Building structure intelligent design method and system based on 3D modeling, electronic equipment and storage medium

Abstract

The invention provides a building structure intelligent design method and system based on 3D modeling, electronic equipment and a storage medium, and relates to the field of building intelligent design, and the method comprises the following steps: obtaining three-dimensional data through an unmanned aerial vehicle, oblique photography and laser scanning, and carrying out data cleaning, conversion and screening; performing feature extraction and semantic segmentation on the point cloud data by using a deep learning model; performing intelligent registration and fusion on the two types of point clouds by adopting a deep learning model; a three-dimensional model is generated based on the registration data, a parameter form is generated in combination with the building planning data, and a component model is constructed; collision detection, smooth verification and mechanical simulation are carried out on the whole model, parameter adjustment and optimization are carried out according to verification results, through automatic data collection, multi-source point cloud fusion, intelligent parameter modeling and multi-dimensional verification and optimization, the problems that a traditional method is low in efficiency and insufficient in precision are solved, and the design efficiency and precision are remarkably improved.

Description

Technical Field

[0001] The present application relates to the field of intelligent building design, and specifically to a method, system, electronic device and storage medium for intelligent building structure design based on 3D modeling. Background Art

[0002] In the field of architectural design, traditional methods often rely on manual labor, which is not only time-consuming and labor-intensive, but also difficult to model complex building structures. With the continuous development of the construction industry, higher requirements are placed on the accuracy and efficiency of building structure design. Existing design methods are no longer able to meet actual needs. Summary of the Invention

[0003] The purpose of the embodiments of the present application is to provide a method, system, electronic device and storage medium for intelligent design of building structures based on 3D modeling, which can solve the above-mentioned technical problems.

[0004] The present invention provides a method for intelligent design of building structures based on 3D modeling, comprising the following steps:

[0005] Step 1: Data collection and preprocessing: Relevant data of the building structure is obtained through various data collection methods, including but not limited to 360° drone acquisition of three-dimensional data of the building structure, oblique photogrammetry, and laser scanning to obtain point cloud data.

[0006] Perform data cleaning, data transformation, and data screening on the collected 3D data to obtain pre-processed 3D data; data cleaning includes missing value processing, outlier judgment and processing, and duplicate data removal; data transformation includes generating derivative variables, improving variable distribution characteristics, binning interval variables, and standardizing interval variables;

[0007] Step 2: Feature extraction and semantic segmentation:

[0008] Use deep learning models to perform feature extraction and semantic segmentation on the pre-processed point cloud data, dividing the point cloud data into different building structure parts;

[0009] Step 3: Point cloud registration and fusion:

[0010] A deep learning model is used to intelligently register and fuse two types of point clouds obtained from oblique photogrammetry and laser scanning;

[0011] Step 4: 3D model creation:

[0012] The 3D model is built on the registered and fused point cloud data through multi-view image joint aerial triangulation, dense matching point cloud generation, Delaunay triangulation construction and texture mapping;

[0013] Combined with building planning data, parameter generation rules are used to obtain parameter tables for building components, including a component parameter table and a docking parameter table. The component parameter table is a list of structural and mechanical parameters of building components, and the docking parameter table is a list of structural and mechanical parameters of docking between building components. A generative model is used to generate a three-dimensional model of the building component based on the parameter tables.

[0014] Step 5: Model verification and optimization:

[0015] According to the docking parameter form, the three-dimensional models of the building components are spliced ​​to obtain the overall three-dimensional model, and the splicing positions in the overall three-dimensional model are subjected to collision detection and smoothness verification. The smoothness verification verifies the surface smoothness of the outer surface of the splicing positions in the overall three-dimensional model; solid mechanics simulation and stress analysis are performed on the overall three-dimensional model to determine whether the analysis results meet the requirements of the building planning data; if the verification fails, the number of verification failures in this round of the process is determined. If the number of verification failures is greater than the preset value, the number verification fails, the process exits and returns an error message; if the number of verification failures is less than or equal to the preset value, the number verification succeeds and the process returns to the step of generating the three-dimensional model of the building component to adjust the parameters and optimize the model.

[0016] As a preference, the process of missing value processing in step 1 is to identify the coordinate points with missing values ​​in the three-dimensional data set. Search the complete data set for the k nearest neighbor points N={P1, P2, ..., P k}, calculate the Euclidean distance between each neighboring point and the missing point:

[0017]

[0018] Calculate the weight according to the inverse of the distance and perform weighted filling:

[0019] (prevent division by zero);

[0020] (prevent division by zero);

[0021] (prevent division by zero);

[0022] The process of outlier judgment and processing is as follows: Calculate the mean vector of the data set And covariance matrix Σ; for each data point P(x,y,z), calculate its Mahalanobis distance:

[0023]

[0024] Set the significance level α = 0.05α = 0.05 and determine the threshold value through chi-square distribution: Judgment condition: When D Mahalanobis > threshold, mark the point as an outlier;

[0025] The process of removing duplicate data is as follows: define the coordinate matching tolerance ε=10 -6 , for any two points P i (x i ,y i , z i ) and P j (x j ,y j , z j ), calculate the absolute difference of coordinates: Δx=|x i -x j |, Δy=|y i -y j |, Δz=|z i -z j When Δx≤ε, Δy≤ε, and Δz≤ε are satisfied at the same time, the data is determined to be duplicate data. For the duplicate data group, only the first appearing data point is retained and the rest of the duplicate points are deleted.

[0026] Preferably, the feature extraction and semantic segmentation of point cloud data in step 2 includes A. point cloud data preprocessing:

[0027] Use voxelized grid to downsample the original point cloud. The grid side length is calculated as:

[0028]

[0029] Where N is the preset resolution parameter (range 50-200), δ is the minimum voxel threshold (0.01-0.1 m);

[0030] The noise points are filtered out by the statistical outlier removal algorithm, and the standard deviation of the neighborhood distance of point pi is defined as:

[0031]

[0032] If σ i >α·median({σ1,...,σ n}), then remove P i , where α = 1.5-3.0;

[0033] Then normalize the coordinates

[0034]

[0035] where p min , pmax is the global coordinate extreme value of the point cloud;

[0036] B. Multi-scale hierarchical feature extraction: Build a backbone network based on the improved PointNet++, each layer contains a local feature aggregation module,

[0037]

[0038] Where φ is 1D convolution and MLP is multi-layer perceptron;

[0039] Introducing the channel attention mechanism, the weight calculation is:

[0040] w c =σ(w2·ReLU(w1·GAP(f)))

[0041] Where GAP is global average pooling, w1 and w2 are learnable parameters;

[0042] Adopting adaptive feature fusion decoder, the upsampling formula is:

[0043]

[0044] Where ⊕ represents the feature concatenation of the jump connection;

[0045] Define the multi-task loss function:

[0046]

[0047] in is the cross entropy loss, is Lovasz-Softmax loss, λ2=0.7, λ2=0.3, λ3=0.001;

[0048] D. Post-processing optimization: Conditional random field (CRF) is used to optimize the segmentation results. The energy function is:

[0049]

[0050] in θ α =0.1,θ β =0.01;

[0051] Holes are filled in door and window components, and the morphological closing operation kernel size is 3×3×3 voxels.

