Radar and deep learning-based pavement internal disease three-dimensional morphology identification method

Abstract

The invention discloses a radar and deep learning-based road surface internal disease three-dimensional morphology identification method, which integrates a novel technical path of qualitative identification and quantitative reconstruction, provides a high-precision and quantifiable reference for subsequent non-damage assessment by constructing a disease three-dimensional geometric model and reversely calibrating radar map characteristic parameters, and improves the accuracy and the accuracy of the three-dimensional morphology identification of a road surface internal disease. The problem that the three-dimensional morphology characteristics of the internal diseases of the asphalt pavement are difficult to accurately obtain at present is solved, and a calibration method is provided for obtaining the three-dimensional morphology characteristics of the internal diseases of the pavement based on a ground penetrating radar technology. And meanwhile, geometric data support is provided for mechanical response analysis and performance attenuation law research of the existing asphalt pavement structure containing internal diseases.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of asphalt pavement disease morphology feature acquisition, and particularly relates to a pavement internal disease three-dimensional morphology identification method based on radar and deep learning. BACKGROUND

[0002] Three-dimensional morphology features can comprehensively and accurately reflect the spatial distribution, shape, size, and depth of pavement diseases, and other key information. This helps to accurately determine the disease type, severity, and development trend, thereby providing a scientific basis for subsequent maintenance decisions. By comparing three-dimensional morphology data at different time periods, the development speed of the disease can also be analyzed, future potential problems can be predicted, and preventive maintenance can be achieved. Traditional two-dimensional detection methods can only obtain information on the surface of the pavement, and it is difficult to detect diseases hidden inside. Three-dimensional detection methods can penetrate into the internal part of the pavement and capture more subtle disease information, reducing the possibility of missed detection and false detection. At the same time, three-dimensional data have higher precision and resolution, making the detection results more reliable. Three-dimensional morphology features also have potential application value in pavement material research, structure design, and construction process improvement. By analyzing and comparing the three-dimensional morphology features of different pavement materials and structures, the material ratio and structure design can be optimized to improve the durability and disease resistance of the pavement. At the same time, three-dimensional morphology features can also be used to evaluate the pros and cons of the construction process and guide the improvement and optimization of the construction process.

[0003] Currently, the main means to obtain internal diseases of pavement structure is to use three-dimensional ground penetrating radar to reconstruct the space of disease area. However, the ground penetrating radar signal is greatly affected by factors such as material heterogeneity, water change, and disease morphology, and the echo signal is complex, which causes the three-dimensional reconstruction result to have problems such as positioning deviation and fuzzy boundary, and the overall reconstruction accuracy is low. Although the reconstruction algorithm of three-dimensional ground penetrating radar has been widely studied and developed in recent years, there is still a lack of stable and objective quantitative evaluation index to verify and calibrate the reconstruction result.

[0004] In addition, at present, pavement coring is still widely used for type identification and structure performance evaluation of internal diseases of asphalt pavement. Through core sample sectioning or cross-section observation, qualitative analysis of diseases such as cracks and voids can be performed. However, the structure of the core sample is often disturbed or the edge is damaged during drilling, which limits its integrity, making it difficult to accurately describe the true geometric morphology and spatial distribution characteristics of the disease, especially when the disease has complex boundaries or hidden expansion. At the same time, core sample acquisition has the disadvantages of strong destructiveness, high operation cost, and limited detection range, which cannot meet the urgent needs of modern pavement maintenance for non-destructive, high-precision, and visual disease evaluation methods.

[0005] In summary, the existing non-destructive reconstruction method based on three-dimensional ground penetrating radar has limited accuracy, and lacks effective topographic calibration means. The traditional coring method also has problems such as structural disturbance and lack of geometric information, making it difficult to achieve high-fidelity acquisition of disease topography. Therefore, a new technical path that combines qualitative identification and quantitative reconstruction is needed. By constructing a three-dimensional geometric model of the disease, the radar map feature parameters are calibrated in reverse, providing a high-precision, quantifiable reference benchmark for subsequent non-destructive evaluation, thereby effectively improving the accuracy of disease identification and the intelligent level of road detection. SUMMARY

[0006] In view of the above problems in the prior art, the three-dimensional topography recognition method for internal diseases of pavement based on radar and deep learning provided by the present application is not only suitable for various internal diseases of asphalt pavement, but also has quantifiable three-dimensional topographic features, high accuracy, and small external interference, and serves the mechanical response analysis of the asphalt pavement structure in a real service state and the calibration of the three-dimensional reconstruction results of internal diseases based on ground penetrating radar technology.

