A large model-based rockfill skeleton modeling method

Abstract

The application discloses a rockfill skeleton modeling method based on a large model, relates to the technical field of three-dimensional modeling, and solves the problem that existing rockfill skeleton modeling is mostly dependent on simplified geometric body filling or random algorithm model generation and is difficult to reflect the diversity and complexity of the block stone morphology in nature. The rockfill skeleton modeling method comprises the following steps: image collection of rockfills with different morphologies from different angles to obtain multi-view rockfill images; construction of a plurality of rockfill particle models according to the multi-view rockfill images, simultaneous measurement of geometric characteristic parameters of all views of the rockfills, and obtaining of a particle-feature atlas; according to the rockfill skeleton image of a dam, adoption of a skeleton structure reasoning mechanism in combination with the geometric characteristic parameters of all views of the rockfills to infer all rockfill contact surfaces in the rockfill skeleton, and construction of a contact surface data set; construction of complete solid rockfill particles according to the contact surface data set in combination with the rockfill particle models, and completion of rockfill skeleton modeling.

Description

Technical Field

[0001] This invention relates to the field of 3D modeling technology, and in particular to a method for modeling a rockfill skeleton based on a large model. Background Technology

[0002] Rockfill skeleton modeling is a key technology in fields such as hydraulic engineering, geological engineering, and geological disaster prevention. Its core objective is to provide a reliable basis for engineering stability analysis and disaster prediction by digitally representing the shape, fragmentation characteristics, and accumulation structure of the rocks. However, traditional modeling methods face multiple technical bottlenecks, necessitating the introduction of large-scale modeling techniques for innovation. With breakthroughs in generative AI and multimodal large language models (MLLM), the advantages of large-scale models in complex data processing and automated generation are becoming increasingly apparent. For example, the combination of diffusion models and autoregressive frameworks can generate high-fidelity 3D structures, while visual tokenizers (such as VQGAN) can discretize images into semantic tokens, providing a new path for large-scale models to understand rockfill morphology.

[0003] In engineering practice, rockfill modeling needs to balance accuracy and efficiency. For example, seepage stability analysis of rockfill dams in hydraulic engineering requires models that can accurately depict the pore structure of the rocks; geological disaster prevention requires rapid assessment of the mechanical behavior of collapsed rockfill masses. Traditional methods are difficult to meet these needs, while large-scale models, by integrating physical simulation and data-driven technologies, can construct intelligent modeling frameworks that combine interpretability and generalization capabilities. For example, combining generative adversarial networks (GANs) and discrete element methods can generate high-fidelity numerical models of rockfill masses and simulate their fragmentation dynamics.

[0004] Patent No. CN202410941704.2 discloses a system for analyzing the rationality of a rockfill skeleton based on digital twins, relating to the field of electronic digital data processing. The system includes a data acquisition module, a digital modeling module, a skeleton analysis module, and a rationality analysis module. The data acquisition module collects actual rockfill information. The digital modeling module constructs a digital twin model of the rockfill based on the collected information. The skeleton analysis module extracts skeleton information from the digital twin model. The rationality analysis module analyzes the rationality of the rockfill based on the skeleton information. This system can perform stress analysis on the rockfill structure and obtain stability information based on the stress conditions, allowing for timely adjustments to the rockfill and reducing the probability of hazards.

[0005] Although the systems used in the above patents can complete the modeling of the rockfill skeleton, existing rockfill skeleton modeling mostly relies on simplified geometric filling or random algorithm generation of models, which makes it difficult to reflect the contact state of all the internal blocks during the dam construction process, resulting in low dam stability and the rockfill skeleton being prone to collapse and breakage. Summary of the Invention

[0006] The purpose of this invention is to provide a large-scale modeling method for rockfill skeletons, which can enhance the evolution of rockfill particles through a large model, and infer and display the contact surfaces of rockfill particles inside the rockfill skeleton, thereby demonstrating the diversity and complexity of the rock morphology in the rockfill skeleton and verifying the robustness of the rockfill skeleton of the dam.

[0007] This invention utilizes the following technical solution: A method for modeling a riprap skeleton based on a large model includes the following steps in sequence; S1: Collect images of rock piles of different shapes from different angles to obtain multi-view images of rock piles; S2: Construct several riprap particle models based on multi-view riprap images, and simultaneously calculate the geometric feature parameters of all views of the riprap to obtain a particle-feature map. S3: Based on the rockfill skeleton image of the dam, the skeleton structure inference mechanism is used in combination with the geometric feature parameters of all views of the rockfill to infer all rockfill contact surfaces inside the rockfill skeleton and construct a contact surface dataset. S4: Based on the contact surface dataset and the riprap particle model, construct a complete solid riprap particle model, and then complete the riprap skeleton modeling.

