Fabricated building construction site arrangement method based on generative adversarial network

Abstract

The invention provides a fabricated building construction site arrangement method based on a generative adversarial network, and relates to the technical field of building construction site arrangement. Historical construction site arrangement drawings are collected to form an initial data set, a genetic algorithm is adopted to carry out iterative optimization on the initial data set to generate a training data set, and a generative adversarial network is trained based on the training data set; acquiring a site constraint condition of a current project and a bounding box, coordinates and categories of a fixed facility, converting the site constraint condition and the bounding box, coordinates and categories into binary mask representation, inputting the binary mask representation into the trained generative adversarial network, and outputting optimized layout data; and extracting coordinates, bounding boxes and categories of all facilities in the optimized layout data, and carrying out modeling registration based on the initial site model to generate a construction site layout model. The technical problems that an existing construction site arrangement method depends on artificial experience, lacks data support and is low in automation degree can be solved.

Description

Technical Field

[0001] This invention relates to the field of construction site layout technology, and in particular to a method for arranging prefabricated building construction sites based on adversarial generative networks. Background Technology

[0002] The traditional cast-in-place construction method, which the construction industry has long relied on, is facing increasing bottlenecks in resource consumption, labor supply, and quality control, becoming an intrinsic driving force for the development of prefabricated construction technology. Meanwhile, the layout of prefabricated building construction sites is particularly important. However, in actual engineering scenarios, due to multiple constraints such as planning concepts, spatial conditions, management levels, and technology applications, many problems still need to be solved, directly affecting subsequent construction efficiency and safety. The layout of prefabricated building construction sites must simultaneously meet multi-dimensional constraints, including spatial boundary restrictions, facility spacing safety standards, crane coverage requirements, and construction process connection logic. This is a multi-objective, multi-constraint combinatorial optimization problem. Traditional manual planning cannot traverse all feasible solution spaces within a finite time. Existing optimization algorithms, such as genetic algorithms and ant colony algorithms, suffer from slow convergence speeds, susceptibility to local optima, and difficulty in incorporating industry standard knowledge when dealing with complex site boundaries. Prefabricated building construction site layout based on adversarial generative systems provides a new solution to these problems. In prefabricated building construction site layout, the optimized layout of various facilities plays a crucial role in improving the efficiency of prefabricated construction transportation. Currently, the advantages of prefabricated buildings rely on the coordinated management of factory prefabrication and on-site assembly, but the application of construction site layout optimization in prefabricated buildings is negligible. In terms of methods, existing inventions mostly collect site information through oblique photography and optimize the layout using ant colony algorithms and genetic algorithms, lacking data support for initial site layout plans and a system for generating prefabricated building construction site layouts. Construction site layout and optimization are prerequisites for prefabricated buildings to realize their advantages of high efficiency, low carbon emissions, and high quality. Their role spans the entire construction cycle, directly impacting not only the achievement of core objectives such as schedule, cost, safety, and quality, but also maximizing the industrial value of prefabricated buildings.

[0003] Chinese patent application CN111581818A discloses a construction site layout optimization method based on BIM technology. This method decomposes the construction site layout drawing into individual drawings of buildings, roads, construction equipment, material storage areas, etc., and establishes BIM models for each. It uses 3D animation simulation to model the construction process, identifying unreasonable site layouts and optimizing the layout for different construction stages to achieve seamless integration between the foundation, main structure, and decoration stages. While this patent can intuitively display the construction site layout and identify potential problems through BIM 3D simulation, this method still relies on manual modeling and layout adjustments of individual drawings, lacking the ability to automatically generate layout schemes. Furthermore, it fails to utilize historical construction site layout data for intelligent learning, thus failing to achieve automatic generation of data-driven site layout schemes. Summary of the Invention

[0004] In view of this, the present invention aims to solve the technical problems of existing construction site layout methods relying on manual experience, lacking data support, and having a low degree of automation. It provides a prefabricated building construction site layout method based on adversarial generative networks. By learning from historical layout data and optimizing multiple objectives, it can automatically generate prefabricated building construction site layout schemes under the conditions of meeting site constraints and construction specifications.

[0005] The technical solution of this invention is implemented as follows: This invention provides a method for arranging construction site layout for prefabricated buildings based on generative adversarial networks, comprising: S1. Collect multi-view images of the construction site and perform three-dimensional reconstruction to obtain a point cloud mesh model. Extract geometric parameters and spatial coordinates to construct an initial site model. S2. Collect historical construction site layout drawings, simplify the facilities by rectangularization, and mark the center coordinates, bounding boxes and categories to form an initial dataset. Use a genetic algorithm based on a multi-objective function to iteratively optimize the initial dataset to generate a training dataset. Train an adversarial generative network based on the training dataset. S3. Obtain the site constraints and bounding boxes, coordinates, and categories of fixed facilities for the current project, convert them into binary mask representations, and input them into the trained Generative Adversarial Network (GAN). The GAN uses a shape encoder to encode and decode the binary mask to obtain encoded features containing constraint and fixed facility information. The generator generates candidate layout data for temporary facilities based on the encoded features and sampling vectors. The discriminator judges and standardizes the complete layout containing fixed and temporary facilities and provides feedback for adjustment. The optimizer iteratively eliminates facility overlaps and meets spacing requirements in the latent space, and outputs optimized layout data. S4. Extract the coordinates, bounding boxes, and categories of each facility from the optimized layout data, and perform modeling and registration based on the initial site model to generate a construction site layout model.