[0007] To achieve the above-mentioned purposes, the technical scheme adopted by the present application is as follows:

[0008] The present application provides a three-dimensional topography recognition method for internal diseases of pavement based on radar and deep learning, comprising the following steps:

[0009] S1, preliminary identification of disease area and type analysis:

[0010] Select a test section and scan it using a three-dimensional ground penetrating radar to identify the area with internal diseases and record it as a disease area. Preliminary identification of disease type is performed through the characteristic parameters of the radar map;

[0011] S2, drill a grouting hole in the detected disease area, inject a high-performance two-component polyurethane material with good fluidity and fast curing speed, and make it fully penetrate into the disease area and complete curing. Then, drill a core sample containing the cured polyurethane material;

[0012] S3, CT scanning and image preprocessing of core sample:

[0013] CT scanning of the core sample is performed by layer-by-layer scanning perpendicular to the axial direction of the core sample to obtain a continuous sequence of two-dimensional slice images;

[0014] Deep learning or advanced image processing algorithms are used to filter, enhance and pixel-level segment the images to extract the distribution information of the polyurethane material in the disease area, and obtain a binary segmentation image after accurate identification of the disease area;

[0015] S4, three-dimensional topography reconstruction and feature extraction of disease:

[0016] Based on the binary segmentation image, the image is reconstructed to construct a three-dimensional voxel model of the polyurethane filling area, and a three-dimensional reconstruction algorithm is used to extract an isosurface structure to obtain a three-dimensional geometric feature of the disease area, wherein the quantitative indexes of the three-dimensional geometric feature include a spatial volume, a penetration height, and a contact area with a structure interface.

[0017] The data of the three-dimensional geometric feature of the disease is matched with the corresponding radar map to construct a calibration sample library.

[0018] S5, calibrate the radar map and the three-dimensional feature of the disease:

[0019] Based on a large number of calibrated radar maps and three-dimensional geometric feature sample data in the sample library, an end-to-end mapping model from the radar map to the disease feature is constructed.

[0020] Based on deep learning methods such as convolutional neural network (CNN), Transformer structure, and graph neural network (GNN), the spatiotemporal features of the radar map are automatically extracted to realize high-precision prediction between the radar map input and the three-dimensional geometric parameters of the disease, wherein the three-dimensional geometric parameters of the disease at least include a spatial volume, a penetration height, and an interface contact area.

[0021] S6, intelligent recognition of the three-dimensional feature of the road disease based on the mapping model:

[0022] In a continuous road section, real-time radar map data can be obtained by three-dimensional ground penetrating radar scanning, and combined with the trained end-to-end mapping model from the radar map to the disease feature, the three-dimensional geometric feature of the disease area can be quickly inverted and recognized, and the non-destructive spatial positioning and three-dimensional reconstruction of the internal disease of the asphalt pavement can be achieved.

[0023] In the step S1, the ground penetrating radar should use a high-frequency three-dimensional ground penetrating radar to facilitate intuitive identification of the position of the road disease area.

[0024] In the step S2, the injected high-performance polyurethane material should be a two-component polyurethane material, and the viscosity of each component material should be 200-500 mPa·s to ensure that the polyurethane material can better penetrate into the disease; the tensile strength after curing is greater than 5 MPa to facilitate rapid curing of the material at the disease position and to obtain a complete core sample.

[0025] In the step S2, the injection of the high-performance polyurethane material can be achieved by drilling a hole at the determined road disease position area, and injecting the high-performance polyurethane material into the road disease position area by a grouting machine.

