RC bridges is established to train RandLA-BridgeNet, a semantic segmentation network for large-scale bridge point clouds based on deep learning framework. According to the 3D-to-2D projection principle, an image ROI extraction program is written, which can process the bridge images taken by an unmanned aerial vehicle (UAV) into images containing only the ROI. A deep learning convolutional neural network (CNN) called the grid-based classification and box-based detection fusion model (GCBD) is utilized to identify cracks in the processed images, and the background misrecognition problem is solved. An UAV-based inspection experiment was conducted on a real highway bridge to validate the proposed method.(2) A fiber beam model updating method for quantitative service performance evaluation of RC flexural members is established. The method is verified by some case studies on typical RC beams and slabs.(3) Experimental study on an existing RC hollow slab beam bridge was conducted, aiming to apply and verify the proposed crack detection method together with the model updating and service performance evaluation method. The appearance defect detection, lateral load distribution test, field destructive test, laboratory destructive test, and cross-sectional configuration and material property test were carried out. The tests reveal the lateral load distribution performance of the bridge and the full-range mechanical performance of hollow slab beams subjected to uniaxial monotonic and hysteretic loads.(4) Model updating and service performance evaluation of existing RC hollow slab beam bridges are carried out. An improved hinge-connected slab/beam method considering the damage of hinge joints and girders is proposed. The actual stiffnesses of hinge joints and hollow slab beams are estimated according to the lateral load distribution test results. Based on the experimental data, the stiffnesses and ultimate capacities of hollow slab beams are evaluated, and the applicability of the calculation method in typical standards is discussed. In addition, the fiber beam model updating is conducted to invert the material properties and accurately reproduce the full-range structural response of hollow slab beams.(5) A set of UAV path planning method for comprehensive bridge surface crack detection mission is established. A path planning example of a bridge conceptual model is presented to validate the feasibility of this method.This dissertation is supported by the National Natural Science Foundation of China (52192662, 51978376).
