A method, apparatus, and medium for modeling a damaged concrete structure

Abstract

The application provides a damaged concrete structure modeling method, device and medium, wherein the method comprises: acquiring an image sequence of a concrete structure surface, and generating a semantic point cloud; extracting a disease semantic point set based on the semantic point cloud, fitting a bounding box, and generating a binary point cloud slice image; performing contour reconstruction using the slice image, and generating a complete structure three-dimensional model; voxelizing the disease semantic point set and the complete model after merging, generating a semantic voxel model, establishing a three-dimensional digital mapping matrix after voxel-level correction, analyzing the matrix through finite element software, and generating a finite element grid model with disease information embedded. The application can solve the technical problems that the existing method relies on model registration to cause cumulative positioning errors and cannot synchronously process multiple types of complex diseases.

Description

Technical Field

[0001] This document relates to the field of modeling technology for damaged concrete structures, and in particular to a method, equipment and medium for modeling damaged concrete structures. Background Technology

[0002] Concrete structures are widely used in infrastructure fields such as bridges, tunnels, buildings, water conservancy projects, and ports. Under the combined effects of long-term loads, environmental erosion, and material degradation, these concrete structures exhibit varying degrees of surface defects. These defects begin on the structural surface but tend to accumulate and extend into the interior of the structure. Therefore, timely and accurate detection and assessment of these surface defects are crucial for the safe operation of concrete structures.

[0003] In recent years, related research has attempted to integrate intelligent defect detection with finite element analysis, using defect images and point cloud data obtained from detection to generate refined finite element geometric models of damaged concrete structures, thereby improving the reliability and accuracy of defect assessment.

[0004] However, the existing technologies still have the following shortcomings: First, directly converting the overall point cloud into a finite element model by voxelization or patching is highly dependent on the quality of the point cloud, resulting in high computational costs and difficulty in guaranteeing accuracy in complex areas; second, local geometric update methods rely on existing finite element basic models or manual modeling, which limits their application; third, disease embedding requires spatial registration, which can easily lead to cumulative positioning errors and affect the accuracy of the assessment; and fourth, existing methods mostly focus on single types of diseases and lack the ability to integrate and model composite diseases.

[0005] Therefore, there is a need for a method for modeling and evaluating damaged concrete structures that does not rely on high-precision point clouds, does not require manual modeling or existing finite element models, avoids spatial registration accumulation errors, and can simultaneously handle multiple types of complex defects. Summary of the Invention

[0006] This invention provides a method, equipment, and medium for modeling damaged concrete structures, aiming to solve the above-mentioned problems.

[0007] According to an embodiment of the present invention, a method for modeling damaged concrete structures is provided, comprising: S1. Obtain the image sequence of the concrete structure surface, and generate a semantic point cloud that contains both structural geometric information and defect semantic information by performing defect semantic segmentation and three-dimensional reconstruction. S2. Based on the semantic point cloud, extract the disease semantic point set, fit a bounding box to the semantic point cloud, and use the bounding box to generate a binarized point cloud slice image along the main axis of the structure. S3. Contour reconstruction is performed using the sliced ​​image to generate a intact three-dimensional structural model in a damage-free state; S4. After merging the disease semantic point set with the intact three-dimensional model using the bounding box, voxelize the result to generate a semantic voxel model. Perform voxel-level correction on the semantic voxel model, establish a three-dimensional digital mapping matrix, and parse the matrix using finite element software to generate a finite element mesh model with embedded disease information.

[0008] According to an embodiment of the present invention, an electronic device is provided, characterized in that it comprises: Processor; and, A memory is configured to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the damaged concrete structure modeling method described above.

[0009] According to an embodiment of the present invention, a storage medium is provided, characterized in that it is used to store computer-executable instructions, which, when executed, implement the steps of the above-described method for modeling damaged concrete structures.

[0010] This invention generates semantic point clouds by combining image semantic segmentation and 3D reconstruction techniques, without relying on other modal data. Disease points are extracted based on the CIELAB color space, requiring fewer parameters and exhibiting strong noise resistance. Bounding boxes provide a unified spatial reference, eliminating the need for registration and avoiding accumulated positioning errors, enabling efficient handling of complex diseases distributed across a plane. A complete structural model is obtained through contour reconstruction of 2D slice images, without relying on manual modeling and exhibiting good noise resistance. Disease and structural information are fused based on a voxel model, and voxel-level noise reduction and correction improve the accuracy of disease embedding. A 3D digital mapping matrix drives the automatic generation of finite element models, resulting in low computational cost and high real-time performance.