[0026] In the step S2, the height of the core sample should include the thickness of the pavement surface layer and the upper and lower base layers, and the diameter of the core sample should be 100-150 mm.

[0027] In step S3, an industrial high-precision CT device with a resolution of not less than 30 μm / voxel should be used to ensure clear imaging of millimeter-level disease cracks; the scanning time should be controlled within 10 minutes to prevent changes in the polyurethane and the environment from causing topographic errors.

[0028] In step S3, a deep semantic segmentation neural network based on the U-Net architecture is recommended for automatic segmentation of the disease area in the CT image. The network includes 4 layers of down-sampling and up-sampling modules, which can effectively capture the boundary feature of the polyurethane grouting area.

[0029] In step S3, to improve the robustness of the model in identifying small targets, it is suggested to use a weighted combination of Dice loss function and BinaryCross-Entropy loss function for optimization during the training phase, and the error of the disease area contour after segmentation should be less than 1 pixel.

[0030] In step S3, image preprocessing should include steps such as Gaussian filtering (kernel size 3x3), histogram equalization, and pseudo-color enhancement to improve the contrast and edge recognizability of the model input image.

[0031] In step S4, the image reconstruction process should be based on a voxel image sequence to construct a three-dimensional voxel model, and a voxel precision of not more than 0.3 mm is recommended to ensure that the geometric reconstruction precision meets the evaluation requirements.

[0032] In step S4, it is suggested to use the MarchingCubes triangular mesh extraction algorithm to obtain a three-dimensional isosurface model of the disease area. The algorithm traverses the regular voxel grid and generates a topologically continuous triangular mesh according to the intensity threshold.

[0033] In step S4, the geometric feature parameters should include but are not limited to: the spatial volume of the disease area, calculated using the integral voxel number method; the maximum penetration height, defined as the maximum projection size of the disease area in the vertical direction; and the contact area with the asphalt layer-base interface, estimated using the triangular mesh area accumulation method.

[0034] In step S4, it is suggested to introduce the Alpha-shape boundary optimization algorithm to reconstruct the edge of complex disease areas and improve the spatial geometric precision, with a boundary error of not more than ±1 mm.

[0035] In step S5, the three-dimensional disease topographic features can include geometric parameters such as spatial volume, maximum penetration height, disease-structure interface contact area, and length-width-height direction projection area, to fully characterize the spatial distribution pattern and development trend of the disease.

[0036] In the step S5, a convolutional neural network (CNN) is preferably used for feature extraction of the radar map, and the neural network structure includes 35 convolutional layers and 23 fully connected layers, which can effectively capture key criteria such as local high-reflection areas, tailing patterns and multi-layer echo structures in the map.

[0037] In the step S5, in order to improve the generalization ability of the model under different radar map patterns, data enhancement can be performed through map rotation, noise addition, time delay disturbance and the like, so as to improve the robustness and actual applicability of the model.

[0038] In the step S5, during the model training process, mean square error (MSE) is used as the loss function, and an L2 regularization term can be introduced to control the model complexity and avoid overfitting.

[0039] In the step S5, in order to enhance the physical interpretability of the model, Grad-CAM or SHAP and the like can be used to label the model recognition area, and the focusing ability of the model on the response characteristics of the disease area is verified.

[0040] In the step S6, the trained mapping model is deployed to the radar detection map of the entire road, and the continuous radar data is processed in a sliding window manner, and the three-dimensional geometric morphology parameters of the corresponding disease area are extracted in sections.

[0041] In the step S6, in order to ensure the integrity of the non-damaged reconstruction, a minimum recognition window for map processing should be set, and the length is preferably 50-100 cm, and can be adaptively adjusted according to the disease spacing or the radar map resolution.

[0042] In the step S6, the reconstructed three-dimensional parameters of the disease should be spatially registered with the road structure level (such as the asphalt layer and the base layer), so as to distinguish the specific occurrence horizon of the disease and improve the structural specificity of the diagnosis.

[0043] In the step S6, the three-dimensional morphology results of the entire road disease can finally be imported into the GIS or BIM platform in a structural layer mode for visual display, so as to facilitate the spatial positioning, disease classification and subsequent maintenance decision of the road maintenance unit.