[0008] Preferably, in step S1, an oblique photography device is used to acquire images of the rock piles of different shapes from different angles to obtain multi-view rock pile images; at the same time, based on the image acquisition time and the rock pile view, combined with N-bit random numbers, a digital view identifier is generated and assigned to all multi-view rock pile images to obtain a digital image identifier table.

[0009] Preferably, step S2 includes the following steps: S21: Use an edge enhancement algorithm to calculate the gradient of the multi-view rock pile image and extract the rock pile edges from different viewpoints; S22: Using an optical flow algorithm combined with a digital image identification table, several rock pile edges are uniquely combined to obtain initial rock pile point cloud data; S23: Use graph neural network algorithm to detect interference points in the initial rockfill point cloud data, and then remove abnormal points to obtain the main rockfill point cloud data; S24: Based on the geological statistical laws, the Diffusion model is used to enhance the point cloud data of the main rockfill point cloud data to obtain polymorphic rockfill point cloud data. S25: Use a non-aliased adversarial network to decouple and controllably generate the polymorphic riprap point cloud data to obtain several riprap view point cloud data. S26: Construct several riprap particle models based on the multi-view point cloud data of the riprap using a 3D reconstruction algorithm; S27: Perform multidimensional characterization calculations on all riprap particle models and riprap view point cloud data in sequence to obtain the geometric feature parameters of all riprap views; Geometric characteristic parameters include surface area, volume, specific surface area, flatness, elongation, roundness, sphericity, and three-dimensional fractal dimension; S28: Based on the digital image identification table, combined with the contact relationship between particles and the path of mechanical property transmission, the geometric feature parameters and the rockfill particle model are associated using a knowledge graph to obtain a particle-feature graph.

[0010] Preferably, step S3 includes the following steps: S31: Aerial photography of the dam during its construction process was conducted using drones to obtain several images of the rockfill skeleton. S32: The rockfill skeleton image is stitched together using an optical fluid algorithm to obtain the main image of the dam; S33: Perform principal component analysis on the main image of the dam, calculate the principal vector of the rockfill skeleton, and obtain the axial-particle packing correlation diagram of the dam body; S34: Based on the particle-feature map and the principal vector of the rockfill skeleton, combined with the dam body axial-particle accumulation correlation diagram, the morphology of the rockfill skeleton is obtained; S35: Based on the shape of the rockfill skeleton along the main direction, the rockfill skeleton is sliced ​​into layers according to a preset fixed thickness, the volume ratio of rockfill particles in each layer is calculated, and a density distribution heat map is generated. S36: A three-dimensional density field is constructed using the Kriging interpolation algorithm, and potential pore concentration areas and dense contact surface areas in the dam's main image are identified based on the density distribution heatmap; S37: A graph attention network is used to construct an initial rockfill map structure based on the geometric feature parameters and particle spatial coordinates in the particle-feature map. S38: The contact probability between pile particles is calculated by using a skeleton structure reasoning mechanism combined with the initial pile diagram structure. S39: Based on the contact probability and preset contact threshold combined with the particle space coordinates, determine all rockfill contact surfaces within the rockfill skeleton, and calculate the contact surface normal vector and contact area.