[0011] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of the damaged concrete structure modeling method according to an embodiment of the present invention; Figure 2This is a flowchart illustrating the overall process of generating embedded FE meshes for disease-related defects according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the slice image generation process according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the complete model generation process according to an embodiment of the present invention; Figure 5 This is a schematic diagram of line segment feature optimization according to an embodiment of the present invention; Figure 6 This is a schematic diagram of point feature optimization according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the voxel-based disease embedding process according to an embodiment of the present invention; Figure 8 This is a schematic diagram of voxel correction in an embodiment of the present invention; Figure 9 This is a schematic diagram of three-dimensional mapping matrix driven modeling according to an embodiment of the present invention; Figure 10 This is a schematic diagram of an RC bridge pier RGB image according to an embodiment of the present invention; Figure 11 This is a schematic diagram of the disease segmentation results according to an embodiment of the present invention; Figure 12 This is a schematic diagram of the semantic point cloud model of the RC beam according to an embodiment of the present invention; Figure 13 This is a schematic diagram of the disease semantic point extraction results in an embodiment of the present invention; Figure 14 This is a schematic diagram of cross-sectional reconstruction of a slice image of an RC bridge pier according to an embodiment of the present invention; Figure 15 This is a schematic diagram of the visualization results of a voxel correction example according to an embodiment of the present invention; Figure 16 This is a schematic diagram of the RC pier defects embedded in the FE mesh model according to an embodiment of the present invention. Detailed Implementation

[0014] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0015] Method Implementation Examples According to an embodiment of the present invention, a method for modeling damaged concrete structures is provided. Figure 1This is a flowchart of the damaged concrete structure modeling method according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the overall process of generating embedded FE meshes for defects according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, the method for modeling damaged concrete structures according to an embodiment of the present invention specifically includes: S1. Obtain an image sequence of the concrete structure surface, and generate a semantic point cloud that simultaneously contains structural geometric information and defect semantic information by performing defect semantic segmentation and 3D reconstruction. Specifically, S1 includes: First, image acquisition is performed by taking pictures around the concrete structure with a camera to obtain a sequence of images of the structural surface. The overlap rate between adjacent images is maintained at more than 70% to ensure the quality and accuracy of subsequent 3D reconstruction. The camera can be a drone or a handheld camera.

[0016] The image sequence is processed using the Structure from Motion (SfM-MVS) algorithm to obtain camera parameters and depth maps. Specifically, the SfM algorithm is used to recover the camera's pose and position parameters from the image sequence, while the MVS algorithm generates a dense depth map for each image. Combining these two algorithms yields the camera's intrinsic and extrinsic parameters for each image, as well as a depth map containing dense depth information, while simultaneously generating an initial point cloud model of the structure.

[0017] The image sequence is semantically segmented using a disease segmentation neural network to generate semantic images. Specifically, a pre-trained disease segmentation neural network is used to process each original RGB image. This neural network can employ deep learning models such as SegFormer, DeepLabv3+, or U-Net. During processing, disease pixels in the image are labeled with preset specific colors, such as red for crack pixels, yellow for peeling pixels, and magenta for exposed ridge corrosion pixels, thereby completing disease segmentation and generating semantic images that correspond one-to-one with the original images.

[0018] The semantic image replaces the original image, and the depth map fusion is performed again while keeping the camera parameters and depth map unchanged, so that the 3D reconstructed points inherit the pixel colors in the semantic image, thereby generating the semantic point cloud.

[0019] Specifically, the original RGB image is replaced with the semantic image generated in the previous steps, while keeping the camera parameters and depth map calculated in the previous steps unchanged. The depth map fusion calculation is then performed again. During this process, the color information of each 3D point in the initial point cloud model is updated from the RGB color of the original image to the disease color marked in the corresponding semantic image. This allows the 3D points to inherit the disease category label, completing the mapping from 2D semantic segmentation to 3D space disease mapping. Finally, a semantic point cloud is generated that simultaneously contains structural geometric information (3D coordinates) and disease semantic information (color labels).

[0020] S2. Based on the semantic point cloud, extract the disease semantic point set, fit a bounding box to the semantic point cloud, and use the bounding box to generate a binarized point cloud slice image along the main axis of the structure. S2 specifically includes: S21. Using the RGB color of the target disease as the target color, convert it to the CIELAB color space through a mapping function, calculate the Euclidean distance between the color of each point in the semantic point cloud and the target color, and when the Euclidean distance is less than a preset threshold, determine the point as a disease semantic point and extract it. Specifically, after step S1, the diseased points have been marked with a uniform semantic color, which serves as the target color. Extract semantic points of the disease.