[0044] Compared with the prior art, the present application has the following advantages:

[0045] The three-dimensional morphology recognition method of internal diseases of the road surface based on radar and deep learning provided by the present application solves the problem of accurately obtaining the three-dimensional morphology characteristics of internal diseases of the asphalt road surface, and provides a calibration method for obtaining the three-dimensional morphology characteristics of internal diseases of the road surface based on the ground penetrating radar technology. At the same time, it also provides geometric data support for the mechanical response analysis and performance attenuation law research of the existing asphalt road surface structure containing internal diseases. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the present application, the embodiments of the present application are described in detail below with reference to the accompanying drawings. It should be understood that the accompanying drawings are only used to illustrate one embodiment of the present application and do not constitute a limitation on the protection scope of the present application. Based on the accompanying drawings, those skilled in the art can also construct other similar embodiments without creative labor.

[0047] Figure 1 is a flowchart of an embodiment of the present application;

[0048] Figure 2 is a flowchart of disease mapping - three-dimensional topography mapping modeling and identification. DETAILED DESCRIPTION

[0049] In order to make the application purposes, features and advantages of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings of the embodiments of the present application. Obviously, the embodiments described below are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0050] The present application provides a three-dimensional topography identification method for internal diseases of a road surface based on radar and deep learning, comprising the following steps:

[0051] S1, preliminary identification of disease area and type analysis:

[0052] GER-SLICE software is used to image and process data of the radar image, to identify the internal disease position and disease type of the road. The subsequent grouting and coring area is determined.

[0053] According to the radar detection results, the internal diseases of the road section are mainly divided into: cracks, poor interlayer bonding, interlayer void, loose and broken, and high water content between layers. According to the disease influence degree, formation reason and other factors, they are preliminarily divided into the following three categories:

[0054] Base cracks are mainly bottom-up reflection cracks;

[0055] Loose structure layers mainly include: surface layer loose, upper base layer loose and lower base layer loose, etc.

[0056] Poor interlayer bonding of structure layers mainly includes: poor interlayer bonding between surface layer and base layer, poor interlayer bonding between upper and lower base layers, and poor interlayer bonding between lower base layer and bottom base layer.

[0057] S2, Polyurethane injection in disease area and core sample acquisition

[0058] According to the radar detection results, a grouting scheme is formulated, a grouting hole (hole diameter 3 cm, depth to the top surface of the base layer) is set at the center of the disease, and high-performance two-component polyurethane material is injected using a grouting device. The selected material is a high-performance two-component polyurethane material produced by a certain company, wherein the viscosity of component A is 250 mPa·s, the viscosity of component B is 280 mPa·s, the mixing ratio is 1:1, the working time is about 60 seconds, and the complete curing time is not more than 10 minutes.

[0059] The grouting pressure is controlled within 0.2 MPa to ensure that the material does not overflow to the non-disease area. After 30 minutes, when the material is completely cured, a Ф100 mm diamond hollow drill is used to drill a complete core sample from the grouting area, with a length of about 65 cm, containing the asphalt surface layer and the upper and lower base layer structure.

[0060] S3, Core sample CT scanning and image preprocessing

[0061] The complete core sample obtained by drilling is scanned layer by layer perpendicular to the core sample axis using an industrial high-resolution CT scanning device (model: YXLON FF20 CT) with a resolution of 15 μm, to obtain a complete sequence of two-dimensional grayscale images.

[0062] The scanning layer thickness is set to 0.1 mm, the core sample height is 65 cm, and finally about 6500 two-dimensional slice images (TIFF format) can be obtained for subsequent image preprocessing and three-dimensional reconstruction operations.

[0063] To prevent the appearance deviation caused by the change of polyurethane over time, the scanning should be completed within 10 minutes.

