A method, device and storage medium for intelligently generating a three-dimensional cast steel node entity model

Abstract

The present invention discloses a method, device, and storage medium for intelligently generating a three-dimensional cast steel node entity model. The method for intelligently generating a three-dimensional cast steel node entity model comprises the following steps: (1) generating a topologically optimized structure of a three-dimensional cast steel node; (2) using the topologically optimized structure of the three-dimensional cast steel node to create a training set for intelligently generating a three-dimensional cast steel node model; (3) importing the training set obtained in step (2) into a three-dimensional generative adversarial network to generate a three-dimensional cast steel node model and performing visualization processing on the model; (4) performing entity and splicing processing on the visualized three-dimensional cast steel node model. This method can solve the problems of insufficient video memory of the computer during editing operations in the process of generating a three-dimensional cast steel node entity model using a computer, insufficient surface smoothness of the generated three-dimensional cast steel node entity model, and lack of complete properties of the entity model.

Description

Technical Field

[0001] The present invention relates to a method for intelligently generating a three-dimensional joint entity model (3D-JointGAN), and in particular to a method, device and storage medium for intelligently generating a three-dimensional cast steel joint entity model. Background Art

[0002] In recent years, with the rapid development of artificial intelligence (AI) technology in civil engineering, structural design has gradually shifted from experience-based, manual design to automated, digital, and modular intelligent design. Intelligent model generation technology, as a core component of intelligent structural design, plays a key role in improving design efficiency and quality. Currently, intelligent model generation technologies primarily encompass two-dimensional (2D) and three-dimensional (3D). 2D generation technology is relatively mature and has been successfully applied in some structural design projects. Research has been conducted on the intelligent generation of 2D vehicle body shapes using deep generative models; a deep learning-based 2D wheel generative design method has been proposed; and intelligent generative design of cast steel support base plates has been conducted using topology optimization and deep learning techniques. All of these studies involve image-based 3D structural design. These structures are formed by extending a 2D image along its normal direction—a 3D volume formed by stacking a series of planes. These structures do not possess true 3D characteristics and are essentially 2D structures. However, with the advancement of modern engineering technology, the scale of structures continues to expand, and the complexity of systems is increasing, placing higher demands on structural design. Structural design based solely on 2D image generation technology is no longer able to meet current engineering needs. Therefore, in order to realize the intelligent design of true three-dimensional structures and adapt to the needs of the development of the times, it is very necessary to study the intelligent generation method of three-dimensional models.

[0003] Scholars at home and abroad have long explored this issue, but progress has been slow. In recent years, deep learning, as a mainstream development in artificial intelligence, has demonstrated significant value in the field of intelligent 3D model generation. Numerous deep learning algorithms for 3D model generation have been developed, including GAN, DeepVO, BA-Net, and CodeSLAM. GAN, with its powerful feature learning and representation capabilities, has become the most dominant generation method. GANs primarily employ three representation methods in various 3D generation applications: voxels, point clouds, and meshes. Voxels offer a simpler representation, but increasing their resolution significantly increases memory usage. Compared to voxels, point clouds offer a simpler, more unified 3D representation and are easier to manipulate during geometric transformations and deformations. However, point clouds lack connectivity and, due to the lack of surface information, the surface features of generated point clouds are unclear. The emergence of 3D mesh models addresses many of the shortcomings of voxels and point clouds. Meshes are lightweight, offer rich shape details, and importantly, maintain connectivity between adjacent points. For example, 3D-RecGAN can reconstruct the complete 3D structure of a given object from a single view at arbitrary depth; the human mesh recovery (HMR) method can reconstruct a 3D human mesh from a single RGB image; 3DFaceGAN can accurately represent, generate, and transform 3D faces; and Total3DUnderstanding integrates object recognition and mesh generation to transform a single RGB image into a complex 3D scene. However, mesh representation methods also have drawbacks. Since meshes describe 3D objects using vertices and faces, capturing more detailed variations in the 3D object requires more vertices and faces, which can lead to insufficient graphics memory on a computer's graphics processor, resulting in unusual lags when operating on the mesh model. Furthermore, mesh models exhibit sudden curvature changes at the intersections of their faces, resulting in insufficient surface smoothness. Most importantly, mesh models lack the completeness of solid models and cannot meet the requirements of mechanical performance analysis models for engineering structures, thus failing to provide a sound model foundation for intelligent structural design. Summary of the Invention