[0052] As a preferred method, the processing process in step 3 is as follows: and laser point cloud Multi-scale features are extracted through a dual-branch deep network:

[0053] Texture branch: Use 3D sparse convolution kernel K on S tex Extract texture features F S =σ(W tex *S+b tex )

[0054] Geometry branch: Adaptive radius MLP is used to extract geometric features from TT

[0055] Where σ is the LeakyReLU activation function, φ r Represents the maximum pooling operation within the neighborhood of radius r;

[0056] Differentiable Quasi-Optimization: Constructing a Similarity Matrix Based on Attention Mechanism:

[0057]

[0058] Solving the optimal transmission matrix using the Sinkhorn algorithm And calculate the affine transformation parameters:

[0059]

[0060] The joint loss function calculates the total loss, which includes the improved Chamfer distance, cosine similarity loss, and the quasi-error distribution distributed by the discriminator DD area. The total loss function is

[0061]

[0062] Among them, λ1, λ2, λ3 are dynamically adjusted weights, satisfying λ1+λ2+λ3=1. Then, the fusion quality index is calculated in level ll. and The next level of refinement is triggered when , otherwise it returns to the differentiable allocation quasi-optimization.

[0063] Preferably, in step 4, the projection matrices of at least two cameras and the matching image points are used to calculate the coordinates of the three-dimensional space point using the following formula:

[0064]

[0065] Among them, (u, v) and (u′, v′) are the coordinates of the matching image points, P and P′ are the camera projection matrices, and X=[X x , X y , X z , 1] T is the homogeneous coordinate of a three-dimensional point;

[0066] The iterative closest point (ICP) algorithm is used to solve the point cloud registration parameters by minimizing the following error function:

[0067]

[0068] in, and is the point cloud to be registered, R is the rotation matrix, and t is the translation vector;

[0069] Based on the registered point cloud data, a divide-and-conquer algorithm is used to generate the Delaunay triangulation, ensuring that the circumscribed circle of any triangle does not contain other points.

[0070] The texture coordinates of any point inside the triangle are calculated by interpolating the barycentric coordinates. The formula is as follows:

[0071] V=αV A +βV B +γV C (α+β+γ=1)

[0072] Among them, V A ,V B ,V C are triangle vertex attributes (such as texture coordinates), α, β, γ are barycentric coordinates.

[0073] As a preferred method, in step 4, the component parameter form and the docking parameter form are calculated using parameter generation rules in combination with the building planning data: The component parameter form includes the structural parameters (such as beam cross-sectional area and length) and mechanical parameters (such as volume and bearing capacity) of the component:

[0074] V beam = cross-sectional area × length, butt joint parameter form: describes the structural parameters and mechanical parameters of the butt joint between components;

[0075] Based on the parameter form, a generative model (such as a deep learning network) is used to generate a 3D model of the building component. The model output satisfies the following constraints:

[0076] Model accuracy ≥∈ and mechanical parameter error ≤δ, where ∈ is the geometric accuracy threshold and δ is the mechanical parameter error threshold.

[0077] An intelligent design system for building structures based on 3D modeling includes: a data acquisition module: used to obtain relevant data of building structures through multiple data acquisition methods, including drone modules, oblique photogrammetry equipment and laser scanning equipment, to realize the acquisition of three-dimensional data of building structures.

[0078] Preprocessing module: used to clean, transform and filter the collected 3D data to obtain preprocessed 3D data; this module has a data processing algorithm that can automatically complete operations such as missing value processing, outlier judgment and processing, and duplicate data removal, and perform data transformation to meet subsequent processing requirements;

[0079] Feature extraction and semantic segmentation module: Uses a deep learning model to perform feature extraction and semantic segmentation on preprocessed point cloud data, dividing the point cloud data into different building structure parts; this module contains a trained deep learning model that can accurately analyze and process the input point cloud data.

[0080] Point cloud registration and fusion module: This module uses a deep learning model to intelligently register and fuse two types of point clouds acquired through oblique photogrammetry and laser scanning. This module can automatically achieve accurate registration and fusion of point clouds without manual intervention.

[0081] 3D model building module: This module is used to build a 3D model of the registered and fused point cloud data through multi-view image joint aerial triangulation, dense matching point cloud generation, Delaunay triangulation construction, and texture mapping. It also combines architectural planning data with parameter generation rules and generative models to generate 3D models of building components.

[0082] Model verification and optimization module: splice the 3D models of building components according to the docking parameter form to obtain the overall 3D model, perform collision detection and smoothness verification on the splicing positions in the overall 3D model; perform solid mechanics simulation and stress analysis on the overall 3D model to determine whether the analysis results meet the requirements of the building planning data; if the verification fails, perform corresponding processing based on the number of verification failures, and return to the step of generating the 3D model of the building components to adjust parameters and optimize the model.

[0083] An electronic device comprises: a memory for storing one or more programs; a processor; when the one or more programs are executed by the processor, any one of the 3D modeling-based intelligent design methods for building structures as described in claims 1-7 is implemented.

[0084] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements any one of the 3D modeling-based intelligent design methods for building structures as described in claims 1-7.

