围绕道路结构检测与维护 研究，本期精选发表在《中国公路学报》5篇双语文章 ，欢迎阅读！

精选文章·Selected Articles

01

三维探地雷达道路隐性病害检测分析与数字化技术综述

Road Structural Defects Detection and Digitalization Based on 3D Ground Penetrating Radar Technology: A State-of-the-art Review

【摘要】道路服役数据信息是交通基础设施数字化建设与养护的关键，先进探测设备是获取道路服役数据信息的途径。近年来，三维探地雷达(3D Ground Penetrating Radar, 3D GPR)因其高效、无损等检测优势得到广泛应用，可为道路隐性病害数据获取提供重要支撑。基于此，对道路典型隐性病害类型与检测手段进行总结归纳。梳理了三维探地雷达技术原理、数据采集方法、数据处理及在道路工程检测中的应用；根据三维探地雷达图谱隐性病害特征与识别手段，重点分析人工智能技术在探地雷达图谱识别技术中的应用与发展。针对交通基础设施发展进程，展望基于探地雷达数据的数字孪生技术，主要介绍基于三维探地雷达数据的建模与模型仿真方法。该综述可为三维探地雷达道路隐性病害检测提供基础理论知识与实践方法借鉴，同时为基于三维探地雷达数据的数字化交通基础设施建设与道路养护决策提供参考。

【Abstract】Road information is the key to the decision-making of digital traffic infrastructure and maintenance behavior. The application of advanced detection equipment is the way to obtain road information. The high-efficiency, and non-destructive detection characteristics of 3D Ground Penetrating Radar (3D GPR) can support the acquisition of road structural defects data. This paper reviews the typical types of road structural defects, detection methods, etc. The major principles of 3D GPR technology, and its application on road engineering are introduced afterwards. Subsequently, this paper concludes the application and development of artificial intelligence technology in GPR image recognition technology according to the structural defect characteristics and recognition methods. According to the development of traffic infrastructure, this paper looks forward to the digital twin technology based on 3D GPR data, mainly introduces the modeling and model simulation methods. The review can provide basic theoretical knowledge and practical methods for 3D GPR road structural defections detection and provide guidance for digital transport infrastructure construction and road maintenance based on 3D GPR data.

02

基于深度学习的土木基础设施裂缝检测综述

Review of Deep Learning-based Crack Detection for Civil Infrastructures

【摘要】基于深度学习的裂缝检测对于降低基础设施运营风险、节约运维成本并推进中国土木工程行业智能化转型具有重要意义。算法、数据集和评价指标是构建深度学习裂缝检测模型的关键要素；裂缝检测模型集成于机器人平台，从而实现对土木基础设施的全自动裂缝检测。为此，从以上4个方面对当前研究进行了系统梳理。首先，回顾了深度学习的发展历程，重点介绍了深度卷积神经网络在计算机视觉领域的应用及其在图像处理方面较传统算法所具有的显著优势。接着，详细介绍了3类基于深度学习的裂缝检测主流算法，包括分类算法、目标检测算法和语义分割算法。然后，对现有裂缝图像数据集以及模型性能评价指标进行了归纳。最后，总结了土木基础设施的各类裂缝检测机器人平台。综合分析表明：基于卷积神经网络主干结构的深度学习算法已被广泛用于土木基础设施表面裂缝的精准定位与分类，而裂缝的尺寸信息仍需依靠传统图像处理技术进行提取；由于像素级标注的成本和专业性高，大型的裂缝语义分割数据集相对缺乏，致使当前基于语义分割算法的裂缝检测模型鲁棒性较差；目前多数研究人员采用个人建立的裂缝数据集进行模型训练且采用不同的指标进行模型性能评价，缺乏统一的基准测试数据集和评价指标体系，无法对不同模型的性能进行平行比较；目前针对不同基础设施已相应开发了一些裂缝检测机器人，提高裂缝检测机器人的多场景适应性，并降低其应用成本是未来的发展方向。