[0004] The technical problem solved by the present invention is to provide an intelligent generation method, device and storage medium for a three-dimensional cast steel node solid model, thereby solving the problems of abnormal computer freezes when performing editing operations in the process of generating a three-dimensional cast steel node solid model, the surface of the generated three-dimensional cast steel node solid model not having sufficient smoothness, and not having the complete properties of a solid model.

[0005] The technical solution adopted in the present invention is as follows:

[0006] The present invention discloses an intelligent generation method of a three-dimensional cast steel node entity model, comprising the following steps:

[0007] (1) Generate the topological optimization structure of the three-dimensional cast steel node;

[0008] (2) Using the topological optimization structure of the 3D cast steel node to create a training set for intelligent generation of the 3D cast steel node model;

[0009] (3) importing the training set obtained in step (2) into a three-dimensional generative adversarial network to generate a three-dimensional cast steel node model and perform visualization processing on it;

[0010] (4) The visualized three-dimensional cast steel node model is solidified and spliced.

[0011] Furthermore, the method for generating the topological optimization structure of the three-dimensional cast steel node in step (1) specifically includes the following process:

[0012] (1.1) Using SolidWorks, an initial model of the 3D cast steel node was created. The entire model was divided into optimized and non-optimized areas. The solid sphere was divided into the optimized area, and the main and branch pipes were divided into the non-optimized area.

[0013] (1.2) Using the SIMP method, the initial model of the three-dimensional cast steel node is topologically optimized to obtain the topological optimized structure of the three-dimensional cast steel node;

[0014] (1.2.1) Import the initial model of the 3D node into HyperMesh, create an isotropic material and assign it PSOLID entity properties. The material is cast steel with the corresponding elastic modulus E = 206 GPa, Poisson's ratio μ = 0.3, and density ρ = 7.85×10 -6 kg / mm 3 ;

[0015] (1.2.2) Apply loads and constraints to the meshed model based on the most unfavorable working conditions and boundary conditions of the initial node model;

[0016] (1.2.3) Topology optimization is performed with volume fraction as the constraint condition and flexibility minimization as the objective function to obtain the three-dimensional cast steel node topology optimization structure.

[0017] Furthermore, the method for preparing a training set for intelligently generating a three-dimensional cast steel node model in step (2) comprises the following steps:

[0018] (2.1) Export the 3D cast steel node topology optimization structure generated by SIMP method into a .step entity file;

[0019] (2.2) By adjusting the unit density threshold and the working condition type and incorporating the optimization parameters into them, a certain number of three-dimensional cast steel node topology optimization models are obtained. These models are cut off from the main pipes and branch pipes using transparent solid spheres, and splicing joints are retained at the roots where they are connected to the middle design area, so that the intelligently generated node models can be spliced ​​with the main pipes and branch pipes.

[0020] (2.3) The resected model is rotated with the center line of the model as the axis, thereby increasing the number of models;

[0021] (2.4) Convert the .step format solid file into a .off format mesh file using 3D modeling software;

[0022] (2.5) The .off file is voxelized using MATLAB to obtain a training set for intelligently generating three-dimensional cast steel node models.

[0023] Furthermore, in step (3), before using the three-dimensional generative adversarial network to generate the three-dimensional cast steel node model, the voxel values ​​of the voxelized images in the training set need to be scaled from [0, 1] to the range of [-1, 1] of the tanh activation function; the generator and discriminator of the three-dimensional generative adversarial network both use the Adam optimizer, and all weights are initialized using the Glorot normal distribution initialization method, and binary cross entropy is selected as the loss function.