[0011] Preferably, the operation flow of the skeleton structure reasoning mechanism is as follows: A: The watershed algorithm is used to perform three-dimensional edge segmentation on the overlapping areas of the rockfill skeleton image based on the potential pore concentration area and the dense contact surface area, so as to obtain the inter-particle overlapping domain features and rockfill overlapping point cloud. B: Morphological calculations are performed on the features of the overlapping regions between particles to obtain the area and volume of the overlapping regions. At the same time, a point cloud segmentation algorithm is used to segment the overlapping point cloud of the rock pile according to a preset overlapping threshold to obtain several particle contact point clouds. C: Calculate the three-dimensional fractal dimension of each particle contact point cloud, and at the same time calculate the morphological matching degree between the particle contact point cloud and the view geometric feature point cloud, and judge it with the preset morphological threshold. D: If the shape matching degree is less than the shape threshold, change the view geometric feature point cloud and compare again until the shape matching degree of the two is greater than or equal to the shape threshold. If all view geometric feature point clouds do not match, then combine the n view geometric feature point clouds, recalculate the shape matching degree, and make a judgment again with the preset shape threshold. E: If the morphological matching degree is greater than or equal to the morphological threshold, then the surface area and specific surface area of ​​the particle contact point cloud are compared with those of the view geometric feature point cloud: F: If the specific surface area is equal but the surface area is not equal, then the view geometric feature point cloud is determined to be similar to the particle contact point cloud, and the view geometric feature point cloud is scaled proportionally. At the same time, the corresponding pile particle model is obtained according to the digital image identification table. If the specific surface area and the surface area are not equal, then combine the m view geometric feature point clouds, calculate the combined specific surface area and the combined surface area, and compare them again with the specific surface area and surface area of ​​the view geometric feature point clouds. If the specific surface areas are unequal but the surface areas are equal, then reverse reasoning is performed on the current particle contact point cloud, and the view geometric feature point cloud is corrected. At the same time, a new rockfill particle model is generated based on the digital image identification table and geometric characteristic parameters. G: If the specific surface area and the surface area are equal, then calculate and judge the flatness, elongation, roundness and sphericity between the geometric feature point cloud of the view and the contact point cloud of the particles according to the contact force model. H: If any of the flatness, elongation, roundness and sphericity between the two are not matched, the view geometric feature point cloud is corrected according to the dynamic correction weight, and the contact probability between the pile particles is obtained by combining the initial pile map structure. If the flatness, elongation, roundness, and sphericity of the two are all matched, the geometric feature point cloud of the view is reduced in dimensionality according to the view orientation to obtain all the contact surfaces of the rockfill skeleton, and then the contact surface dataset is constructed.

[0012] Preferably, step S4 includes the following steps: S41: Based on the set of contact surfaces of the riprap, establish a contact adjacency table between riprap particles, record the contact surface normal vector, contact area and riprap stiffness, and generate a contact network diagram. S42: Employing geometric Boolean operations based on boundary representation, the independent riprap particle model is integrated with the contact surface geometry. Local mesh refinement is performed on the overlapping domains between particles, and the gaps in the riprap particle model are filled by combining a non-manifold topology stitching algorithm. S43: Based on the lithology of the particles, the elastic modulus, density and internal friction angle are matched with the rockfill material library and bound to the rockfill particle model. At the same time, the properties of the contact surface are dynamically assigned to generate solid rockfill particles. S44: Use multi-resolution mesh to model the skeleton of solid rockfill particles to obtain the rockfill skeleton; S45: Perform geometric consistency verification and mechanical rationality verification on the rockfill skeleton to optimize the skeleton structure reasoning mechanism and complete the rockfill skeleton modeling.

[0013] This invention uses a skeleton structure reasoning mechanism combined with geometric feature parameters and a large model to derive and enhance the rockfill particles, and reasons and displays the contact surface of the rockfill particles inside the rockfill skeleton, thereby demonstrating the diversity and complexity of the rock morphology in the rockfill skeleton and verifying the robustness of the rockfill skeleton of the dam. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0015] Figure 1 A schematic diagram illustrating the principle of the method for modeling the riprap skeleton; Figure 2 A schematic diagram for inferring all the contact surfaces of the rockfill within the rockfill skeleton. Detailed Implementation

[0016] The present invention will now be described in detail with reference to the accompanying drawings and embodiments: like Figures 1-2 As shown, the method for modeling a rockfill skeleton based on a large model according to the present invention includes the following steps in sequence; S1: Collect images of rock piles of different shapes from different angles to obtain multi-view images of rock piles; S2: Construct several riprap particle models based on multi-view riprap images, and simultaneously calculate the geometric feature parameters of all views of the riprap to obtain a particle-feature map. S3: Based on the rockfill skeleton image of the dam, the skeleton structure inference mechanism is used in combination with the geometric feature parameters of all views of the rockfill to infer all rockfill contact surfaces inside the rockfill skeleton and construct a contact surface dataset. S4: Based on the contact surface dataset and the riprap particle model, construct a complete solid riprap particle model, and then complete the riprap skeleton modeling.

[0017] In this invention, in step S1, an oblique photography device is used to acquire images of piled stones of different shapes from different angles to obtain multi-view images of piled stones; at the same time, based on the image acquisition time and the view of the piled stones, combined with N-bit random numbers, a digital view identifier is generated and assigned to all multi-view images of piled stones to obtain a digital image identifier table.