[0085] Beneficial effects of the present invention:

[0086] The present application provides a method, system, electronic device and storage medium for intelligent design of building structures based on 3D modeling, including the following steps: acquiring three-dimensional data through drones, oblique photography and laser scanning, and performing data cleaning, transformation and screening; using a deep learning model to perform feature extraction and semantic segmentation on point cloud data; using a deep learning model to perform intelligent registration and fusion of two types of point clouds; generating a three-dimensional model based on the registered data, generating a parameter form and constructing a component model in combination with the building planning data; performing collision detection, smoothing verification and mechanical simulation on the overall model, and adjusting and optimizing parameters according to the verification results. Through automated data collection, multi-source point cloud fusion, intelligent parameter modeling and multi-dimensional verification and optimization, the problems of low efficiency and insufficient accuracy of traditional methods are solved, and the design efficiency and accuracy are significantly improved. DETAILED DESCRIPTION

[0087] Furthermore, terms such as "horizontal," "vertical," and "overhanging" do not necessarily imply that a component must be absolutely horizontal or overhanging, but rather that it can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but rather that it can be slightly tilted.

[0088] It should also be noted that, in the description of this application, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in this application based on the specific circumstances.

[0089] Example 1

[0090] A method for intelligent design of building structures based on 3D modeling, comprising the following steps:

[0091] Step 1: Data collection and preprocessing: Relevant data of the building structure is obtained through various data collection methods, including but not limited to 360° drone acquisition of three-dimensional data of the building structure, oblique photogrammetry, and laser scanning to obtain point cloud data.

[0092] Perform data cleaning, data transformation, and data screening on the collected 3D data to obtain pre-processed 3D data; data cleaning includes missing value processing, outlier judgment and processing, and duplicate data removal; data transformation includes generating derivative variables, improving variable distribution characteristics, binning interval variables, and standardizing interval variables;

[0093] Step 2: Feature extraction and semantic segmentation:

[0094] Use deep learning models to perform feature extraction and semantic segmentation on the pre-processed point cloud data, dividing the point cloud data into different building structure parts;

[0095] Step 3: Point cloud registration and fusion:

[0096] A deep learning model is used to intelligently register and fuse two types of point clouds obtained from oblique photogrammetry and laser scanning;

[0097] Step 4: 3D model creation:

[0098] The 3D model is built on the registered and fused point cloud data through multi-view image joint aerial triangulation, dense matching point cloud generation, Delaunay triangulation construction and texture mapping;

[0099] Combined with building planning data, parameter generation rules are used to obtain parameter tables for building components, including a component parameter table and a docking parameter table. The component parameter table is a list of structural and mechanical parameters of building components, and the docking parameter table is a list of structural and mechanical parameters of docking between building components. A generative model is used to generate a three-dimensional model of the building component based on the parameter tables.

[0100] Step 5: Model verification and optimization:

[0101] According to the docking parameter form, the three-dimensional models of the building components are spliced ​​to obtain the overall three-dimensional model, and the splicing positions in the overall three-dimensional model are subjected to collision detection and smoothness verification. The smoothness verification verifies the surface smoothness of the outer surface of the splicing positions in the overall three-dimensional model; solid mechanics simulation and stress analysis are performed on the overall three-dimensional model to determine whether the analysis results meet the requirements of the building planning data; if the verification fails, the number of verification failures in this round of the process is determined. If the number of verification failures is greater than the preset value, the number verification fails, the process exits and returns an error message; if the number of verification failures is less than or equal to the preset value, the number verification succeeds and the process returns to the step of generating the three-dimensional model of the building component to adjust the parameters and optimize the model.

[0102] 360-degree drone photography, combined with oblique photography and laser scanning, forms a complementary data source, covering details of the building structure from all angles. In the preprocessing stage, the k-nearest neighbor algorithm is used to fill in missing point cloud data, and outliers are identified and removed using the Mahalanobis distance threshold to ensure the quality of the data for subsequent processing. The feature extraction network uses an improved PointNet++ architecture and introduces a channel attention mechanism to enhance the recognition of irregular-shaped components. During point cloud registration, a dual-branch network of texture and geometry is constructed, and the Sinkhorn algorithm is used to solve the optimal affine transformation parameters to achieve millimeter-level alignment accuracy. In the parameter form generation stage, component dimensions and mechanical indicators are automatically derived based on building planning data, and a physically consistent three-dimensional model is constructed using a generative adversarial network. In the verification phase, a collision detection algorithm is used to identify component interface interference, and curvature continuity analysis is used to verify splicing smoothness. Finite element simulation results are combined to determine whether rollback adjustments are necessary.

[0103] Compared with existing technologies, traditional methods rely on a single data source, resulting in incomplete modeling. This solution achieves comprehensive data coverage through multi-sensor collaborative acquisition. Existing segmentation algorithms struggle to handle irregularly shaped building components. This solution utilizes a multi-scale network with an attention mechanism to improve segmentation accuracy. Conventional parametric modeling lacks physical property constraints. This solution ensures mechanical compatibility between component interfaces by integrating parameter tables. Traditional verification processes rely on manual inspection. This solution establishes an automated multi-level verification system for closed-loop optimization.

[0104] Through the above-mentioned technical solution, this application effectively addresses the inefficiency and inaccurate modeling issues of traditional methods. Multi-source data fusion and intelligent registration technologies improve input data quality, parametric form-driven generative modeling ensures the accuracy of component physical properties, and automated verification mechanisms shorten design iteration cycles. This approach significantly improves the design efficiency and model reliability of complex building structures, providing a viable technical path for intelligent building design.

[0105] In this embodiment, the process of missing value processing in step 1 is to identify the coordinate point P (x i ,y i , z i ); Search the k nearest neighbor points N={P1, P2, ..., P k}, calculate the Euclidean distance between each neighboring point and the missing point:

[0106]

[0107] Calculate the weight according to the inverse of the distance and perform weighted filling:

[0108] (prevent division by zero);

[0109] (prevent division by zero);

[0110] (prevent division by zero);

[0111] The process of outlier judgment and processing is as follows: Calculate the mean vector of the data set And covariance matrix Σ; for each data point P(x,y,z), calculate its Mahalanobis distance:

[0112]

[0113] Set the significance level α = 0.05α = 0.05 and determine the threshold value through chi-square distribution: Judgment condition: When D Mahalanobis > threshold, mark the point as an outlier;

[0114] The process of removing duplicate data is as follows: define the coordinate matching tolerance ε=10 -6 , for any two points P i (x i ,y i , z i ) and P j (x j ,y j , z j ), calculate the absolute difference of coordinates: Δx=|x i -x j |, Δy=|y i -y j |, Δz=|z i -z j When Δx≤ε, Δy≤ε, and Δz≤ε are satisfied at the same time, the data is determined to be duplicate data. For the duplicate data group, only the first appearing data point is retained and the rest of the duplicate points are deleted.