[0024] Furthermore, the process of visualizing the three-dimensional cast steel node model in step (3) includes the following steps:

[0025] (3.1) Set the normal distribution as a random seed and input it into the trained 3D cast steel node model to obtain voxelized images of different 3D nodes;

[0026] (3.2) Process the voxel values ​​of the voxelized image. First, scale the voxel values ​​from [-1, 1] back to [0, 1]. Then, set the voxel values ​​less than 0.5 to 0 and the voxel values ​​greater than 0.5 to 1. Leave the voxel values ​​equal to 0.5 unchanged. Obtain the three-dimensional coordinates of all points whose voxel values ​​are 1.

[0027] (3.3) Use Mayavi’s visualization module to visualize the intelligently generated nodes.

[0028] Furthermore, in step (4), the process of materializing and splicing the visualized three-dimensional cast steel node model specifically includes the following steps:

[0029] (4.1) Convert the voxel model of the intelligently generated node into a mesh model, and perform smoothing, hole filling, extraction, and other processing on it to obtain a beautiful and relatively smooth mesh model;

[0030] (4.2) Fit the mesh model surface using NURBS surface and convert it into a CAD model using 3D reconstruction technology;

[0031] (4.3) Export the CAD model in step (4.2) into a solid file in .step format and then perform the main and branch splicing.

[0032] Furthermore, in step (2.3), the resected model is rotated at least 30 times, with the same rotation angle each time, and the total rotation angle is 360 degrees; in step (2.5), the resolution selected when voxelizing the .off file is 64×64×64.

[0033] Furthermore, the normal distribution selected in step (3.1) has a mean of 0, a standard deviation of 0.33, and a tensor shape of 64×100.

[0034] The present invention also discloses a three-dimensional cast steel node entity model intelligent generation device, including a topology optimization structure generation module for generating a topology optimization structure of a three-dimensional cast steel node, a training set generation module for intelligently generating a three-dimensional cast steel node model training set, a model generation and visualization module for generating a three-dimensional cast steel node model and visualizing it, and a solidification and splicing module for solidifying and splicing the visualized three-dimensional cast steel node model.

[0035] The present invention also discloses a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the device where the computer-readable storage medium is located executes the intelligent generation method of the three-dimensional cast steel node entity model of the present invention.

[0036] The beneficial effects of the present invention are as follows: the method can solve the problems of insufficient video memory of the computer during editing operations in the process of generating a three-dimensional cast steel node solid model using a computer, insufficient smoothness of the surface of the generated three-dimensional cast steel node solid model, and sudden curvature changes at the intersection of surfaces. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a flow chart of the method of the present invention;

[0038] Figure 2 is the geometric feature of the trifurcated node;

[0039] Figure 3 This is an example of a topology optimized three-fork node after the main branch is removed;

[0040] Figure 4 is the binary cross entropy loss function of the generator and the discriminator;

[0041] Figure 5 Some examples of intelligently generating trifurcation nodes

[0042] Figure 6 Generate a detailed diagram of node 4 intelligently;

[0043] Figure 7 It is the materialization process of the three-fork node;

[0044] Figure 8 It is the process of splicing the materialized three-fork node and the main branch;

[0045] Figure 9 It is the materialization and splicing processing of the representative node model 1;

[0046] Figure 10 It is the materialization and splicing processing of the representative node model 2;

[0047] Figure 11 It is the materialization and splicing processing of the representative node model 3;

[0048] In the figure, a represents the process of fitting the mesh model of the node, b represents the process of converting the fitted mesh model into a CAD model, and c represents the process of splicing the node model. DETAILED DESCRIPTION

[0049] The present invention will be further described in detail below with reference to the embodiments.