[0018] In this invention, step S2 includes the following steps: S21: Use an edge enhancement algorithm to calculate the gradient of the multi-view rock pile image and extract the rock pile edges from different viewpoints; S22: Using an optical flow algorithm combined with a digital image identification table, several rock pile edges are uniquely combined to obtain initial rock pile point cloud data; S23: Use graph neural network algorithm to detect interference points in the initial rockfill point cloud data, and then remove abnormal points to obtain the main rockfill point cloud data; S24: Based on the geological statistical laws, the Diffusion model is used to enhance the point cloud data of the main rockfill point cloud data to obtain polymorphic rockfill point cloud data. S25: Use a non-aliased adversarial network to decouple and controllably generate the polymorphic riprap point cloud data to obtain several riprap view point cloud data. In this embodiment, feature decoupling refers to separating the mixed features of particulate materials, such as morphology and mechanics, into independent variables through mathematical or algorithmic means, eliminating redundant correlations between features, and achieving accurate mapping between microscopic parameters and macroscopic performance. Controllable generation refers to the synthesis of particle systems that meet specific engineering performance requirements based on decoupled feature parameters and through generative models (such as GANs and diffusion models).

[0019] S26: Construct several riprap particle models based on the multi-view point cloud data of the riprap using a 3D reconstruction algorithm; S27: Perform multidimensional characterization calculations on all riprap particle models and riprap view point cloud data in sequence to obtain the geometric feature parameters of all riprap views; In this embodiment, multidimensional characterization calculation refers to constructing a high-dimensional feature space by integrating multidimensional feature parameters such as geometry, mechanics, and topology of particulate materials, and using dimensionality reduction algorithms or machine learning models to quantify its correlation with macroscopic engineering performance. Geometric characteristic parameters include surface area, volume, specific surface area, flatness, elongation, roundness, sphericity, and three-dimensional fractal dimension; In this embodiment, the specific surface area is the total surface area of ​​particles per unit mass or per unit volume. Flatness refers to the degree of flatness of particles in a certain direction. Flat particles are prone to forming layered accumulation, which may reduce shear strength but improve permeability anisotropy. Elongation is the ratio of particle length to width. Particles with high elongation tend to be oriented, leading to anisotropy in mechanical properties, such as significant differences in compressive strength in different directions. Roundness refers to the smoothness of the edges and corners of particles. Particles with high roundness have a lower friction angle and better flowability, and are commonly found in river sediments; while angular particles have strong interlocking properties and more stable depositional structures. Sphericity refers to the degree to which particles approximate an ideal sphere. Spherical particles roll easily and have low porosity; non-spherical particles improve shear strength through interlocking. The three-dimensional fractal dimension quantifies the complexity and irregularity of the particle surface or overall morphology. The higher the fractal dimension, the rougher the surface, which may enhance the interlocking effect between particles, but it will also increase the permeation resistance.

[0020] S28: Based on the digital image identification table, combined with the contact relationship between particles and the path of mechanical property transmission, the geometric feature parameters and the rockfill particle model are associated using a knowledge graph to obtain a particle-feature graph.

[0021] In this embodiment, the interparticle contact relationship refers to the mechanical interaction mode formed by discrete particles under stacking or stress, including contact type, contact force transmission mechanism and geometric topological characteristics, which are key microscopic factors that determine the macroscopic mechanical behavior of particulate materials. The mechanical property transfer path refers to the force chain structure through which external loads diffuse into the interior of the medium via a particle contact network, reflecting the microscopic mechanical mechanism of macroscopic stress-strain response.

[0022] In this invention, step S3 includes the following steps: S31: Aerial photography of the dam during its construction process was conducted using drones to obtain several images of the rockfill skeleton. S32: The rockfill skeleton image is stitched together using an optical fluid algorithm to obtain the main image of the dam; S33: Perform principal component analysis on the main image of the dam, calculate the principal vector of the rockfill skeleton, and obtain the axial-particle packing correlation diagram of the dam body; S34: Based on the particle-feature map and the main vector of the rockfill skeleton, combined with the dam body axial-particle accumulation correlation diagram, the morphology of the rockfill skeleton is obtained; S35: Based on the shape of the rockfill skeleton along the main direction, the rockfill skeleton is sliced ​​into layers according to a preset fixed thickness, the volume ratio of rockfill particles in each layer is calculated, and a density distribution heat map is generated. S36: A three-dimensional density field is constructed using the Kriging interpolation algorithm, and potential pore concentration areas and dense contact surface areas in the dam body image are identified based on the density distribution heat map; S37: A graph attention network is used to construct an initial rockfill map structure based on the geometric feature parameters and particle spatial coordinates in the particle-feature map. S38: The contact probability between pile particles is calculated by using a skeleton structure reasoning mechanism combined with the initial pile diagram structure. S39: Based on the contact probability and preset contact threshold combined with the particle space coordinates, determine all rockfill contact surfaces within the rockfill skeleton, and calculate the contact surface normal vector and contact area.