[0021] Calculate the Euclidean distance between the color information of each point and the target color in the semantic point cloud in the CIELAB color space. and with preset threshold Comparison. When Less than When this point is identified as a disease semantic point, it is extracted. The mathematical expression of this process is as follows: ; ; ; in, This represents a mapping function that converts RGB color space to CIELAB color space. The semantic point cloud represents the first i The RGB color of each point The target RGB color.

[0022] Different disease types correspond to different target colors and distance thresholds. For example, the target color of a crack has an RGB value of [128,0,0], and the distance threshold... Set to 25. By setting different target colors and threshold parameters, this method can simultaneously extract semantic points of multiple types of diseases, achieving parallel processing of complex diseases.

[0023] S22, Fit the minimum directed bounding box to the semantic point cloud; Specifically, the get_minimal_oriented_bounding_box() function from the OPEN3D library is used to generate a minimum oriented bounding box for the semantic point cloud, with all parameters of the function kept at their default values.

[0024] Specifically, the generated bounding box can closely fit the overall outline of the point cloud, and its eight spatial corner coordinates record the boundary position of the point cloud in three-dimensional space.

[0025] This bounding box provides a unified spatial reference system for subsequent merging of disease semantic points with intact models, restoration of 3D coordinates after slice image contour processing, and generation of 3D voxel models, eliminating the need for additional model registration or coordinate transformation, thus avoiding accumulated positioning errors.

[0026] S23. Voxelize within the bounding box at a set resolution to generate a voxel model. Extract each voxel slice along the main axis of the structure. Assign the first pixel value to voxels containing three-dimensional points and the second pixel value to voxels not containing three-dimensional points to generate a binarized point cloud slice image. Figure 3 This is a schematic diagram of the slice image generation process according to an embodiment of the present invention.

[0027] Specifically, the semantic point cloud within the bounding box is first voxelized at a set resolution to generate a rectangular voxel model with a fixed resolution.

[0028] Depending on whether each voxel contains a 3D point, each voxel is assigned one of two states: occupied or empty. Voxels containing 3D points are marked as occupied, and voxels not containing 3D points are marked as empty.

[0029] Then, voxel slices of each layer are extracted along the direction of the maximum size of the structure, i.e., the principal axis of the structure. The extracted voxel slices are converted into two-dimensional images through binarization. The pixel value corresponding to the occupied voxel is set to 255 as the first pixel value, and the pixel value corresponding to the empty voxel is set to 0 as the second pixel value, thus generating a binarized point cloud slice image. The number and resolution of the slice images are determined by the bounding box size and the set voxel resolution.

[0030] S3. Contour reconstruction is performed using the sliced ​​image to generate a intact three-dimensional structural model in a damage-free state; Figure 4 This is a schematic diagram of the complete model generation process according to an embodiment of the present invention. Specifically, step S3 includes: S31. Perform concave hull detection on the binarized point cloud slice image to generate a closed polygon that fits the original contour and fill it with solid to obtain a solid region. Specifically, the `concave_hull()` function from the Shapely library is used to identify feature points of the contour on the sliced ​​image. The concavity ratio parameter of this function is set to 1 to identify feature points of the contour on the sliced ​​image and generate a closed polygon that fits the original contour. Then, the area enclosed by the closed polygon is filled with solid fill to obtain a solid region. M The solid area represents the complete cross-section of the concrete structure at the slice location, where the area with a pixel value of 255 is the structural solid area, and the area with a pixel value of 0 is the background or the area with missing defects.

[0031] S32. Fit a circumscribed regular polygon based on the geometry of the solid region as the basic contour; Specifically, a circumscribed regular polygon is fitted based on the geometry of the solid region. For example, when the solid region is approximately rectangular, a circumscribed rectangle is fitted; when the solid region is approximately circular, a circumscribed circle is fitted. Simultaneously, the geometric feature points of this polygon, such as corner coordinates, are obtained. This regular polygon serves as the basic contour in subsequent contour optimization steps. The total number of corner points of the basic contour is denoted as... n Each corner point is arranged relative to the center point. p c The polar angles are sorted and numbered to define the adjacency relationships between corner points. The selection of the basic contour is not limited to a specific type. As long as the geometric primitives of the basic contour, such as line segments, corner points, or arcs, can be defined, and the overlap relationship between them and the solid region can be evaluated, the same optimization strategy can be used for contour refinement.