[0050] like Figure 1 As shown, a method for intelligently generating a three-dimensional cast steel node solid model includes the following steps:

[0051] (1) Generate the topological optimization structure of the three-dimensional cast steel node;

[0052] (1.1) The initial model of the 3D cast steel node was established using SolidWorks. The entire model was divided into optimization and non-optimization areas. The solid sphere was divided into the optimization area, and the main pipe and branch pipe were divided into the non-optimization area (one main pipe and three branch pipes). The regional division and geometric characteristics of the node are as follows: Figure 2 As shown;

[0053] (1.2) Using the SIMP method, the initial model of the three-dimensional cast steel node is topologically optimized to obtain the topological optimized structure of the three-dimensional cast steel node;

[0054] (1.2.1) Import the initial model of the 3D node into HyperMesh, create an isotropic material and assign it PSOLID entity properties. The material is cast steel, and the corresponding elastic modulus E = 206 GPa, Poisson's ratio μ = 0.3, and density ρ = 7.85×10 -6 kg / mm 3 ;

[0055] (1.2.2) Apply loads and constraints to the meshed model based on the most unfavorable working conditions and boundary conditions of the initial node model;

[0056] Taking into account the combined effects of nodal axial force, shear force, and bending moment, as well as the limitations of existing laboratory instruments for nodal loading tests, the loads are simplified and combined here. A vertical load of 615 kN is applied to the top surfaces of the three meshed branch pipes, for a total load of 1845 kN. The bottom surface of the main pipe is then set as the fixed end.

[0057] (1.2.3) With a volume fraction of 40% as the constraint condition and minimizing flexibility as the objective function, topology optimization is performed to obtain the three-dimensional cast steel node topology optimization structure.

[0058] (2) Using the topological optimization structure of the 3D cast steel node to create a training set for intelligent generation of the 3D cast steel node model;

[0059] (2.1) Export the 3D cast steel node topology optimization structure generated by SIMP method into a .step entity file;

[0060] (2.2) By adjusting the unit density threshold and the working condition type and incorporating the optimization parameters into them, a certain number of three-dimensional cast steel node topology optimization models are obtained. These models are cut off from the main pipes and branch pipes using transparent solid spheres, and splicing joints are retained at the roots where they are connected to the middle design area, so that the intelligently generated node models can be spliced ​​with the main pipes and branch pipes.

[0061] The cell density threshold is adjusted to ρ = 0.10 ~ 0.85. The optimization parameters include manufacturing process constraints such as penalty factor, checkerboard control, and minimum member size. After the optimization parameters are integrated into the .step entity file, 270 topology optimization models are obtained, such as Figure 3 As shown, the 270 models were cut out of the main pipe and branch pipe using a transparent solid sphere, and a 20 mm splicing joint was retained at the root where it was connected to the middle design area.

[0062] The resection operation is primarily performed for two reasons: a) The main and branch pipes occupy too much voxel space. Given a fixed total voxel space, this results in a smaller voxel space in the central design area, resulting in a less clear and smooth result. b) The main and branch pipes in the non-design area also generate new structures, but these structures differ significantly from the original ones and do not meet the non-design requirements for the main and branch pipes.

[0063] (2.3) Rotate the resected model at least 30 times with the centerline of the model as the rotation axis, with each rotation of equal degrees, for a total rotation angle of 360 degrees;

[0064] In this embodiment, the resected model is rotated 30 times, 12 degrees each time, about the model's centerline. This enhancement increases the number of trifurcated node training sets from 270 to 3240, improving deep learning performance. The primary purpose of this invention is to obtain diverse 3D models using a neural network. Therefore, validation and test sets are not divided to increase the amount of training data and the number of extracted features.

[0065] (2.4) Convert the .step format solid file into a .off format mesh file using 3D modeling software;

[0066] (2.5) Using a custom MATLAB program, voxelize the .off file at a resolution of 64×64×64 and save the 3D 0-1 matrix corresponding to the optimized 3D cast steel node topology structure after the main and branch pipes are removed in .mat format. This 3D 0-1 matrix in .mat format serves as the training set for intelligently generating the 3D cast steel node model.