[0023] In this invention, the operation flow of the skeleton structure reasoning mechanism is as follows: A: The watershed algorithm is used to perform three-dimensional edge segmentation on the overlapping areas of the rockfill skeleton image based on the potential pore concentration area and the dense contact surface area, so as to obtain the inter-particle overlapping domain features and rockfill overlapping point cloud. B: Morphological calculations are performed on the features of the overlapping regions between particles to obtain the area and volume of the overlapping regions. At the same time, a point cloud segmentation algorithm is used to segment the overlapping point cloud of the rock pile according to a preset overlapping threshold to obtain several particle contact point clouds. In this embodiment, the interparticle overlap domain feature refers to the overlapping spatial domain formed by the deformation or interlocking of adjacent particles in the contact area. Its geometric and mechanical properties dominate the microscopic force transmission behavior of the particle system and are the key transition scale connecting discrete particle contact and continuous medium mechanical response. Morphological measurement refers to the establishment of a correlation model between microscopic morphology and macroscopic mechanical behavior by quantitatively analyzing the geometric morphological parameters (such as particle size, shape, surface texture, etc.) and spatial distribution characteristics of particulate materials. C: Calculate the three-dimensional fractal dimension of each particle contact point cloud, and at the same time calculate the morphological matching degree between the particle contact point cloud and the view geometric feature point cloud, and judge it with the preset morphological threshold. D: If the shape matching degree is less than the shape threshold, change the view geometric feature point cloud and compare again until the shape matching degree of the two is greater than or equal to the shape threshold. If all view geometric feature point clouds do not match, combine the n view geometric feature point clouds, recalculate the shape matching degree, and then compare it with the preset shape threshold. E: If the morphological matching degree is greater than or equal to the morphological threshold, then the surface area and specific surface area of ​​the particle contact point cloud are compared with those of the view geometric feature point cloud: F: If the specific surface area is equal but the surface area is not equal, then the view geometric feature point cloud is determined to be similar to the particle contact point cloud, and the view geometric feature point cloud is scaled proportionally. At the same time, the corresponding pile particle model is obtained according to the digital image identification table. If the specific surface area and the surface area are not equal, then combine the m view geometric feature point clouds, calculate the combined specific surface area and the combined surface area, and compare them again with the specific surface area and surface area of ​​the view geometric feature point clouds. If the specific surface areas are unequal but the surface areas are equal, then reverse reasoning is performed on the current particle contact point cloud, and the view geometric feature point cloud is corrected. At the same time, a new rockfill particle model is generated based on the digital image identification table and geometric characteristic parameters. G: If the specific surface area and the surface area are equal, then calculate and judge the flatness, elongation, roundness and sphericity between the geometric feature point cloud of the view and the contact point cloud of the particles according to the contact force model. H: If any of the flatness, elongation, roundness and sphericity between the two are not matched, the view geometric feature point cloud is corrected according to the dynamic correction weight, and the contact probability between the pile particles is obtained by combining the initial pile map structure. If the flatness, elongation, roundness, and sphericity of the two are all matched, the geometric feature point cloud of the view is reduced in dimensionality according to the view orientation to obtain all the contact surfaces of the rockfill skeleton, and then the contact surface dataset is constructed.