[0032] S33. The basic contour is iteratively adjusted through line segment feature optimization and corner point feature optimization to generate the optimal cross-sectional contour; Specifically, the line segment feature optimization adjusts the position of each edge based on the overlap between each edge of the basic contour and the solid fill area (pixel value 255), such as... Figure 5 As shown. If the overlap of a certain edge is insufficient, it is translated inward along its normal direction towards the geometric center of the section. After all edges are adjusted, a refined profile is obtained. Through continuous iteration, the profile is gradually adjusted to the outer boundary of the actual concrete structure section.

[0033] The specific steps for adjusting the feature position of each line segment are as follows: 1. Determining line segment characteristics. Based on the position of each corner point of the basic contour relative to the center point. p c The polar angles are sorted and numbered to define the adjacency relationships between corner points. Therefore, the solid region... M And the sorted point pairs ( p i , p i+1 The current line segment is defined.L i It can be represented as: ; in, t These are parameters used to determine the position of points on a line segment. n This represents the total number of corner points of the basic outline.

[0034] 2. Overlap rate calculation. Based on the filled area. M With line segment L i The overlap rate between the two pixels is calculated by taking the number of pixels they occupy together, as shown in the following expression: ; 3. Determine the direction of line segment movement. The overlap rate is below a specific threshold. When, line segment L i The line segment is gradually translated along the geometric center of the basic contour, and the overlap rate is recalculated after each translation step. When the overlap rate reaches a threshold, the line segment stops moving. Therefore, using the center point... p c and p c Online segment L i Projection point on p f The vector formed determines the direction of movement. Its expression is as follows: ; ; 4. Calculation of endpoint coordinates. After the line segment is translated, the coordinates of its two endpoints change, and are calculated using the following formula: ; ; in, d This indicates the translation magnitude applied during each movement.

[0035] After optimizing the line segment features as described above, further point feature optimization is performed to eliminate the small, approximately triangular gaps between the contour and the solid area near the corner points, thereby improving the fitting accuracy of the contour in the corner point region. The basic idea is to generate two new points for each corner point and determine the positions of the two new corner points based on the positions of the pixel values ​​of the corresponding contour edges that change abruptly. The two new corner points replace the original corner points, thus updating the contour shape near the corner points, such as... Figure 6 As shown. The specific steps for corner feature optimization are as follows: 1. Direction vector determination. Based on the current corner point. p i and two neighboring pointsp i+1 , p i-1 Determine two vectors v1( p i+1 , p i ) and v2( p i-1 , p i ), and normalize to obtain two direction vectors. 1 and 2, its expression is as follows: ; 2. Calculate the minimum distance. Using 1 and 2. Determine the corner points Minimum distance to the nearest pixel in the filled region and This refers to the distance traveled from the 0-value pixel to the initial position of pixel 255, expressed as follows: ; ; 3. Calculate the coordinates of the candidate corner points. Using the direction vectors and minimum distances obtained earlier, the coordinates of two candidate corner points are initially determined, as shown in the following expression: ; 4. Triangle Area Calculation. After completing steps 1-3 for all corner points, calculate the area of ​​each corner point and the area of ​​the triangle formed by it and two candidate corner points. The area of ​​each triangle is calculated using vectors. and The calculation is expressed as follows: ; 5. Chamfer Evaluation. Considering that the gaps between triangles at corner points may also originate from reconstruction errors or surface defects, it is necessary to evaluate whether the chamfer features formed by these triangles can be considered as genuine structural chamfers. The average relative error of all triangle areas calculated using the triangle area is used as the discrimination index, and the average value of non-zero triangle areas is used as the standard value. If the calculated triangle areas at three or more corner points are non-zero, and their average relative error is lower than a preset threshold, then these chamfer features are considered valid. In this case, candidate corner points will replace the original corner points for contour reshaping. Otherwise, the point feature optimization step is skipped.

[0036] 6. Determining the Optimal Chamfer. Most practical civil engineering structures have cross-sections with a certain degree of symmetry, and the chamfers at all corners should maintain a uniform size. The triangle with the smallest area is selected as the optimal reference for generating the structural chamfer, and the chamfer shape at all other corners is updated using the minimum distance between the two candidate corner points corresponding to this triangle, thus completing the point feature optimization.