[0067] (3) importing the training set obtained in step (2) into a three-dimensional generative adversarial network to generate a three-dimensional cast steel node model and perform visualization processing on it;

[0068] In order to present more details and changes in the design area of ​​the 3D model, the present invention uses a high resolution of 64×64×64 as the output, and before training, the voxel values ​​of the voxelized images in the training set are scaled from [0,1] to the range of [-1,1] of the tanh activation function. The experiments of the present invention were tested in Windows 10 using Python 3.7.11 and Tensorflow 2.3. Both the generator and the discriminator use the Adam optimizer with β1 = 0.5. Since the learning speed of the discriminator is usually much faster than that of the generator, two methods are used to limit the discriminator: one is to set the learning rate of the generator to 0.0015 and the learning rate of the discriminator to 0.0001; the second is that for each round, the discriminator will only be updated if the accuracy of the current batch is not higher than 80%. All models were trained using mini-batch stochastic gradient descent (SGD). Smaller batch sizes have lower memory utilization and introduce greater randomness, making it difficult for the loss function to converge. Furthermore, larger batch sizes (such as 128 or 256) do not significantly improve training results, but consume more memory and slow down the training process. Therefore, after selecting several reasonable batch sizes for training and comparison, the batch size was set to 24. All weights were initialized using the Glorot normal distribution. The alpha parameter of all LeakyReLU activation functions was set to 0.2 to ensure that information on the negative axis is not lost.

[0069] In order to avoid the decrease in learning rate caused by gradient dissipation, the present invention selects binary cross entropy as the loss function. 64 100-dimensional random noises are initialized with a normal distribution with a mean of 0 and a standard deviation of 0.33, and are input into 3D-JointGAN together with the batched training set for training. During the training process of 3D-JointGAN, the loss function is divided into the generator loss function Generator Loss and the discriminator loss function Discriminator Loss. The evolution process of the cross entropy loss function of the generator and the discriminator is as follows: Figure 4 As shown in the figure, the loss functions of the two models show increasing and decreasing trends, respectively, and both are accompanied by varying degrees of fluctuation. After 260 rounds of training, the loss functions of the two models gradually stabilize and eventually converge. It is worth noting that after 260 rounds, the loss functions of the two models fluctuate around a certain value and gradually stabilize rather than converge to a straight line. This is due to the neural network algorithm, but it does not affect its convergence characteristics as an adversarial process. At this point, the generator and discriminator reach a Nash equilibrium in their adversarial learning.

[0070] The process of visualizing the 3D cast steel node model is as follows:

[0071] (3.1) Inputting different random seeds into the trained generator model can generate a variety of trifurcated nodes. This can be achieved by changing the distribution of random variables or related parameters within the same distribution, typically using a priori distributions such as uniform or normal distributions. Furthermore, changing the shape of the tensor in the random seed can control the number of trifurcated nodes generated at a time. The present invention sets a normal distribution with a mean of 0, a standard deviation of 0.33, and a tensor shape of 64×100 as the random seed and inputs it into the trained three-dimensional cast steel node model, obtaining voxelized images of 64 different three-dimensional nodes.

[0072] (3.2) Process the voxel values ​​of the voxelized image. First, scale the voxel values ​​from [-1, 1] back to [0, 1]. Then, set the voxel values ​​less than 0.5 to 0 and the voxel values ​​greater than 0.5 to 1. Leave the voxel values ​​equal to 0.5 unchanged. Obtain the three-dimensional coordinates of all points whose voxel values ​​are 1.

[0073] (3.3) Use Mayavi’s visualization module to visualize the 64 intelligently generated nodes, then sort them by voxel integrity and list the top 24 nodes in Figure 5 middle.