[0115] Specifically, during the missing value processing phase, a spatial index structure is established to quickly locate valid data around missing points, and a distance-weighted mechanism is used to generate fill-in values. This method can better maintain the surface continuity of building structures than traditional mean-filling methods. Outlier detection uses a dynamic threshold judgment based on statistical distribution, reflecting the data distribution characteristics through the covariance matrix, overcoming the adaptability defects of fixed thresholds on different data sets. Duplicate data processing uses a multi-dimensional coordinate joint judgment mechanism to eliminate redundant data while retaining necessary details, providing topologically correct point cloud data for subsequent modeling. The three processing links form a progressive data cleaning process, in which spatial proximity filling ensures data integrity, statistical anomaly detection improves data reliability, and tolerance matching deduplication optimizes data storage efficiency.

[0116] Compared with existing technologies, traditional methods for handling missing values ​​often use global averaging or interpolation, which are difficult to adapt to the local geometric characteristics of building structures and can easily lead to surface distortion in subsequent modeling. Conventional anomaly detection relies on a fixed Euclidean distance threshold, which cannot effectively distinguish true outliers from data acquisition noise. Common deduplication methods use precise coordinate matching, which is easily limited by sensor accuracy and leads to excessive deletion. This solution maintains spatial structure through k-nearest neighbor local filling, improves anomaly recognition accuracy through a dynamic Mahalanobis distance threshold, and uses a tolerance matching mechanism to balance data accuracy and redundancy elimination, forming a preprocessing method that adapts to the characteristics of 3D point clouds.

[0117] Through the above technical solution, this application can improve the spatial consistency of missing value filling in 3D building structure data, enhance the adaptability of outlier detection to complex data distributions, and effectively reduce the false deletion rate in duplicate data processing, thereby providing high-quality input data for subsequent modeling. This preprocessing method, through multi-stage collaborative processing, overcomes the technical bottlenecks of traditional methods in local geometry preservation, statistical anomaly identification, and data redundancy control.

[0118] In this embodiment, the feature extraction and semantic segmentation of point cloud data in step 2 include A. point cloud data preprocessing:

[0119] Use voxelized grid to downsample the original point cloud. The grid side length is calculated as:

[0120]

[0121] Where N is the preset resolution parameter (range 50-200), δ is the minimum voxel threshold (0.01-0.1 m);

[0122] The noise points are filtered out by the statistical outlier removal algorithm, and the standard deviation of the neighborhood distance of point pi is defined as:

[0123]

[0124] If σ i >α·median({σ1,...,σ n}), then remove P i , where α = 1.5-3.0;

[0125] Then normalize the coordinates

[0126]

[0127] where p min , p max is the global coordinate extreme value of the point cloud;

[0128] B. Multi-scale hierarchical feature extraction: Build a backbone network based on the improved PointNet++, each layer contains a local feature aggregation module,

[0129]

[0130] Where φ is 1D convolution and MLP is multi-layer perceptron;

[0131] Introducing the channel attention mechanism, the weight calculation is:

[0132] w c =σ(w2·ReLU(w1·GAP(f)))

[0133] Where GAP is global average pooling, w1 and w2 are learnable parameters;

[0134] Adopting adaptive feature fusion decoder, the upsampling formula is:

[0135]

[0136] Where ⊕ represents the feature concatenation of the jump connection;

[0137] Define the multi-task loss function:

[0138]

[0139] in is the cross entropy loss, is Lovasz-Softmax loss, λ2=0.7, λ2=0.3, λ3=0.001;

[0140] D. Post-processing optimization: Conditional random field (CRF) is used to optimize the segmentation results. The energy function is:

[0141]

[0142] in θ α =0.1,θ β =0.01;

[0143] Holes are filled in door and window components, and the morphological closing operation kernel size is 3×3×3 voxels.

[0144] Specifically, during the missing value processing phase, a spatial index structure is established to quickly locate valid data around missing points, and a distance-weighted mechanism is used to generate fill-in values. This method can better maintain the surface continuity of building structures than traditional mean-filling methods. Outlier detection uses a dynamic threshold judgment based on statistical distribution, reflecting the data distribution characteristics through the covariance matrix, overcoming the adaptability defects of fixed thresholds on different data sets. Duplicate data processing uses a multi-dimensional coordinate joint judgment mechanism to eliminate redundant data while retaining necessary details, providing topologically correct point cloud data for subsequent modeling. The three processing links form a progressive data cleaning process, in which spatial proximity filling ensures data integrity, statistical anomaly detection improves data reliability, and tolerance matching deduplication optimizes data storage efficiency.

[0145] Compared with existing technologies, traditional methods for handling missing values ​​often use global averaging or interpolation, which are difficult to adapt to the local geometric characteristics of building structures and can easily lead to surface distortion in subsequent modeling. Conventional anomaly detection relies on a fixed Euclidean distance threshold, which cannot effectively distinguish true outliers from data acquisition noise. Common deduplication methods use precise coordinate matching, which is easily limited by sensor accuracy and leads to excessive deletion. This solution maintains spatial structure through k-nearest neighbor local filling, improves anomaly recognition accuracy through a dynamic Mahalanobis distance threshold, and uses a tolerance matching mechanism to balance data accuracy and redundancy elimination, forming a preprocessing method that adapts to the characteristics of 3D point clouds.

[0146] Through the above technical solution, this application can improve the spatial consistency of missing value filling in 3D building structure data, enhance the adaptability of outlier detection to complex data distributions, and effectively reduce the false deletion rate in duplicate data processing, thereby providing high-quality input data for subsequent modeling. This preprocessing method, through multi-stage collaborative processing, overcomes the technical bottlenecks of traditional methods in local geometry preservation, statistical anomaly identification, and data redundancy control.

[0147] Specifically, in the point cloud data preprocessing stage, the point cloud density is adjusted through voxelized grid downsampling, and noise points are dynamically eliminated by combining a statistical outlier removal algorithm. Subsequently, global coordinate normalization is performed to form standardized input data. In the multi-scale hierarchical feature extraction stage, the improved PointNet++ backbone network extracts geometric features layer by layer through a local feature aggregation module. The channel attention mechanism automatically selects effective feature channels, and the adaptive feature fusion decoder integrates multi-scale features to generate semantic segmentation results. In the post-processing optimization stage, the conditional random field uses spatial constraints to optimize the continuity of segmentation boundaries, and morphological closing operations repair hole defects in door and window components. This solution optimizes input quality through data preprocessing, enhances feature expression capabilities through multi-scale feature extraction, and improves segmentation accuracy through post-processing optimization, forming a complete processing closed loop.