[0064] After scanning, the images are imported into the Python image processing environment (OpenCV+scikit-image+NumPy), and the following preprocessing operations are performed in sequence:

[0065] Gray scale normalization: all pixel gray scales are normalized to the interval;

[0066] Gaussian filter denoising: the convolution kernel size is set to 5x5, the standard deviation is removed to remove scanning noise;

[0067] Histogram equalization: enhance the overall contrast of the image and improve the boundary clarity;

[0068] A deep semantic segmentation neural network based on U-Net architecture is used to automatically identify the polyurethane grouting area in the CT image. The network structure contains 4 layers of down-sampling and up-sampling modules, and fuses cross-layer feature connection, which has good boundary extraction capability and is suitable for precise positioning of disease area.

[0069] To enhance the robustness of the model in identifying small size grouting area, the weighted combination of Dice loss function and BinaryCross-Entropy (BCE) loss function is used for optimization in the training stage. The definition is as follows:

[0070]

[0071] Wherein, is the weight coefficient,

[0072] represents the binary cross-entropy loss, and the calculation formula is: , n is the number of samples;

[0073] represents the Dice loss function, which is used to measure the overlap between the predicted area and the real area, and its expression is:

[0074]

[0075] Wherein, is the true label of the ith pixel, is the prediction probability of the corresponding pixel, and N is the total number of pixels, is a small constant to avoid division by zero, which is set according to the empirical value.

[0076] The training set is constructed by manual annotation, and the model prediction output is a binary segmentation image consistent with the original image size.

[0077] The segmentation result of each slice image is output in the form of a binary image, where the pixel value of the polyurethane grouting area is 255 (white), and the rest of the area is 0 (black). The model prediction result is processed by morphological dilation-erosion, which further improves the boundary continuity and segmentation accuracy. The segmentation contour error is controlled within 1 pixel to meet the spatial accuracy requirements of subsequent three-dimensional reconstruction.

[0078] S4, three-dimensional morphology reconstruction and feature extraction of disease

[0079] The binary image sequence obtained in step S3 is stacked layer by layer according to the slice in the axial direction of the core sample (defined as Z axis) to generate a three-dimensional voxel array , wherein represents the polyurethane grouting area, and V=0 represents the background area.

[0080] The Marching Cubes algorithm is used to extract the isosurface of the three-dimensional voxel model to generate a three-dimensional geometric model represented by a triangular mesh wherein is a set of vertices, is a set of triangular facets. This algorithm can accurately depict the complex disease morphology while maintaining the continuity of the boundary.

[0081] Reconstructing the volume of the voxel model The calculation formula is: wherein, is the number of voxels with a voxel value of 1, are the actual sizes of the voxels in the X, Y, and Z directions (0.1 mm in this embodiment).

[0082] Maximum penetration height The maximum penetration height is obtained by finding the maximum continuous non-zero section of the polyurethane region in the Z-axis direction. It is defined as:

[0083] , ,

[0084] That is, within the interval and , for all z, if the function representing the polyurethane region is equal to 1 (i.e., the region is continuous and non-zero in this Z-axis interval), then the maximum penetration height is the maximum value among all such .

[0085] Contact area calculation: Extract the facet set of the polyurethane region near the asphalt-base layer interface , and convert it to the total area , i.e.:

[0086] , represents the facet set , represents the area of the i-th facet.

[0087] The three-dimensional reconstruction and parameter extraction process can be completed on the Python platform with the help of third-party libraries such as scikit-image, PyMCubes, and trimesh. Finally, the three-dimensional geometric parameters of the disease are obtained, including but not limited to: spatial volume, maximum penetration height, and contact area.

[0088] S5, Radar map and calibration of disease three-dimensional morphology

[0089] ​To achieve quantitative interpretation and morphological feature identification of 3D ground-penetrating radar (GPR) images, and to establish a mapping relationship between radar images and the 3D geometric morphology of internal defects, a map calibration model based on a convolutional neural network (CNN) is proposed. This model uses the 3D morphological parameters of the polyurethane-filled region reconstructed from CT scans as calibration labels, and the corresponding radar image as input, enabling the direct prediction of defect structural features from radar data.