[0074] (4) Solidify and splice the visualized three-dimensional cast steel node model;

[0075] The present invention uses reverse engineering technology to materialize the intelligently generated three-fork node. The details of node 4 are as follows Figure 6 As shown. Figure 7 As shown, the instantiation and splicing process of node 4 is as follows:

[0076] (4.1) The voxel model of the intelligently generated node is converted into a mesh model, and after smoothing, filling holes, and extraction, a beautiful and relatively smooth mesh model is obtained, that is, Figure 6 The leftmost figure in the figure;

[0077] The mesh model obtained in this step has many mesh faces and their distribution is dense.

[0078] (4.2) Use NURBS surface to fit the mesh model surface Figure 6 The middle figure is converted into a CAD model using 3D reconstruction technology. Figure 6 The rightmost figure in ;

[0079] Using NURBS surfaces to fit the mesh model surface at node 4, the 203,702 mesh faces were converted into 207 facets, making the model surface smoother and more uniform. 3D reconstruction technology was then used to convert it into a CAD model.

[0080] (4.3) Export the CAD model into a solid file in .step format and then perform the main and branch splicing.

[0081] The splicing process is as follows Figure 8 As shown in the figure, the lower left corner shows the main and branch pipes pre-cut when the dataset was created, the upper left corner shows the intelligently generated node core, and the upper right corner shows the completed model. The joints of the spliced ​​node model are well-connected and the transition is smooth, meeting the architectural aesthetic requirements for joints. The splicing process is easy to implement and achieves excellent results without spending a lot of time.

[0082] right Figure 5 The properties of the 24 intelligently generated nodes in this paper were summarized to outline the common characteristics of node models generated by the intelligent generation method and to verify the feature extraction and model generation capabilities of the 3D GAN. Three excellent nodes were then selected and their materialization and splicing processes were demonstrated to further verify the rationality of the intelligent generation method. Analysis of the generated results revealed the following properties:

[0083] a) Diversity and innovation. The trifurcated nodes shown in the figure vary in symmetry, voxel distribution, and hole details. These models differ significantly from one another, with no similar or identical appearance features, thus demonstrating their diversity. Furthermore, they differ from the models in the training set, each extending some innovative features while retaining the basic characteristics of the training set models, demonstrating significant innovation.

[0084] b) Complete, symmetrical, and aesthetically pleasing appearance. The trifurcated nodes represented by the voxels are complete, without any damage, missing pieces, or distortion, and the voxel distribution is uniform and continuous. Nodes 1, 2, 3, 4, and 21 are symmetrical about the xy, xz, and yz planes, while the remaining nodes are symmetrical about one or two planes, demonstrating excellent symmetry and aesthetics.

[0085] c) Hole details are fully restored. Hole details at the junctions of the main and branch pipes at nodes 3, 8, 10, 11, 18, 19, 20, 22, 23, and 24, as well as within the interior areas of each node, are fully restored. Whether small or large, round, oval, square, or irregularly shaped, the edges are smooth and gently transitioned, free of impurities or unusual derivatives.

[0086] The above analysis demonstrates that the intelligently generated models possess numerous properties, including symmetry, aesthetics, and novelty, demonstrating the powerful feature extraction and model generation capabilities of 3D-JointGAN. However, the vast majority of these models are beyond the designer's imagination, making it difficult for them to conceive such node configurations based solely on experience and imagination.

[0087] Three nodes with beautiful appearance were selected from the 24 trifurcated nodes displayed, namely node 1, node 2 and node 3, and were materialized and spliced.

[0088] The process of intelligent generation of node 1 is as follows: Figure 9 As shown in the figure, letter a represents meshing, letter b represents solidification, and letter c represents splicing. As can be seen from the figure, the main pipe joint necks upward for a distance, then emanates three branches, each connected to the lower portion of the three branch pipe joints. The three branches have a "U"-shaped cross-section. The upper portions of the three branch pipes are connected and supported by each other, forming a slightly concave connection surface. The bifurcation has a slightly convex connection surface, and a solid cylinder stands within the cavity formed by this connection surface and the inner walls of the three branches. The entire node is beautiful, novel, and well-symmetrical.