[0148] Compared with existing technologies, traditional methods using fixed threshold denoising are prone to loss of valid data, single-scale feature extraction struggles to capture complex structural details, and post-segmentation processing lacks spatial continuity constraints. This solution significantly improves the segmentation accuracy of complex building structures through dynamic threshold denoising to retain valid data points, multi-level feature aggregation to capture geometric features at different scales, conditional random field optimization to ensure segmentation boundaries conform to physical structural continuity, and morphological processing to repair topological defects.

[0149] Through the above technical solution, this application effectively reduces the impact of noise interference on semantic segmentation, enhances the ability to express multi-scale features, improves the accuracy of building structure division, ensures that the generated three-dimensional model has complete structural details and reasonable topological connections, and provides high-quality semantic segmentation results for subsequent model establishment.

[0150] In this embodiment, the processing process in step 3 is: and laser point cloud Multi-scale features are extracted through a dual-branch deep network:

[0151] Texture branch: Use 3D sparse convolution kernel K on S tex Extract texture features F S =σ(W tex *S+b tex )

[0152] Geometry branch: Adaptive radius MLP is used to extract geometric features from TT

[0153] Where σ is the LeakyReLU activation function, φ r Represents the maximum pooling operation within the neighborhood of radius r;

[0154] Differentiable Quasi-Optimization: Constructing a Similarity Matrix Based on Attention Mechanism:

[0155]

[0156] Solving the optimal transmission matrix using the Sinkhorn algorithm And calculate the affine transformation parameters:

[0157]

[0158] The joint loss function calculates the total loss, which includes the improved Chamfer distance, cosine similarity loss, and the quasi-error distribution distributed by the discriminator DD area. The total loss function is

[0159]

[0160] Among them, λ1, λ2, λ3 are dynamically adjusted weights, satisfying λ1+λ2+λ3=1. Then, the fusion quality index is calculated in level ll. and The next level of refinement is triggered when , otherwise it returns to the differentiable allocation quasi-optimization.

[0161] Specifically, to address the issue of feature discrepancies between heterogeneous point clouds, a feature complementation mechanism is constructed through a two-branch network: the texture branch leverages the anisotropic properties of 3D sparse convolutions to capture building surface details, while the geometry branch accurately extracts structural contours through an adaptive radius mechanism. The differentiable registration module establishes cross-modal feature associations with the help of an attention mechanism and solves the optimal transfer matrix using the Sinkhorn algorithm to overcome the sensitivity of traditional registration methods to initial position. The improved Chamfer distance in the joint loss function strengthens global structural alignment, while cosine similarity enhances local feature matching. A discriminator-driven error distribution loss specifically optimizes system error. Dynamic weight adjustment automatically balances loss terms based on the hierarchical registration status. Hierarchical refinement is triggered when the fusion quality indicator reaches a predetermined threshold, forming a progressive optimization mechanism.

[0162] Compared with existing technologies, traditional point cloud registration methods rely on iterative closest point algorithms, which are sensitive to initial position and have low feature utilization. This solution achieves end-to-end feature alignment through a deep learning network, uses an attention mechanism to establish cross-modal associations, and combines optimal transmission theory to achieve robust registration. Existing technologies use fixed-radius search, resulting in incomplete geometric feature extraction. This solution's adaptive radius mechanism effectively adapts to changes in point cloud density and improves feature representation capabilities. Traditional methods use a single loss function to make it difficult to balance different optimization objectives. This solution's dynamic weighting mechanism achieves multi-objective collaborative optimization.

[0163] Through the above-mentioned technical solution, this application effectively solves the technical problem of low registration accuracy of heterogeneous point cloud data and achieves high-precision fusion of building structural features. Specifically, the dual-branch network architecture effectively separates and strengthens the advantageous features of heterogeneous data, the adaptive radius mechanism improves the integrity of geometric feature extraction, the optimal transmission algorithm enhances the robustness of the registration process, and the dynamic weighting strategy ensures the coordination of multi-objective optimization. This solution significantly improves the point cloud fusion quality of complex building structures and can adapt to the needs of heterogeneous point cloud fusion with different acquisition conditions and different accuracies.

[0164] In this embodiment, in step 4, the projection matrices of at least two cameras and the matching image points are used to calculate the coordinates of the three-dimensional space point using the following formula:

[0165]

[0166] Among them, (u, v) and (u′, v′) are the coordinates of the matching image points, P and P′ are the camera projection matrices, and X=[X x , Xy , X z , 1] T is the homogeneous coordinate of a three-dimensional point;

[0167] The iterative closest point (ICP) algorithm is used to solve the point cloud registration parameters by minimizing the following error function:

[0168]

[0169] in, and is the point cloud to be registered, R is the rotation matrix, and t is the translation vector;

[0170] Based on the registered point cloud data, a divide-and-conquer algorithm is used to generate the Delaunay triangulation, ensuring that the circumscribed circle of any triangle does not contain other points.

[0171] The texture coordinates of any point inside the triangle are calculated by interpolating the barycentric coordinates. The formula is as follows:

[0172] V=αV A +βV B +γV C (α+β+γ=1)

[0173] Among them, V A ,V B ,V C are triangle vertex attributes (such as texture coordinates), α, β, γ are barycentric coordinates.

[0174] In this embodiment, in step 4, the component parameter form and the docking parameter form are calculated using parameter generation rules in combination with the building planning data: The component parameter form includes the structural parameters (such as beam cross-sectional area and length) and mechanical parameters (such as volume and bearing capacity) of the component:

[0175] V beam = cross-sectional area × length, butt joint parameter form: describes the structural parameters and mechanical parameters of the butt joint between components;

[0176] Based on the parameter form, a generative model (such as a deep learning network) is used to generate a 3D model of the building component. The model output satisfies the following constraints:

[0177] Model accuracy ≥∈ and mechanical parameter error ≤δ, where ∈ is the geometric accuracy threshold and δ is the mechanical parameter error threshold.