[0024] In this invention, step S4 includes the following steps: S41: Based on the set of contact surfaces of the riprap, establish a contact adjacency table between riprap particles, record the contact surface normal vector, contact area and riprap stiffness, and generate a contact network diagram. S42: Employing geometric Boolean operations based on boundary representation, the independent riprap particle model is integrated with the contact surface geometry. Local mesh refinement is performed on the overlapping domains between particles, and the gaps in the riprap particle model are filled by combining a non-manifold topology stitching algorithm. S43: Based on the lithology of the particles, the elastic modulus, density and internal friction angle are matched with the rockfill material library and bound to the rockfill particle model. At the same time, the properties of the contact surface are dynamically assigned to generate solid rockfill particles. In this embodiment, the lithology includes sedimentary rocks such as sandstone (quartz sandstone, feldspar sandstone), conglomerate, limestone, dolomite, etc.; igneous rocks such as granite, basalt, andesite, etc.; and metamorphic rocks such as gneiss, quartzite, marble, etc.; and the properties include normal stiffness, tangential stiffness, and cohesion. S44: Use multi-resolution mesh to model the skeleton of solid rockfill particles to obtain the rockfill skeleton; S45: Perform geometric consistency verification and mechanical rationality verification on the rockfill skeleton to optimize the skeleton structure reasoning mechanism and complete the rockfill skeleton modeling.

[0025] In this embodiment, geometric consistency verification refers to verifying the degree of matching between the geometric morphological parameters of the rockfill (such as stone size, stacking angle, porosity, etc.) and the design target or theoretical model by quantitatively analyzing the geometric morphological parameters of the rockfill. Mechanical rationality verification refers to evaluating whether the stress transfer, stability and deformation characteristics of the rockfill skeleton under load meet safety requirements by comparing mechanical model calculations and experimental data. Example: By using an oblique photography device, images of rock piles of different shapes are acquired from different angles to obtain multi-view images of rock piles. At the same time, based on the image acquisition time and the rock pile view, combined with N-bit random numbers, digital view identifiers are generated and assigned to all multi-view rock pile images to obtain a digital image identifier table.

[0026] This paper employs an edge enhancement algorithm to calculate gradients in multi-view rockfill images, extracting rockfill edges from different perspectives. An optical flow algorithm combined with a digital image identifier table is used to uniquely combine several rockfill edges, obtaining initial rockfill point cloud data. A graph neural network algorithm is then used to detect interference points in the initial rockfill point cloud data, removing outliers to obtain the main rockfill point cloud data. Based on geostatistical principles, a difffusion model is used to enhance the main rockfill point cloud data, obtaining polymorphic rockfill point cloud data. Finally, a non-aliased adversarial network is used to extract features from the polymorphic rockfill point cloud data. Coupled with controllable generation, several point cloud data of riprap views are obtained; using a 3D reconstruction algorithm, several riprap particle models are constructed based on the riprap multi-view point cloud data; multi-dimensional characterization calculations are performed on all riprap particle models and riprap view point cloud data in sequence to obtain the geometric feature parameters of all riprap views: surface area, volume, specific surface area, flatness, elongation, roundness, sphericity, and 3D fractal dimension; based on the digital image identification table combined with the contact relationship between particles and the path of mechanical property transmission, the geometric feature parameters and riprap particle models are associated using a knowledge graph to obtain a particle-feature graph.

[0027] Aerial photography of the dam during its construction was conducted using drones, yielding several images of the rockfill skeleton. These images were then stitched together using an optical fluid dynamics algorithm to obtain the main image of the dam. Principal component analysis was performed on the main dam image to calculate the principal vector of the rockfill skeleton, resulting in an axial-particle packing correlation map of the dam body. Based on the particle-feature map and the principal vector of the rockfill skeleton, combined with the axial-particle packing correlation map, the morphology of the rockfill skeleton was determined. The rockfill skeleton was then sliced ​​along its principal direction according to a predetermined fixed thickness, and the volume ratio of the rockfill particles in each layer was statistically analyzed to generate a density distribution heatmap. A three-dimensional density field was constructed using a Kriging interpolation algorithm, and the potential porosity concentration areas and contact surface density areas in the main dam image were identified based on the density distribution heatmap. A graph attention network was used to construct an initial rockfill map structure based on the geometric feature parameters and particle spatial coordinates in the particle-feature map. A skeleton structure inference mechanism was used in conjunction with the initial rockfill map structure to calculate the contact probability between rockfill particles. Based on the contact probability, a predetermined contact threshold, and particle spatial coordinates, all contact surfaces within the rockfill skeleton were determined, and the contact surface normal vector and contact area were calculated.