[0089] The process of intelligent generation of node 2 is as follows: Figure 10 As shown in the figure, letter a represents meshing, letter b represents solidification, and letter c represents splicing. As can be seen from the figure, the node's outer contour is approximately hyperbolic, with a smooth, distinct outline and a natural transition. The cross-sections of the three branches are thin-walled "U" shapes. The upper parts of the three branch pipes are connected to each other at the joint to form a closed loop. The interior of the node is conical, and the three oblique supports extending from the cone's apex are connected to the connecting supports between the three branch pipes. The entire node has a unique shape and a reasonable material distribution, which makes it novel and beautiful.

[0090] The materialization and splicing process of intelligent generation node 3 is as follows Figure 11 As shown in the figure, letter a represents meshing, letter b represents solidification, and letter c represents splicing. As can be seen from the figure, the interior of the node consists of three thin-walled structures divided equally by 120 degrees. The lower edge of the thin wall is semicircular, extending upward to form three branches that connect to the lower part of the three branch pipe joints. The upper edge of the thin wall is circular and connected to the inner walls of the three branches. The thin-walled structure extends downward to form four evenly distributed supports that connect to the main pipe joint. The upper parts of the three branch pipe joints each extend an acute-angle support that converges at a point. The entire node has a novel appearance, high material utilization, and is symmetrical and beautiful.

[0091] The above are just some examples. In specific operations, the trifurcated nodes represented by voxels can be screened according to mass size or appearance characteristics. The selected models can then be processed using reverse engineering technology to obtain a three-dimensional solid model that can be used for mechanical performance analysis of engineering structures. This realizes the intelligent generation of three-dimensional solid models and lays a good foundation for intelligent design of structures.

[0092] The intelligent generation device of a three-dimensional cast steel node solid model includes a topology optimization structure generation module for generating a topology optimization structure of a three-dimensional cast steel node, a training set generation module for intelligently generating a training set of a three-dimensional cast steel node model, a model generation and visualization module for generating a three-dimensional cast steel node model and visualizing it, and a solidification and splicing module for solidifying and splicing the visualized three-dimensional cast steel node model.

[0093] A computer-readable storage medium stores a computer program, which, when executed by a processor, enables a device where the computer-readable storage medium is located to execute the intelligent generation method of a three-dimensional cast steel node entity model of the present invention.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents, and these modifications or replacements 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 intelligently generating a three-dimensional cast steel node entity model, characterized in that: The following steps are involved: (1) Generate the topological optimization structure of the three-dimensional cast steel node; (2) Using the topological optimization structure of the three-dimensional cast steel node to prepare a training set for intelligently generating a three-dimensional cast steel node model; the method for preparing the training set for intelligently generating a three-dimensional cast steel node model in step (2) comprises the following steps: (2.1) Export the 3D cast steel node topology optimization structure generated by SIMP method into a .step entity file; (2.2) By adjusting the cell density threshold and the operating condition type and incorporating the optimization parameters, a certain number of three-dimensional cast steel node topology optimization models are obtained. These models are then cut off from the main and branch pipes using transparent solid spheres, and splicing joints are retained at the roots where they connect to the intermediate design area, so that the intelligently generated node models can be spliced ​​with the main and branch pipes. (2.3) Rotate the cut model with the center line of the model as the rotation axis to increase the number of models; (2.4) Convert the .step format solid file into a .off format mesh file using 3D modeling software; (2.5) Voxelizing the .off file using MATLAB yields a training set for intelligently generating a 3D cast steel node model. (3) The training set obtained in step (2) is imported into a three-dimensional generative adversarial network to generate a three-dimensional cast steel node model and perform visualization processing on it; before using the three-dimensional generative adversarial network to generate the three-dimensional cast steel node model in step (3), the voxel value of the voxelized image in the training set needs to be scaled from [0, 1] to the range [-1, 1] of the tanh activation function; the generator and discriminator of the three-dimensional generative adversarial network both use the Adam optimizer, take β1 = 0.5, all weights are initialized using the Glorot normal distribution initialization method, and select Binary Crossentropy as the loss function; the process of visualizing the three-dimensional cast steel node model in step (3) includes the following steps: (3.1) Set the normal distribution as a random seed and input it into the trained 3D cast steel node model to obtain voxelized images of different 3D nodes; (3.2) Process the voxel values ​​of the voxelized image. First, scale the voxel values ​​from [-1, 1] back to [0, 1]. Then, set the voxel values ​​less than 0.5 to 0 and the voxel values ​​greater than 0.5 to 1. Leave the voxel values ​​equal to 0.5 unchanged. Obtain the three-dimensional coordinates of all points whose voxel values ​​are 1. (3.3) Use Mayavi’s visualization module to visualize the intelligently generated nodes; (4) The visualized three-dimensional cast steel node model is solidified and spliced.