[0037] 7. Solid Fill. The optimized contour-enclosed area is filled with solid fill to complete the reconstruction of the solid cross-section of the structure.

[0038] S34. Using the spatial corner coordinates of the bounding box, recover the three-dimensional coordinates for all reconstructed optimal cross-sectional contours to generate a complete three-dimensional structural model.

[0039] Specifically, using the coordinates of the eight 3D spatial corner points of the bounding box obtained in step S22 as a reference, the 3D coordinates of each pixel in the reconstructed section are calculated, thereby restoring the separated 2D slice images into a 3D point cloud and generating a complete 3D structural model in a damage-free state. This complete 3D structural model eliminates the missing information in the diseased area and constructs a complete, damage-free structural geometric model, which serves as the basis model for subsequent disease embedding.

[0040] S4. After merging the disease semantic point set with the intact structural 3D model using the bounding box, voxelize the result to generate a semantic voxel model. Perform voxel-level correction on the semantic voxel model, establish a 3D digital mapping matrix, and parse the matrix using finite element software to generate a finite element mesh model with embedded disease information. S4 specifically includes: S41. Using the bounding box, merge the disease semantic point set with the intact three-dimensional model, and voxelize it according to the set resolution to generate a semantic voxel model. Figure 7 This is a schematic diagram of the voxel-based disease embedding process according to an embodiment of the present invention.

[0041] Specifically, the intact model generated in step S34 and the disease semantic point set extracted in step S21 are read together and merged into the bounding box space obtained by fitting in step S22. Then, the merged point cloud model is processed. PC m Voxelization is performed at a set resolution to generate a merged voxel model. This merged voxel model simultaneously contains concrete information of the intact structure and semantic information of defects, laying the foundation for subsequent voxel-level fusion processing.

[0042] S42. Perform voxel denoising and voxel correction. The voxel denoising is used to remove intact model points from mixed voxels that simultaneously contain intact model points and disease semantic points. The voxel correction is used to correct the voxels on the line connecting the disease voxel and the nearest contour voxel to the disease category. In a merged voxel model, some voxels simultaneously contain structural concrete information and defect semantic information, representing the intact model. This causes the voxel colors to deviate from the values ​​when only defect semantic information is included, affecting the accuracy of subsequent 3D mapping matrix generation. Therefore, concrete 3D points in such mixed voxels containing both types of information are considered noise points. P N and from the merged point cloud model PC m Remove from the image to ensure uniform color of the diseased voxels. The mathematical representation of the denoising process is as follows: ; in, Indicates from point x The function that indexes the voxel it contains. V D , V S and V N These represent the sets of voxels related to disease, concrete, and noise, respectively.

[0043] Areas surrounded by defects (especially spalling voxels), which are areas of missing concrete, are misclassified as healthy concrete voxels in the model. Therefore, these voxels need to be corrected, such as... Figure 8 As shown. The specific steps are as follows: 1. Define the search region. For each layer of the voxel model, traverse all disease voxels in that layer. For each disease voxel... D v Centered on the location, construct with a radius of R spherical search area S Its mathematical definition is as follows: ; 2. Extract target contour voxels. The `findContours()` function from the OpenCV library is used to detect the outermost contour points of the current layer, and the voxels containing these contour points are extracted as contour voxels. Then, the `query_ball_points()` function from the SciPy library is used to identify contour voxels falling within the search area. This function has a radius... R Perform a local query, define the identified voxels as target contour voxels, and add them to the target contour voxel set. C v .

[0044] 3. Determine the nearest contour voxel. Calculate all target contour voxels down to the current defect voxel. D v The European distance, the distance D vThe most recent target contour voxel is identified as the most recent contour voxel. N cv Its mathematical definition is as follows: ; in, c i For the target contour voxel set C v The first in i Individual factors.

[0045] 4. Voxel type change. (Using voxels...) N cv and D v Construct line segments using each as an endpoint. L Voxels located on this line segment will be identified as misclassified voxels. M v and update its category to match D v This ensures consistency, thus completing the voxel correction. The mathematical definition of this step is as follows: ; in, T ( v ) indicates the use of voxels for acquisition v Functions of type [type].

[0046] After completing the voxel correction, the background space within the bounding box space is also voxelized to generate the final voxel containing multiple components representing structural concrete, defects, and background.

[0047] S43. Create an empty array matrix of the same size as the semantic voxel model, and numerically encode each voxel according to its color or semantic label to generate a three-dimensional digital mapping matrix.