【Abstract】Crack detection using deep learning (DL) is important for reducing infrastructure operation risks, saving operation and maintenance costs, and promoting the intelligent transformation of the civil engineering industry. In current practice, algorithms, datasets, and evaluation metrics are the key components of DL-based crack detection. Hence, in this article, four aspects of current research are reviewed systematically. First, the development of the DL method is reviewed, and the applications of deep convolution neural networks in the field of computer vision and their significant advantages over conventional algorithms in image data processing are introduced. Second, three popular DL algorithms for crack detection are described in detail. Third, the available crack image datasets and current evaluation metrics are reviewed. Finally, recent research outcomes in this field are summarized, and future research needs are discussed. The comprehensive analyses show that, based on the backbone of convolutional neural networks, DL algorithms have been widely used to locate and classify cracks on the surface of civil infrastructure with good accuracy, although obtaining quantitative information about cracks still requires auxiliary extraction using traditional image-processing technology. Because of the high cost and technical skills required for pixel-level annotation, there is a lack of large-scale crack semantic segmentation datasets, resulting in crack detection models with poor robustness. Moreover, most researchers created their own datasets for model training and used different metrics to evaluate their model performance, highlighting the need to establish a benchmark dataset for model training and use a set of popular indices to compare the performances of different models. Crack detection robots have been developed for different types of infrastructure, and it is the development trend to improve multiscene adaptability and reduce the application cost of crack detection robots.

03

内置应变传感器与沥青路面结构协同工作性能研究

Research on Cooperative Performance of Built-in Strain Sensor and Asphalt Pavement Structure

【摘要】为研究传感器与沥青路面结构的协同工作性能，揭示传感器的感知效能及与混合料协同工作的性能演化规律，通过对融合内置传感器的沥青混合料试件，开展多重模式下的力学响应试验，分析不同加载模式和环境工况下的应变响应规律特征；采用ABAQUS软件，建立融合内置传感器的梁试件有限元模型，通过数值模拟与计算，并结合四点弯曲疲劳试验，对内置传感器与沥青混合料协同工作下的交互影响行为和耐久性进行分析。研究结果表明：在逐级加载和动载作用下，传感器表现出良好的感知时效性；实测应变响应受混合料黏弹性质影响存在明显的滞后效应，加载频率f愈大，滞后效应产生的不利影响愈明显；当信息采集频率f′为加载频率f的120倍时，数据采集效果最佳；应变传感器在温度变化、水-热耦合及不同速率等工况下能保持稳定的应变感知功能，实测应变响应曲线符合沥青路面结构的力学行为特征；与传感器法兰相接触的混合料和传感器自身测力杆处存在明显的应力集中现象，导致梁试件发生疲劳开裂而降低混合料的疲劳寿命；荷载反复作用下，裂缝的形成和拓展及传感器与基体材料发生滑移，造成实测应变难以反映混合料的真实形变，且感知功能呈现协同工作、感知失准和感知失效3阶段特征。

【Abstract】 To study the cooperative performance of the sensor and the asphalt pavement structure, reveal the sensor’s perceived efficiency and the evolution of the performance of the sensor’s cooperative work with the mixture. Through the fusion of the asphalt mixture specimens with built-in sensors, the mechanical response test under multiple modes is carried out, and the characteristics of the strain response law under different loading modes and environmental conditions are analyzed. By using ABAQUS software, a finite element model of beam specimens with built-in sensors is established. Combined with numerical simulation and four-point bending fatigue test, the interactive influence behavior and durability under the built-in strain sensor works with asphalt mixture are explored. The results show that the sensor exhibits a good sense of timeliness under progressive loading and dynamic loading. Because of the influence of the viscoelasticity of the mixture, the strain response has obvious hysteresis. The greater the loading frequency f, the more obvious the adverse effects of hysteresis. When the information collection frequency f′ is 120 times the loading frequency f, the data collection effect is the best. The strain sensor can maintain a stable strain sensing function under working conditions such as temperature changes, water-heat coupling, and different speeds. And the measured strain response curve accords with the mechanical behavior characteristics of the asphalt pavement structure. There is obvious stress concentration in the mixture in contact with the sensor flange and the sensor’s force measuring rod, which leads to fatigue cracking of the beam specimen and reduces the fatigue life of the mixture. Under repeated loads, the formation and expansion of cracks, and the slippage of the sensor and the matrix material, which make the measured strain difficult to reflect the true deformation of the mixture. And the sensing function presents three-stage characteristics of cooperative work, sensing inaccuracy, and sensing failure.