2. The intelligent generation method of a three-dimensional cast steel node entity model according to claim 1, characterized in that: The method for generating the topology optimization structure of the three-dimensional cast steel node in step (1) specifically includes the following process: (1.1) Using SolidWorks, an initial model of the 3D cast steel node was created. The entire model was divided into optimized and non-optimized areas. The solid sphere was divided into the optimized area, and the main and branch pipes were divided into the non-optimized area. (1.2) Using the SIMP method, the initial model of the three-dimensional cast steel node is topologically optimized to obtain the topological optimized structure of the three-dimensional cast steel node; (1.2.1) Import the initial model of the 3D cast steel node into HyperMesh, create an isotropic material and assign it PSOLID entity properties. Select cast steel as the material and set the appropriate elastic modulus, Poisson's ratio, and density. (1.2.2) Apply loads and constraints to the meshed model based on the most unfavorable working conditions and boundary conditions of the initial node model; (1.2.3) Topology optimization is performed with volume fraction as the constraint condition and flexibility minimization as the objective function to obtain the three-dimensional cast steel node topology optimization structure.

3. The intelligent generation method of a three-dimensional cast steel node entity model according to claim 1, characterized in that: In step (4), the process of materializing and splicing the visualized three-dimensional cast steel node model specifically includes the following steps: (4.1) Convert the voxel model of the intelligently generated node into a mesh model, and then perform smoothing, hole filling, and extraction on it to obtain a beautiful and relatively smooth mesh model; (4.2) Fit the mesh model surface using NURBS surface and convert it into a CAD model using 3D reconstruction technology; (4.3) Export the CAD model into a solid file in .step format and then perform the main and branch splicing.

4. The intelligent generation method of a three-dimensional cast steel node entity model according to claim 1, characterized in that: In step (2.3), the resected model is rotated at least 30 times, with the same rotation angle each time, and the total rotation angle is 360 degrees; in step (2.5), the resolution selected when voxelizing the .off file is 64×64×64.

5. The intelligent generation method of a three-dimensional cast steel node entity model according to claim 1, characterized in that: The normal distribution selected in step (3.1) has a mean of 0, a standard deviation of 0.33, and a tensor shape of 64×100.

6. A device for intelligently generating a three-dimensional cast steel node entity model for implementing any one of the methods of claims 1-5, characterized in that: It includes a topology optimization structure generation module for generating a topology optimization structure of a three-dimensional cast steel node, a training set generation module for intelligently generating a three-dimensional cast steel node model training set, a model generation and visualization module for generating a three-dimensional cast steel node model and visualizing it, and a solidification and splicing module for solidifying and splicing the visualized three-dimensional cast steel node model.

7. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the device where the computer-readable storage medium is located executes the method according to any one of claims 1 to 5.

Images