[0048] Specifically, first, a three-dimensional empty array matrix of the same size as the semantic voxel model is created to store the encoded values ​​of different component voxels. Then, the voxels in each voxel slice of the three-dimensional voxel volume are traversed, and a specific numerical code is assigned to each voxel based on its color information. Pixels corresponding to background voxels are assigned a value of 0, pixels corresponding to concrete voxels are assigned a value of 1, and voxels for various diseases are assigned values ​​of 2, 3, 4, etc., respectively. If the model contains multiple types of diseases, different diseases are assigned different encoded values ​​that correspond one-to-one with their colors, thereby generating a three-dimensional numerical mapping matrix. Figure 9 This is a schematic diagram of three-dimensional mapping matrix driven modeling according to an embodiment of the present invention.

[0049] S44. Read the three-dimensional digital mapping matrix through the application programming interface of the finite element software, delete the background unit according to the encoding value, add the concrete unit and various disease units to the corresponding set respectively, and generate a finite element mesh model with embedded disease information.

[0050] Specifically, firstly, a rectangular geometry with the same dimensions as the 3D voxel is created using ABAQUS finite element software. The rectangular geometry is then divided according to the voxel mesh resolution, and hexahedral elements are used to mesh the geometry, generating the basic finite element mesh model. Next, a 3D digital mapping matrix is ​​read in through the ABAQUS Python application interface. The numerical codes in the 3D matrix correspond one-to-one with the elements at the same location. Elements corresponding to code 0 are identified as background elements, excluded from calculation, and deleted. Elements corresponding to code 1 are identified as concrete elements and added to the concrete element set. Elements with codes 2 and above are identified as various types of defect elements and added to their corresponding defect element sets, ultimately generating a finite element mesh model with embedded defect information.

[0051] Furthermore, in the three-dimensional digital mapping matrix, different disease types are assigned different numerical codes, and the elements corresponding to different codes in the finite element mesh model are automatically grouped into element sets of different disease types, thereby realizing the distinguishable modeling of multiple types of diseases.

[0052] To more clearly illustrate the technical solution and effects of the present invention, a specific reinforced concrete bridge pier is used as an example for detailed explanation below. The dimensions of this bridge pier, L×W×H, are 2350 mm × 700 mm × 2500 mm. It should be noted that this example is only used to illustrate the principles and implementation methods of the present invention and should not be considered as a limitation thereof.

[0053] The RC bridge pier is modeled and evaluated according to steps S1 to S4 above. The specific process is as follows: In step S1, a DJI Phantom 4 RTK quadcopter drone was used to fly around the bridge pier, collecting 412 clear images from different heights and angles as an image sequence. For example... Figure 10 The image shown is an RGB image of an RC bridge pier. A trained multi-defect segmentation model is used to process the RGB image of the RC bridge pier, as follows: Figure 11The diagram illustrates the segmentation results of the defects. Crack pixels in the original image are marked in red (RGB value [128,0,0]), peeling pixels in yellow (RGB value [255,195,0]), and exposed reinforcement / corrosion pixels in magenta (RGB value [195,0,195]), outputting a semantic image of the defects. Then, the SfM-MVS algorithm is used to process the RGB image sequence of the RC pier, obtaining the camera intrinsic and extrinsic parameters and a depth map containing dense depth information for each image, generating an initial point cloud model of the structure. Finally, the original RC pier image is replaced with the semantic image of the defects, maintaining the original camera parameters and depth map, and the depth map fusion calculation is re-performed to generate the semantic point cloud model of the RC pier.

[0054] In step S2, semantic points of defects are extracted according to the parameters shown in Table 1. The target color for cracks is [128,0,0] with a distance threshold of 25; the target color for spalling is [255,195,0] with a distance threshold of 25; and the target color for exposed rebar corrosion is [195,0,195] with a distance threshold of 32. When generating bounding boxes for the semantic point clouds of RC piers, the parameters of the get_minimal_oriented_bounding_box() function are kept at their default values. The space within the bounding box is voxelized at a resolution of 1 cm (i.e., a voxel side length of 1 cm), and 251 slice images are generated along the principal axis, i.e., the beam length direction. The model size error is controlled within one voxel, i.e., approximately 1 cm. Figure 12 This is a schematic diagram of the semantic point cloud model of an RC beam. Figure 13 This is a schematic diagram of the semantic point extraction results for the disease.