04

沥青路面表面纹理重构与构造深度预测模型

Surface Texture Reconstruction and Mean Texture Depth Prediction Model of Asphalt Pavement

【摘要】为了实现通过调整混合料的级配设计来获得期望的路面平均构造深度的目标，采用高精度三维激光扫描技术，采集了AC、SMA、OGFC三种典型级配的沥青混合料试件表面纹理特征信息。通过邻域插值法对采样数据的异常值和离群值进行替换，并通过均值滤波对采样数据进行降噪处理后，三维重构了试样表面；在采用傅里叶变换得到重构表面频域信息的基础上，根据宏观纹理的波长对应的频率设计带通滤波器，从重构表面中分离并提取出了路面宏观纹理。应用蒙特卡罗算法计算了路面的平均构造深度，通过采用筛上质量比-粒径积同时考虑了混合料的粒径和筛孔通过率对平均构造深度的影响。采用多元线性回归、随机森林和人工神经网络的方法，建立筛上质量比-粒径积与平均构造深度的预测模型，研究了混合料级配对沥青路面平均构造深度的影响。研究结果表明：均值滤波在去除噪声信号的同时也比较完整地保留了高程轮廓特征，三维重构的试样表面特征与原始表面特征一致；平均构造深度会受到级配曲线中除最大公称粒径外的其他粒径及筛孔通过率的影响；通过多元线性回归、随机森林和人工神经网络3种模型建立了以各筛孔尺寸的筛上质量比-粒径积为自变量，平均构造深度为因变量的回归模型，得到的预测值与实测值的决定系数R2在0.95以上。

【Abstract】To achieve the desired pavement mean texture depth by adjusting the gradation design of an asphalt mixture, high-precision three-dimensional laser scanning technology was used to collect the surface texture feature information of three typical gradation asphalt mixture specimens: Asphalt Concrete, Stone Matrix Asphalt, and Open Graded Friction Course. After the exception values and outliers were processed by neighborhood interpolation and the sampled data was denoised by mean filtering, the sample surface was reconstructed in three dimensions. A band-pass filter was designed according to the frequency corresponding to the wavelength of the macro texture based on the frequency domain information of the reconstructed surface obtained by the Fourier transform; the macro texture of the pavement was separated and extracted from the reconstructed surface. A Monte Carlo algorithm was used to calculate the mean texture depth of the pavement. The influence of the mixture particle size and passing rate of the sieve on the mean texture depth was considered by using the product of the mass ratio on the sieve and particle size. The prediction models of the product of the mass ratio on sieve and particle size, and the mean texture depth were established using multiple linear regression analysis, random forest, and artificial neural networks; the influence of the mixture gradation on the mean texture depth of the asphalt pavement was studied. The results show that mean filtering not only removes the noise signal but also retains the elevation profile features. The surface features of the three-dimensional reconstructed specimen are consistent with the original surface features. The mean texture depth is affected by other particle sizes in the grading curve, except for the maximum nominal particle size and passing rate of the sieve. The regression model was established using multiple linear regression, random forest, and an artificial neural network, which takes the product of the mass ratio and particle size on the sieve of each mesh size as the independent variable and the mean texture depth as the dependent variable, has an R2 of more than 0.95.