[0090] During the model training phase, multiple typical disease core samples were collected, and their radar atlases and CT reconstruction results were obtained. The radar atlases and 3D reconstructed morphologies of each group of diseases were paired one-to-one to construct a high-quality training sample set. Before inputting the atlases, their dimensions were standardized (e.g., to 512×512 pixels), and normalization, contrast enhancement, and other standardization preprocessing were performed. Target morphology parameters, including disease volume H, maximum penetration height H, and contact area A at the interface with the asphalt structure, served as prediction objects for the model's multi-output regression, enabling geometric calibration of the radar atlases.

[0091] The network model backbone adopts a ResNet18 structure, utilizing its multi-layer residual convolutional modules to efficiently extract key reflection regions, trailing structures, and other disease features from the radar image. Two fully connected layers are connected at the end of the model for regression output, corresponding to the morphology parameters α, H, and A, respectively. Mean Squared Error (MSE) is used as the loss function during training, with the Adam optimization algorithm, an initial learning rate of 0.001, and 200 training epochs. To enhance the model's adaptability to different disease morphologies and radar image variations, data augmentation strategies such as image rotation and additive noise perturbation are introduced during training to improve the model's generalization ability in real radar image scenarios.

[0092] S6. Intelligent Recognition of 3D Morphology of Road Surface Defects Based on Mapping Model

[0093] The trained CNN model was applied to the radar image analysis of the entire road. A sliding window analysis method was used, with the radar image window size set to 1.0m and the step size to 0.5m. Continuous radar scan image segments were extracted sequentially and input into the model, outputting the corresponding predicted geometric features of the road defects.

[0094] The model output includes parameters such as volume, penetration height, and contact area for each map segment. By stitching multiple predicted segments along the road axis and spatially interpolating based on the prediction results, the three-dimensional morphology of the entire road segment's defects is reconstructed. The obtained three-dimensional parameter data can be imported into 3D modeling software (such as Rhino, 3ds Max) or BIM systems to achieve digital visualization of the defect morphology, assisting in structural health assessment and maintenance strategy formulation.

[0095] The method realizes non-damage identification and three-dimensional visual reconstruction of internal diseases of the asphalt pavement, avoids the damage and sample limitation of the traditional core sampling mode, and has good engineering application prospect and popularization value.

[0096] The preferred embodiments of the application are described in detail above, but the application is not limited to the specific details in the above-described embodiments, and various equivalent transformations (such as quantity, shape, position, etc.) can be made to the technical solutions of the application within the technical concept of the application, and these equivalent transformations all belong to the protection of the application.

Claims

1. A radar and deep learning based method for identifying three-dimensional topography of internal pavement diseases, characterized in that, include: S1. Preliminary identification and type analysis of diseased areas: A test section was selected and scanned using three-dimensional ground-penetrating radar to identify areas with internal defects and record them as defect areas. The type of defects was preliminarily determined by the characteristic parameters of the radar map. S2. Obtain core samples after injecting polyurethane into the affected area: In the detected disease area, grouting holes are drilled, high-performance polyurethane material is injected, and after it fully penetrates into the disease area and completes curing, core samples are drilled to obtain complete core samples containing cured polyurethane material. S3, Core CT Scan and Image Preprocessing: The core sample was CT scanned layer by layer perpendicular to the axis of the core sample to obtain a continuous sequence of two-dimensional slice images. Deep learning or advanced image processing algorithms are used to filter, enhance, and segment images at the pixel level to extract the distribution information of polyurethane material in the diseased area, and obtain a binary segmented image after accurate identification of the diseased area. S4. Three-dimensional morphological reconstruction and feature extraction of disease: Image reconstruction is performed based on binary segmentation images to construct a three-dimensional voxel model of the polyurethane filling region. The isosurface structure is extracted using a three-dimensional reconstruction algorithm to obtain the three-dimensional geometric features of the diseased area. The quantitative indicators of the three-dimensional geometric features include spatial volume, penetration height, and contact area with the structural interface. The data of the three-dimensional geometric morphology features of the disease are matched with the corresponding radar images to construct a calibration sample library; S5. Calibration of radar map and three-dimensional morphology of the disease: Based on a large number of calibrated radar maps and three-dimensional geometric feature sample data in the sample library, an end-to-end mapping model from radar maps to disease morphology features is constructed. Based on deep learning methods, the spatiotemporal features of radar images are automatically extracted to achieve high-precision prediction between radar image input and three-dimensional geometric parameters of the disease. The three-dimensional geometric parameters of the disease include at least spatial volume, penetration height and interface contact area. S6. Intelligent recognition of 3D morphology of road surface defects based on mapping model: In continuous road sections, real-time radar map data can be obtained through three-dimensional ground-penetrating radar scanning. Combined with the end-to-end mapping model from radar map to disease morphology features that has been trained, the three-dimensional geometric features of the disease area can be rapidly inverted and identified, achieving non-destructive spatial positioning and three-dimensional reconstruction of internal diseases of asphalt pavement.