[0055] Table 1 Threshold selection for semantic point extraction parameters of disease

[0056] In step S3, the concave_hull() function is used to perform concave hull detection on the sliced ​​image, and the concaveness ratio parameter is set to 1. The solid region of the RC pier is approximately rectangular, so a fitted regular bounding rectangle is used as the basic contour. In the line segment feature optimization, an overlap rate threshold is set. The translation amplitude is 0.75 for each step. d The distance is 1 pixel. In corner feature optimization, the preset threshold for the average relative error is set to 0.3. After optimization, the absolute and relative errors between the length L and width W of all reconstructed sections and the measured values ​​are calculated. The average absolute error of length L is 0.923 cm, and the average relative error is 0.007; the average absolute error of width W is 0.754 cm, and the average relative error is 0.018. The three-dimensional coordinates are recovered using the eight 3D spatial corner coordinates of the bounding box, generating a complete model of the structure. Figure 14 This is a schematic diagram of cross-sectional reconstruction from a slice image of an RC bridge pier. Figure 15 A schematic diagram illustrating the results of voxel correction.

[0057] In step S4, the intact model and the set of semantic points of defects are merged and then voxelized. During voxel correction, the search radius R is set to 5 cm, correcting a total of 255 misidentified voxels. A three-dimensional empty array matrix of the same size as the semantic voxel model is created. In the RC pier, background voxels correspond to code 0, concrete voxels to code 1, crack voxels to code 2, spalling voxels to code 3, and exposed rebar corrosion voxels to code 4. A rectangular geometry of 2350 mm × 700 mm × 2500 mm is constructed using ABAQUS software, and a finite element basic mesh model is generated using hexahedral elements with a side length of 1 cm. The three-dimensional digital mapping matrix is ​​read in through the ABAQUS Python application interface, and background elements are deleted according to the coding values. Concrete elements and various defect elements are added to their corresponding sets. Figure 16 A schematic diagram of embedding FE mesh model for defects in RC bridge piers.

[0058] For example, in the RC bridge pier example, 2849 crack elements, 637 spalling elements, and 9 corrosion elements were ultimately identified. The relative error between the spalling area volume in the finite element model and the actual spalling volume was 3.1%. The finite element software used is not limited to ABAQUS; other modeling software or platforms with Python data interfaces and similar modeling logic are also applicable.

[0059] The above examples verify that the method proposed in this invention can effectively handle multiple types of complex diseases, automatically generate finite element mesh models with embedded disease information, and has high geometric reconstruction accuracy and disease identification accuracy.

[0060] The following are the specific beneficial effects of using the embodiments of the present invention: Semantic point clouds are generated by combining image semantic segmentation and 3D reconstruction techniques, without relying on other modal data. Disease points are extracted based on the CIELAB color space, requiring fewer parameters and exhibiting strong noise resistance. Bounding boxes provide a unified spatial reference, eliminating the need for registration and avoiding accumulated positioning errors, enabling efficient handling of complex diseases distributed across a plane. A complete structural model is obtained through contour reconstruction of 2D slice images, without relying on manual modeling and exhibiting good noise resistance. Disease and structural information are fused based on voxel models, and voxel-level noise reduction and correction improve the accuracy of disease embedding. The automatic generation of finite element models is driven by a 3D digital mapping matrix, resulting in low computational cost and high real-time performance.

[0061] Device Example 1 According to an embodiment of the present invention, an electronic device is provided, characterized in that it comprises: Processor; and, A memory is configured to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the damaged concrete structure modeling method described above.

[0062] Device Example 2 According to an embodiment of the present invention, a storage medium is provided, characterized in that it is used to store computer-executable instructions, which, when executed, implement the steps of the above-described method for modeling damaged concrete structures.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for modeling damaged concrete structures, characterized in that, include: S1. Obtain the image sequence of the concrete structure surface, and generate a semantic point cloud that contains both structural geometric information and defect semantic information by performing defect semantic segmentation and three-dimensional reconstruction. S2. Based on the semantic point cloud, extract the disease semantic point set, fit a bounding box to the semantic point cloud, and use the bounding box to generate a binarized point cloud slice image along the main axis of the structure. S3. Contour reconstruction is performed using the sliced ​​image to generate a intact three-dimensional structural model in a damage-free state; S4. After merging the disease semantic point set with the intact three-dimensional model using the bounding box, voxelize the result to generate a semantic voxel model. Perform voxel-level correction on the semantic voxel model, establish a three-dimensional digital mapping matrix, and parse the matrix using finite element software to generate a finite element mesh model with embedded disease information.