[0028] Based on the set of contact surfaces of the riprap, a contact adjacency table is established between riprap particles, recording the contact surface normal vector, contact area, and riprap stiffness to generate a contact network diagram. Geometric Boolean operations based on boundary representation are used to integrate the independent riprap particle models with the contact surface geometry: local mesh refinement is applied to the overlapping domains between particles, and a non-manifold topology stitching algorithm is used to fill the gaps in the riprap particle models. Based on the particle lithology, elastic modulus, density, and internal friction angle are matched using a riprap material library and bound to the riprap particle models. Simultaneously, attributes are dynamically assigned to the contact surfaces to generate solid riprap particles. Multi-resolution meshes are used to model the skeleton of the solid riprap particles, obtaining the riprap skeleton. Geometric consistency and mechanical rationality verification are performed on the riprap skeleton to provide feedback for optimizing the skeleton structure reasoning mechanism, thus completing the riprap skeleton modeling.

Claims

1. A method for modeling a riprap skeleton based on a large model, characterized in that: The steps are as follows: S1: Collect images of rock piles of different shapes from different angles to obtain multi-view images of rock piles; S2: Construct several riprap particle models based on multi-view riprap images, and simultaneously calculate the geometric feature parameters of all views of the riprap to obtain a particle-feature map. S3: Based on the rockfill skeleton image of the dam, the skeleton structure inference mechanism is used in combination with the geometric feature parameters of all views of the rockfill to infer all rockfill contact surfaces inside the rockfill skeleton and construct a contact surface dataset. S4: Based on the contact surface dataset and the riprap particle model, construct a complete solid riprap particle model, and then complete the riprap skeleton modeling.

2. The method for modeling a riprap skeleton based on a large model according to claim 1, characterized in that: In step S1, an oblique photography device is used to acquire images of rock piles of different shapes from different angles to obtain multi-view rock pile images; at the same time, based on the image acquisition time and rock pile view, combined with N-bit random numbers, a digital view identifier is generated and assigned to all multi-view rock pile images to obtain a digital image identifier table.

3. The method for modeling a riprap skeleton based on a large model according to claim 1, characterized in that: Step S2 includes the following steps: S21: Use an edge enhancement algorithm to calculate the gradient of the multi-view rock pile image and extract the rock pile edges from different viewpoints; S22: Using an optical flow algorithm combined with a digital image identification table, several rock pile edges are uniquely combined to obtain initial rock pile point cloud data; S23: Use graph neural network algorithm to detect interference points in the initial rockfill point cloud data, and then remove abnormal points to obtain the main rockfill point cloud data; S24: Based on the geological statistical laws, the Diffusion model is used to enhance the point cloud data of the main rockfill to obtain polymorphic rockfill point cloud data.

4. The method for modeling a riprap skeleton based on a large model according to claim 1, characterized in that: Step S2 also includes the following steps: S25: The polymorphic rock pile point cloud data is decoupled and generated in a controllable manner using a non-aliased adversarial network to obtain several rock pile view point cloud data. S26: Construct several riprap particle models based on the multi-view point cloud data of the riprap using a 3D reconstruction algorithm; S27: Perform multidimensional characterization calculations on all riprap particle models and riprap view point cloud data in sequence to obtain the geometric feature parameters of all riprap views; Geometric characteristic parameters include surface area, volume, specific surface area, flatness, elongation, roundness, sphericity, and three-dimensional fractal dimension; S28: Based on the digital image identification table, combined with the contact relationship between particles and the path of mechanical property transmission, the geometric feature parameters and the rockfill particle model are associated using a knowledge graph to obtain a particle-feature graph.

5. The method for modeling a riprap skeleton based on a large model according to claim 1, characterized in that: Step S3 includes the following steps: S31: Aerial photography of the dam during its construction process was conducted using drones to obtain several images of the rockfill skeleton. S32: The rockfill skeleton image is stitched together using an optical fluid algorithm to obtain the main image of the dam; S33: Perform principal component analysis on the main image of the dam, calculate the principal vector of the rockfill skeleton, and obtain the axial-particle packing correlation diagram of the dam body; S34: Based on the particle-feature map and the main vector of the rockfill skeleton, combined with the dam body axial-particle accumulation correlation diagram, the morphology of the rockfill skeleton is obtained; S35: Based on the shape of the rockfill skeleton along the main direction, the rockfill skeleton is sliced ​​into layers according to a preset fixed thickness, the volume ratio of rockfill particles in each layer is calculated, and a density distribution heat map is generated.