[0178] Specifically, a three-dimensional point coordinate calculation equation is established through the multi-camera projection matrix, and the pixel coordinates of the matching image points and the camera parameters are input to solve the spatial point position, eliminating the data missing problem caused by single-view observation. An iterative nearest point algorithm with convergence condition constraints is used to optimize the rigid body transformation parameters of the point cloud to be registered. High-precision spatial alignment is achieved by cyclically executing the nearest neighbor search and minimizing the error function. During the Delaunay triangulation generation stage, the point cloud is divided into several subsets for parallel processing. Local triangulations are constructed separately and then merged to form a global topological structure, reducing the computational complexity while maintaining the empty circumscribed circle characteristics. The texture attributes of the triangle vertices are linearly interpolated to any internal point using the barycentric coordinates. The texture coordinates are continuously distributed along the surface according to the geometric weight distribution mechanism, avoiding the stretching deformation problem caused by linear mapping.

[0179] Compared with existing technologies, traditional methods rely on single-sensor data, resulting in incomplete spatial information; fixed-threshold registration is susceptible to noise; serial triangulation algorithms are inefficient in processing large amounts of data; and simple linear interpolation in texture mapping results in noticeable seams. This solution enhances spatial information integrity through collaborative computation of multi-source data, adaptively optimizes registration parameters to improve the alignment accuracy of heterogeneous point clouds, uses a parallel processing mechanism to accelerate the modeling of complex structures, and uses a geometric weighted interpolation algorithm to enhance texture mapping quality.

[0180] Through the above technical solution, this application realizes the accurate reconstruction of multi-perspective three-dimensional coordinates and the efficient registration of heterogeneous point clouds, solves the problem of slow modeling of large-scale building point clouds, effectively avoids the distortion phenomenon in the texture mapping process, and enables the generated three-dimensional model to have the dual advantages of high geometric accuracy and strong visual realism.

[0181] Example 2

[0182] An intelligent design system for building structures based on 3D modeling includes: a data acquisition module: used to obtain relevant data of building structures through multiple data acquisition methods, including drone modules, oblique photogrammetry equipment and laser scanning equipment, to realize the acquisition of three-dimensional data of building structures.

[0183] Preprocessing module: used to clean, transform and filter the collected 3D data to obtain preprocessed 3D data; this module has a data processing algorithm that can automatically complete operations such as missing value processing, outlier judgment and processing, and duplicate data removal, and perform data transformation to meet subsequent processing requirements;

[0184] Feature extraction and semantic segmentation module: Uses a deep learning model to perform feature extraction and semantic segmentation on preprocessed point cloud data, dividing the point cloud data into different building structure parts; this module contains a trained deep learning model that can accurately analyze and process the input point cloud data.

[0185] Point cloud registration and fusion module: This module uses a deep learning model to intelligently register and fuse two types of point clouds acquired through oblique photogrammetry and laser scanning. This module can automatically achieve accurate registration and fusion of point clouds without manual intervention.

[0186] 3D model building module: This module is used to build a 3D model of the registered and fused point cloud data through multi-view image joint aerial triangulation, dense matching point cloud generation, Delaunay triangulation construction, and texture mapping. It also combines architectural planning data with parameter generation rules and generative models to generate 3D models of building components.

[0187] Model verification and optimization module: splice the 3D models of building components according to the docking parameter form to obtain the overall 3D model, perform collision detection and smoothness verification on the splicing positions in the overall 3D model; perform solid mechanics simulation and stress analysis on the overall 3D model to determine whether the analysis results meet the requirements of the building planning data; if the verification fails, perform corresponding processing based on the number of verification failures, and return to the step of generating the 3D model of the building components to adjust parameters and optimize the model.

[0188] Specifically, after the system acquires multi-source three-dimensional data through the data acquisition module, the pre-processing module automatically performs data cleaning and standardization to eliminate noise and inconsistencies. The feature extraction and semantic segmentation module performs deep feature learning on the cleaned point cloud to achieve intelligent recognition and classification of building components. The point cloud registration and fusion module uses deep learning algorithms to achieve spatial alignment and data fusion of oblique photography and laser point clouds, solving the problem of multi-source data collaboration. The three-dimensional model building module combines point cloud reconstruction and parametric generation technology to simultaneously construct geometric models and mechanical parametric models. The model verification and optimization module performs multi-dimensional verification through collision detection, smoothing verification and mechanical simulation. When model defects are detected, parameter adjustment and iterative optimization are automatically triggered to form a closed-loop quality control mechanism.

[0189] Compared with existing technologies, traditional systems typically use a single sensor to collect data and rely on manual cleaning, resulting in limited data integrity and processing efficiency. Existing point cloud registration methods often use ICP algorithms, which have difficulty processing cross-modal data, and the separation of model generation and verification leads to long iteration cycles. This system improves data quality through multi-sensor collaborative acquisition and automated preprocessing, leverages deep learning for efficient registration of multi-source point clouds, and combines generative models with physical simulation to create a closed-loop optimization process, effectively addressing the data fragmentation, manual dependency, and delayed verification issues inherent in traditional methods.

[0190] Through the above technical solutions, this application realizes the full-chain automation of the building structure design process, eliminates the bottleneck of manual operation through the collaborative work of multiple modules, uses deep learning to improve the ability to handle complex structures, and combines a closed-loop verification mechanism to ensure model accuracy and engineering compliance, ultimately solving the technical problems of low efficiency, modeling difficulties and insufficient coordination of multi-source data in traditional methods.

[0191] Example 3:

[0192] An electronic device includes: a memory for storing one or more programs; and a processor. When the one or more programs are executed by the processor, an intelligent design method for building structures based on 3D modeling of embodiment 1 is implemented.

[0193] The program stored in the memory contains data acquisition instructions, point cloud processing algorithms, and model verification logic. When the processor executes the program, it first drives the drone and oblique photography equipment to collect three-dimensional building data. It then fills in missing values ​​and removes outliers in the data, completing point cloud segmentation and fusion through a deep learning model. During the three-dimensional modeling stage, the program automatically generates component parameter forms based on the building planning data and uses generative models to construct three-dimensional models of building components. During model verification, the processor calls the collision detection algorithm and mechanical simulation module to perform stress analysis on the entire model. If the verification fails, the parameter fallback mechanism is triggered to re-optimize the model. The entire process achieves full automation of building structure design through the collaboration of memory and processor.