2. The method according to claim 1, characterized in that, S1 specifically includes: The image sequence is processed using a motion-reconstruction structure multi-view stereo vision algorithm to obtain camera parameters and depth maps; The image sequence is semantically segmented using a disease segmentation neural network to generate a semantic image. The semantic image replaces the original image, and the depth map fusion is performed again while keeping the camera parameters and depth map unchanged, so that the 3D reconstructed points inherit the pixel colors in the semantic image, thereby generating the semantic point cloud.

3. The method according to claim 1, characterized in that, S2 specifically includes: The RGB color of the target disease is used as the target color. It is converted to the CIELAB color space through a mapping function. The Euclidean distance between the color of each point in the semantic point cloud and the target color is calculated. When the Euclidean distance is less than a preset threshold, the point is identified as a disease semantic point and extracted. Fit the minimum directed bounding box to the semantic point cloud; Within the bounding box, voxels are generated at a set resolution to form a voxel model. Each voxel slice is extracted along the main axis of the structure. Voxels containing three-dimensional points are assigned the first pixel value, and voxels not containing three-dimensional points are assigned the second pixel value to generate a binarized point cloud slice image.

4. The method according to claim 1, characterized in that, S3 specifically includes: Perform concave hull detection on the binarized point cloud slice image to generate a closed polygon that fits the original contour and fill it with solid to obtain a solid region; A circumscribed regular polygon is fitted based on the geometry of the solid region as the basic contour. The basic contour is iteratively adjusted by optimizing line segment features and corner point features to generate the optimal cross-sectional contour. Using the spatial corner coordinates of the bounding box, the three-dimensional coordinates are recovered for all reconstructed optimal cross-sectional profiles, generating a complete three-dimensional structural model.

5. The method according to claim 1, characterized in that, S4 specifically includes: The bounding box is used to merge the disease semantic point set with the intact three-dimensional model, and voxelization is performed at a set resolution to generate a semantic voxel model. Voxel denoising and voxel correction are performed. Voxel denoising is used to remove intact model points from mixed voxels that simultaneously contain intact model points and disease semantic points. Voxel correction is used to correct voxels on the line connecting disease voxels and the nearest contour voxel to the disease category. Create an empty array matrix of the same size as the semantic voxel model, and generate a three-dimensional digital mapping matrix by numerically encoding each voxel according to its color or semantic label. The three-dimensional digital mapping matrix is ​​read through the application programming interface of the finite element software. Background elements are deleted according to the encoded values, and concrete elements and various disease elements are added to the corresponding sets to generate a finite element mesh model with embedded disease information.

6. The method according to claim 4, characterized in that, The line segment feature optimization includes: Calculate the overlap rate between each geometric edge of the basic contour and the solid region, where the overlap rate is the ratio of the number of pixels occupied by the geometric edge and the solid region to the total number of pixels of the geometric edge. When the overlap rate is lower than a preset threshold, the geometric edge is gradually translated along the direction toward the geometric center of the cross section. Each translation is a preset amount, and the overlap rate is recalculated after each translation step until the overlap rate reaches the threshold and the translation stops.

7. The method according to claim 4, characterized in that, The corner feature optimization includes: Starting from the current corner point, search outward pixel by pixel along the direction of the two contour edges, and determine the positions where the pixel value first changes from an empty pixel to a solid pixel as two candidate corner points; Calculate the area of ​​the triangle formed by the original corner point and the two candidate corner points; When the calculated triangle areas at multiple corner points meet the condition that the average relative error is below a preset threshold, it is determined to be a valid structural chamfer. The candidate corner point position corresponding to the triangle with the smallest area is selected, and the coordinates of all corner points are updated.

8. The method according to claim 5, characterized in that, The voxel correction includes: Construct a spherical search area centered on each diseased element; The outermost contour points of the current voxel layer are detected, the voxels containing these contour points are extracted as contour voxels, and the contour voxels falling into the spherical search area are identified as target contour voxels. Calculate the Euclidean distance from each target contour voxel to the current disease voxel, and determine the voxel with the closest distance as the nearest contour voxel; Construct a line segment with the current disease voxel and the nearest contour voxel as endpoints, and correct the category of all voxels that the line segment passes through to the same disease category as the current disease voxel.

9. An electronic device, characterized in that, include: processor; as well as, A memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the modeling method for damaged concrete structures as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, Used to store computer-executable instructions, which, when executed, implement the steps of the damaged concrete structure modeling method as described in any one of claims 1 to 8.

Images