2. The method for three-dimensional morphological identification of internal road surface defects based on radar and deep learning according to claim 1, characterized in that, In step S2, the high-performance polyurethane material includes a two-component polyurethane material, and the viscosity of each component material is 200~500 mPa·s, so as to ensure that the polyurethane material can better penetrate into the disease. After curing, its tensile strength is greater than 5 MPa, which facilitates rapid curing of the affected area and makes it easy to drill a complete core sample.

3. The method for three-dimensional morphological identification of internal road surface defects based on radar and deep learning according to claim 1, characterized in that, In step S2, the height of the core sample taken should at least include the thickness of the road surface layer and the upper and lower base layers, and the core sample diameter should be 100mm to 150mm.

4. The method for three-dimensional morphological identification of internal road surface defects based on radar and deep learning according to claim 1, characterized in that, In step S3, the CT scan uses an industrial-grade high-precision CT device with a resolution of not less than 30μm / voxel to ensure clear imaging of millimeter-level defects and cracks. The scanning time is kept within 10 minutes to prevent morphological errors caused by changes in the polyurethane and the environment.

5. The method for three-dimensional morphological identification of internal road surface defects based on radar and deep learning according to claim 1, characterized in that, In step S3, a deep semantic segmentation neural network based on the U-Net architecture is used to automatically segment the disease area of ​​the CT image. The network includes a 4-layer downsampling module and an upsampling module to effectively capture the boundary features of the polyurethane grouting area.

6. The method for three-dimensional morphological identification of internal road surface defects based on radar and deep learning according to claim 5, characterized in that, In step S3, when using deep learning or advanced image processing algorithms to filter, enhance, and segment images at the pixel level, in order to improve the robustness of the model for small target recognition, a weighted combination of the Dice loss function and the Binary Cross-Entropy loss function is used for optimization during the training phase.

7. The method for three-dimensional morphological identification of internal road surface defects based on radar and deep learning according to claim 5, characterized in that, In step S3, the training set is constructed through manual annotation. The model prediction output is a binary segmented image with the same size as the original image. The segmentation result of each slice image is output in the form of a binary image. The model prediction result is post-processed by morphological dilation-erosion to improve boundary continuity and segmentation accuracy. The segmentation contour error is controlled within 1 pixel to meet the spatial accuracy requirements of subsequent 3D reconstruction.

8. The method for three-dimensional morphological identification of internal road surface defects based on radar and deep learning according to claim 1, characterized in that, In step S4, the image reconstruction process should construct a three-dimensional voxel model based on the voxel image sequence, with a voxel accuracy of no more than 0.3 mm, to ensure that the precision of the geometric reconstruction meets the evaluation requirements.

9. The method for three-dimensional morphological identification of internal road surface defects based on radar and deep learning according to claim 1, characterized in that, In step S5, the geometric parameters of the morphological characteristics of the disease include at least the spatial volume, the maximum penetration height, the contact area between the disease and the structure, and the projected area in the length, width and height directions, so as to comprehensively characterize the spatial distribution pattern and development trend of the disease.

10. The method for three-dimensional morphological identification of internal road surface defects based on radar and deep learning according to claim 1, characterized in that, In step S6, the trained end-to-end mapping model is deployed to the radar detection map of the entire road, and the continuous radar data is processed by sliding window to extract the three-dimensional geometric shape parameters of the corresponding disease area segment by segment.

Images