[0194] Compared with existing technologies, traditional architectural design relies on manual operations to complete point cloud processing and model assembly, which is inefficient and difficult to handle complex structures. However, this solution integrates data acquisition, feature segmentation, model generation, and verification into an automated process through programmatic control of electronic equipment. It leverages the parallel computing power of processors to rapidly process massive amounts of point cloud data, while using memory to store intermediate results for iterative optimization, significantly improving modeling accuracy and processing efficiency.

[0195] Through the above technical solution, this application solves the problems of low efficiency of manual modeling and difficulty in processing complex structures, realizes the full process automation of building structure design from data collection to model optimization, reduces human intervention while ensuring the precise matching of component parameters and mechanical properties, and provides hardware support for the intelligent design of complex building structures.

[0196] Example 4:

[0197] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a 3D modeling-based intelligent design method for building structures according to a first embodiment.

[0198] When the computer program is executed by the processor, the three-dimensional data of the building structure is collected and standardized, and then automatically optimized for data quality using a preset preprocessing algorithm. The point cloud feature extraction module calls on a trained deep learning model to perform multi-scale feature analysis on the denoised point cloud data, enhancing the expression of key features through a channel attention mechanism. During the registration and fusion phase, the two types of point cloud data undergo feature extraction through the texture branch and geometry branch, respectively, and are then spatially aligned using a differentiable registration algorithm to form a complete fused point cloud dataset. The parameter form generation module automatically derives key parameters such as component size and connection method based on the design parameters in the building planning data, and uses a generative model to construct a three-dimensional model that meets the parameter constraints. During the verification phase, a programmed detection algorithm is used to detect geometric interference at the model joints, while the finite element analysis module is used to verify the mechanical properties. If the test results do not meet the preset standards, the system automatically triggers parameter adjustment instructions and regenerates the optimized model.

[0199] Compared with existing technologies, traditional computer-aided design systems require manual intervention in key links such as data preprocessing and model registration, resulting in low operational efficiency and unstable execution results. This solution, by solidifying the entire processing flow into a computer executable program, achieves full process automation from data acquisition to model optimization. Existing storage media usually only store design result data, but this solution integrates data processing algorithms and verification programs into the medium, making the design process repeatable. Compared with model optimization methods that rely on manual experience and judgment, this solution achieves closed-loop optimization through a programmed verification-feedback mechanism, ensuring the consistency and reliability of the design results.

[0200] Through the above technical solutions, this application achieves procedural control of the building structure design process, eliminating execution deviations caused by manual operations. The automated execution of the preprocessing algorithm ensures the quality consistency of the input data, and the application of deep learning methods improves the processing accuracy of feature extraction and data alignment. The combination of parametric modeling and procedural verification enables the design optimization process to form a standardized processing flow. The application of computer-readable storage media enables the complete design method to be accurately reproduced and executed in batches, solving the technical problems of low operational efficiency and untraceable execution results in the traditional design process.

[0201] The above are merely preferred embodiments of the present application and are not intended to limit the present application. Those skilled in the art will readily appreciate that various modifications and variations are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A method for intelligent design of building structures based on 3D modeling, characterized in that: The following steps are involved: Step 1: Data collection and preprocessing: Relevant data of the building structure is obtained through various data collection methods, including but not limited to 360° drone acquisition of three-dimensional data of the building structure, oblique photogrammetry, and laser scanning to obtain point cloud data. Perform data cleaning, data transformation, and data screening on the collected 3D data to obtain pre-processed 3D data; data cleaning includes missing value processing, outlier judgment and processing, and duplicate data removal; data transformation includes generating derivative variables, improving variable distribution characteristics, binning interval variables, and standardizing interval variables; Step 2: Feature extraction and semantic segmentation: Use deep learning models to perform feature extraction and semantic segmentation on the pre-processed point cloud data, dividing the point cloud data into different building structure parts; Step 3: Point cloud registration and fusion: A deep learning model is used to intelligently register and fuse two types of point clouds obtained from oblique photogrammetry and laser scanning; Step 4: 3D model creation: The 3D model is built on the registered and fused point cloud data through multi-view image joint aerial triangulation, dense matching point cloud generation, Delaunay triangulation construction and texture mapping; Combined with building planning data, parameter generation rules are used to obtain parameter tables for building components, including a component parameter table and a docking parameter table. The component parameter table is a list of structural and mechanical parameters of building components, and the docking parameter table is a list of structural and mechanical parameters of docking between building components. A generative model is used to generate a three-dimensional model of the building component based on the parameter tables. Step 5: Model verification and optimization: According to the docking parameter form, the three-dimensional models of the building components are spliced ​​to obtain the overall three-dimensional model, and the splicing positions in the overall three-dimensional model are subjected to collision detection and smoothness verification. The smoothness verification verifies the surface smoothness of the outer surface of the splicing positions in the overall three-dimensional model; solid mechanics simulation and stress analysis are performed on the overall three-dimensional model to determine whether the analysis results meet the requirements of the building planning data; if the verification fails, the number of verification failures in this round of the process is determined. If the number of verification failures is greater than the preset value, the number verification fails, the process exits and returns an error message; if the number of verification failures is less than or equal to the preset value, the number verification succeeds and the process returns to the step of generating the three-dimensional model of the building component to adjust the parameters and optimize the model.

2. The method for intelligent design of building structures based on 3D modeling according to claim 1, characterized in that: The process of missing value processing in step 1 is to identify the coordinate point P(x i ,y i , z i ); Search the k nearest neighbor points N={P1, P2, ..., P k }, calculate the Euclidean distance between each neighboring point and the missing point: Calculate the weight according to the inverse of the distance and perform weighted filling: The process of outlier judgment and processing is as follows: Calculate the mean vector of the data set And covariance matrix Σ; for each data point P(x,y,z), calculate its Mahalanobis distance: Set the significance level α = 0.05α = 0.05 and determine the threshold value through chi-square distribution: Judgment condition: When D Mahalanobis > threshold, mark the point as an outlier; The process of removing duplicate data is as follows: define the coordinate matching tolerance ε=10 -6 , for any two points P i (x i ,y i , z i ) and P j (x j ,y j , z j ), calculate the absolute difference of coordinates: Δx=|x i -x j |, Δy=|y i -y j |, Δz=|z i -z j When Δx≤ε, Δy≤ε, and Δz≤ε are satisfied at the same time, the data is determined to be duplicate data. For the duplicate data group, only the first appearing data point is retained and the rest of the duplicate points are deleted.