6. The method for modeling a riprap skeleton based on a large model according to claim 1, characterized in that: Step S3 also includes the following steps: S36: A three-dimensional density field is constructed using the Kriging interpolation algorithm, and potential pore concentration areas and dense contact surface areas in the dam body image are identified based on the density distribution heat map; S37: A graph attention network is used to construct an initial rockfill map structure based on the geometric feature parameters and particle spatial coordinates in the particle-feature map. S38: The contact probability between pile particles is calculated by using a skeleton structure reasoning mechanism combined with the initial pile diagram structure. S39: Based on the contact probability and preset contact threshold combined with the particle space coordinates, determine all rockfill contact surfaces within the rockfill skeleton, and calculate the contact surface normal vector and contact area.

7. The method for modeling a riprap skeleton based on a large model according to claim 1, characterized in that: The operation flow of the skeleton structure reasoning mechanism is as follows: A: The watershed algorithm is used to perform three-dimensional edge segmentation on the overlapping areas of the rockfill skeleton image based on the potential pore concentration area and the dense contact surface area, so as to obtain the inter-particle overlapping domain features and rockfill overlapping point cloud. B: Morphological calculations are performed on the features of the overlapping regions between particles to obtain the area and volume of the overlapping regions. At the same time, a point cloud segmentation algorithm is used to segment the overlapping point cloud of the rock pile according to a preset overlapping threshold to obtain several particle contact point clouds. C: Calculate the three-dimensional fractal dimension of each particle contact point cloud, and at the same time calculate the morphological matching degree between the particle contact point cloud and the view geometric feature point cloud, and judge it with the preset morphological threshold. D: If the shape matching degree is less than the shape threshold, change the view geometric feature point cloud and compare again until the shape matching degree of the two is greater than or equal to the shape threshold. If all view geometric feature point clouds do not match, combine the n view geometric feature point clouds, recalculate the shape matching degree, and then compare it with the preset shape threshold. E: If the morphological matching degree is greater than or equal to the morphological threshold, then the surface area and specific surface area of ​​the particle contact point cloud are compared with those of the view geometric feature point cloud.

8. The method for modeling a riprap skeleton based on a large model according to claim 1, characterized in that: The skeleton structure reasoning mechanism also includes the following steps: F: If the specific surface area is equal but the surface area is not equal, then the view geometric feature point cloud is determined to be similar to the particle contact point cloud, and the view geometric feature point cloud is scaled proportionally. At the same time, the corresponding pile particle model is obtained according to the digital image identification table. If the specific surface area and the surface area are not equal, then combine the m view geometric feature point clouds, calculate the combined specific surface area and the combined surface area, and compare them again with the specific surface area and surface area of ​​the view geometric feature point clouds. If the specific surface areas are unequal but the surface areas are equal, then reverse reasoning is performed on the current particle contact point cloud, and the view geometric feature point cloud is corrected. At the same time, a new rockfill particle model is generated based on the digital image identification table and geometric characteristic parameters. G: If the specific surface area and the surface area are equal, then calculate and judge the flatness, elongation, roundness and sphericity between the geometric feature point cloud of the view and the contact point cloud of the particles according to the contact force model. H: If any of the flatness, elongation, roundness and sphericity between the two are not matched, the view geometric feature point cloud is corrected according to the dynamic correction weight, and the contact probability between the pile particles is obtained by combining the initial pile map structure. If the flatness, elongation, roundness, and sphericity of the two are all matched, the geometric feature point cloud of the view is reduced in dimensionality according to the view orientation to obtain all the contact surfaces of the rockfill skeleton, and then the contact surface dataset is constructed.

9. The method for modeling a riprap skeleton based on a large model according to claim 1, characterized in that: Step S4 includes the following steps: S41: Based on the set of contact surfaces of the riprap, establish a contact adjacency table between riprap particles, record the contact surface normal vector, contact area and riprap stiffness, and generate a contact network diagram. S42: Employing geometric Boolean operations based on boundary representation, the independent riprap particle model is integrated with the contact surface geometry. Local mesh refinement is performed on the overlapping domains between particles, and the gaps in the riprap particle model are filled by combining a non-manifold topology stitching algorithm. S43: Based on the lithology of the particles, the elastic modulus, density and internal friction angle are matched with the rockfill material library and bound to the rockfill particle model. At the same time, the properties of the contact surface are dynamically assigned to generate solid rockfill particles. S44: Use multi-resolution mesh to model the skeleton of solid rockfill particles to obtain the rockfill skeleton; S45: Perform geometric consistency verification and mechanical rationality verification on the rockfill skeleton to optimize the skeleton structure reasoning mechanism and complete the rockfill skeleton modeling.

Images