05

基于深度学习的土木基础设施裂缝检测综述

Review of Deep Learning-based Crack Detection for Civil Infrastructures

【摘要】基于深度学习的裂缝检测对于降低基础设施运营风险、节约运维成本并推进中国土木工程行业智能化转型具有重要意义。算法、数据集和评价指标是构建深度学习裂缝检测模型的关键要素；裂缝检测模型集成于机器人平台，从而实现对土木基础设施的全自动裂缝检测。为此，从以上4个方面对当前研究进行了系统梳理。首先，回顾了深度学习的发展历程，重点介绍了深度卷积神经网络在计算机视觉领域的应用及其在图像处理方面较传统算法所具有的显著优势。接着，详细介绍了3类基于深度学习的裂缝检测主流算法，包括分类算法、目标检测算法和语义分割算法。然后，对现有裂缝图像数据集以及模型性能评价指标进行了归纳。最后，总结了土木基础设施的各类裂缝检测机器人平台。综合分析表明：基于卷积神经网络主干结构的深度学习算法已被广泛用于土木基础设施表面裂缝的精准定位与分类，而裂缝的尺寸信息仍需依靠传统图像处理技术进行提取；由于像素级标注的成本和专业性高，大型的裂缝语义分割数据集相对缺乏，致使当前基于语义分割算法的裂缝检测模型鲁棒性较差；目前多数研究人员采用个人建立的裂缝数据集进行模型训练且采用不同的指标进行模型性能评价，缺乏统一的基准测试数据集和评价指标体系，无法对不同模型的性能进行平行比较；目前针对不同基础设施已相应开发了一些裂缝检测机器人，提高裂缝检测机器人的多场景适应性，并降低其应用成本是未来的发展方向。

【Abstract】Crack detection using deep learning (DL) is important for reducing infrastructure operation risks, saving operation and maintenance costs, and promoting the intelligent transformation of the civil engineering industry. In current practice, algorithms, datasets, and evaluation metrics are the key components of DL-based crack detection. Hence, in this article, four aspects of current research are reviewed systematically. First, the development of the DL method is reviewed, and the applications of deep convolution neural networks in the field of computer vision and their significant advantages over conventional algorithms in image data processing are introduced. Second, three popular DL algorithms for crack detection are described in detail. Third, the available crack image datasets and current evaluation metrics are reviewed. Finally, recent research outcomes in this field are summarized, and future research needs are discussed. The comprehensive analyses show that, based on the backbone of convolutional neural networks, DL algorithms have been widely used to locate and classify cracks on the surface of civil infrastructure with good accuracy, although obtaining quantitative information about cracks still requires auxiliary extraction using traditional image-processing technology. Because of the high cost and technical skills required for pixel-level annotation, there is a lack of large-scale crack semantic segmentation datasets, resulting in crack detection models with poor robustness. Moreover, most researchers created their own datasets for model training and used different metrics to evaluate their model performance, highlighting the need to establish a benchmark dataset for model training and use a set of popular indices to compare the performances of different models. Crack detection robots have been developed for different types of infrastructure, and it is the development trend to improve multiscene adaptability and reduce the application cost of crack detection robots.

期刊推荐·Recommend Journal

《中国公路学报》是由中国科学技术协会主管，中国公路学会主办，长安大学承办的公路交通行业的学术性刊物。

办刊宗旨：提升公路交通领域学术交流质效，促进中国公路交通科技创新，推动中国公路交通事业发展。

联系我们·Contact Us

中国知网“中文精品学术期刊外文版数字出版工程”（简称JTP）自2015年启动，已与400余种学术期刊合作出版了4万余篇双语对照论文，积累了丰富的学术翻译/英语加工/学术推广经验。形成了集双语出版、主题电子书出版、双语讲座视频制作、期刊英文内容编校加工、资讯编译、海外推广为一体的全方位服务体系，全面助力期刊提升国际影响力。