3. The intelligent design method of building structure based on 3D modeling according to claim 1, characterized in that: In step 2, feature extraction and semantic segmentation of point cloud data include A. point cloud data preprocessing: Use voxelized grid to downsample the original point cloud. The grid side length is calculated as: Where N is the preset resolution parameter (range 50-200), δ is the minimum voxel threshold (0.01-0.1 m); The noise points are filtered out by the statistical outlier removal algorithm, and the standard deviation of the neighborhood distance of point pi is defined as: If σ i >α·median({σ1,...,σ n }), then remove P i , where α = 1.5-3.0; Then normalize the coordinates Among them, P min , p max is the global coordinate extreme value of the point cloud; B. Multi-scale hierarchical feature extraction: Build a backbone network based on the improved PointNet++, each layer contains a local feature aggregation module, Where φ is 1D convolution and MLP is multi-layer perceptron; Introducing the channel attention mechanism, the weight calculation is: In c =σ(w2·ReLU(w1·GAP(f))) Where GAP is global average pooling, w1 and w2 are learnable parameters; Adopting adaptive feature fusion decoder, the upsampling formula is: in Feature concatenation representing skip connections; Define the multi-task loss function: in is the cross entropy loss, is Lovasz-Softmax loss, λ2=0.7, λ2=0.3, λ3=0.001; D. Post-processing optimization: Conditional random field (CRF) is used to optimize the segmentation results. The energy function is: in Holes are filled in door and window components, and the morphological closing operation kernel size is 3×3×3 voxels.

4. The method for intelligent design of building structures based on 3D modeling according to claim 1, characterized in that: The processing process in step 3 is: and laser point cloud Extract multi-scale by dual-branch deep network feature: Texture Branch: Apply 3D sparse convolution kernel K to S tex Extract texture features F S =σ(W tex *S+b tex ) Geometry branch: Adaptive radius MLP is used to extract geometric features from TT Where σ is the LeakyReLU activation function, φ r Represents the maximum pooling operation within the neighborhood of radius r; Differentiable Quasi-Optimization: Constructing a Similarity Matrix Based on Attention Mechanism: Solving the optimal transmission matrix using the Sinkhorn algorithm And calculate the affine transformation parameters: The joint loss function calculates the total loss, which includes the improved Chamfer distance, cosine similarity loss, and the quasi-error distribution distributed by the discriminator DD area. The total loss function is Among them, λ1, λ2, λ3 are dynamically adjusted weights, satisfying λ1+λ2+λ3=1. Then, the fusion quality index is calculated in level ll. When HD l <τ geo And TC l <τ tex The next level of refinement is triggered when , otherwise it returns to the differentiable allocation quasi-optimization.

5. The method for intelligent design of building structures based on 3D modeling according to claim 1, characterized in that: In step 4, the projection matrices of at least two cameras and the matching image points are used to calculate the coordinates of the three-dimensional space point using the following formula: Among them, (u, v) and (u′, v′) are the coordinates of the matching image points, P and P′ are the camera projection matrices, and X=[X x , X y , X z , 1] T is the homogeneous coordinate of a three-dimensional point; The iterative closest point (ICP) algorithm is used to solve the point cloud registration parameters by minimizing the following error function: in, and is the point cloud to be registered, R is the rotation matrix, and t is the translation vector; Based on the registered point cloud data, a divide-and-conquer algorithm is used to generate the Delaunay triangulation, ensuring that the circumscribed circle of any triangle does not contain other points. The texture coordinates of any point inside the triangle are calculated by interpolating the barycentric coordinates. The formula is as follows: V=αV A +βV B +γV C (a+b+c=1) Among them, V A ,V B ,V C are triangle vertex attributes (such as texture coordinates), α, β, γ are barycentric coordinates.

6. The method for intelligent design of building structures based on 3D modeling according to claim 5, characterized in that: In step 4, the component parameter form and the docking parameter form are calculated using parameter generation rules in combination with the building planning data: The component parameter form includes the structural parameters (such as beam cross-sectional area and length) and mechanical parameters (such as volume and bearing capacity) of the component: V beam = cross-sectional area × length, butt joint parameter form: describes the structural parameters and mechanical parameters of the butt joint between components; Based on the parameter form, a generative model (such as a deep learning network) is used to generate a 3D model of the building component. The model output satisfies the following constraints: Model accuracy ≥∈ and mechanical parameter error ≤δ, where ∈ is the geometric accuracy threshold and δ is the mechanical parameter error threshold.

7. An intelligent design system for building structures based on 3D modeling, characterized in that: include: Data acquisition module: used to obtain relevant data of building structures through various data acquisition methods, including drone modules, oblique photogrammetry equipment and laser scanning equipment, to realize the collection of three-dimensional data of building structures. Preprocessing module: used to clean, transform and filter the collected 3D data to obtain preprocessed 3D data; this module has a data processing algorithm that can automatically complete operations such as missing value processing, outlier judgment and processing, and duplicate data removal, and perform data transformation to meet subsequent processing requirements; Feature extraction and semantic segmentation module: Uses a deep learning model to perform feature extraction and semantic segmentation on preprocessed point cloud data, dividing the point cloud data into different building structure parts; this module contains a trained deep learning model that can accurately analyze and process the input point cloud data. Point cloud registration and fusion module: This module uses a deep learning model to intelligently register and fuse two types of point clouds acquired through oblique photogrammetry and laser scanning. This module can automatically achieve accurate registration and fusion of point clouds without manual intervention. 3D model building module: This module is used to build a 3D model of the registered and fused point cloud data through multi-view image joint aerial triangulation, dense matching point cloud generation, Delaunay triangulation construction, and texture mapping. It also combines architectural planning data with parameter generation rules and generative models to generate 3D models of building components. Model Verification and Optimization Module: This module splices the 3D models of building components according to the docking parameter form to obtain an overall 3D model, and performs collision detection and smoothness verification on the splicing positions in the overall 3D model. Perform solid mechanics simulation and stress analysis on the overall 3D model to determine whether the analysis results meet the requirements of the building planning data; If the verification fails, corresponding processing is performed according to the number of verification failures, and the process returns to the step of generating the three-dimensional model of the building component to adjust parameters and optimize the model.

8. An electronic device, characterized in that: include: a memory for storing one or more programs; Processor; when the one or more programs are executed by the processor, any one of the 3D modeling-based intelligent design methods for building structures as described in claims 1-7 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements any one of the 3D modeling-based intelligent design methods for building structures as described in claims 1-7.

Images