[0006] Preferably, the process of obtaining the point cloud mesh model in step S1 is as follows: Multi-view images of the construction site were collected around the entire perimeter, with continuity and overlapping areas between adjacent images; Feature point matching is performed on the image to extract common feature points, and dense point cloud is generated based on the common feature points to obtain three-dimensional point cloud data; The discrete point cloud data is connected into a continuous triangular mesh surface by meshing the 3D point cloud data. The image texture is then projected onto the mesh surface through texture mapping to obtain a point cloud mesh model containing geometric parameters and texture information.

[0007] Preferably, the method for simplifying the facility into a rectangular shape in step S2 is as follows: Establish a coordinate system by selecting the boundary of the construction site. The origin of the coordinate system is located at one corner of the boundary of the construction site and covers the entire construction site. Each facility is simplified into a rectangular region with its edges parallel to the coordinate axes. The i-th facility is defined by its central coordinates. ,width and height Description, in which This indicates the horizontal length of the facility's boundary frame parallel to the X-axis. The vertical length of the facility's bounding box parallel to the Y-axis is represented by the initial dataset, which is formed by labeling the center coordinates, bounding box, and category label of each facility.

[0008] Preferably, the iterative optimization process using a genetic algorithm in step S2 includes: Step 1: Encode each layout scheme in the initial dataset as a chromosome. A chromosome consists of N gene segments, each corresponding to a facility and containing the center coordinates of that facility. , bounding box width and height And category labels, forming an initial population with a population size of M; Step 2: Calculate the multi-objective function for each layout scheme in the initial population. ,in The safety objective function is obtained by calculating the minimum distance between facilities; The objective function for noise pollution is obtained by calculating the weighted distance from the noise source facilities to the sensitive area. The objective function for crane construction efficiency is obtained by calculating the average distance from the crane to the precast component storage area and the building under construction. The objective function for the impact on surrounding roads is obtained by calculating the minimum distance from the construction facilities to the road at the site boundary. Step 3: Based on the Pareto dominance relationship, perform non-dominated ranking of the layout schemes in the population. If layout scheme A is not inferior to layout scheme B on all objective functions and is superior to layout scheme B on at least one objective function, then A dominates B, and the population is divided into multiple non-dominated levels. Step 4: Select parent individuals based on non-dominated hierarchy and crowding distance. Individuals with higher non-dominated hierarchy are more likely to be selected, and individuals with larger crowding distance within the same hierarchy are more likely to be selected. Crowding distance is obtained by calculating the distance between an individual and its neighboring individuals in the target space. Step 5: Perform a crossover operation on the selected parent individuals. Randomly select two parent individuals and K facilities as crossover points. Swap the center coordinates of the corresponding facilities in the two parent individuals to generate two child individuals. Step 6: Perform a mutation operation on the offspring individuals. Randomly select facilities from the offspring individuals according to the set mutation probability, and add a random perturbation that follows a Gaussian distribution to the center coordinates of the selected facilities. The perturbation amplitude shall not exceed 10% of the site size. Step 7: Perform boundary constraint checks on the offspring individuals. If the facility's center coordinates exceed the site boundary after adding or subtracting the half-width and half-height of the boundary frame, adjust the center coordinates until the facility is completely within the site boundary. Step 8: Add the offspring individuals to the population, recalculate the multi-objective function values ​​of all individuals and perform non-dominated sorting, retain the M individuals with higher non-dominated levels as the new population; Step 9: Determine whether the maximum number of iterations has been reached or whether the Pareto front has not changed significantly for several consecutive generations. If so, terminate the iteration; otherwise, return to step 4. Step 10: Extract all layout schemes with the first non-dominated level from the final population as the Pareto optimal solution set, and add the Pareto optimal solution set to the initial dataset to expand and form the training dataset.

[0009] Preferably, the site constraints in step S3 include boundary constraints, spacing constraints, and crane coverage constraints. The boundary constraints require that the boundary frames of all facilities must not exceed the construction site boundaries. The spacing constraints require that the minimum distance between facilities be no less than 5m. The crane coverage constraints require that the operating radius of the tower crane completely covers the area of ​​the building under construction. Fixed facilities include building and construction site gates; the boundary of a fixed facility is defined by its central coordinates. ,width ,high And category label description, where j is the fixed facility index.

[0010] Preferably, the training method for the generative adversarial network in step S2 is as follows: A conditional generation strategy is used to divide the layout schemes in the training dataset into fixed facility parts and temporary facility parts; The site constraints and fixed facilities are input into the generator as generation conditions. The generator generates the bounding box, coordinates and category labels of the temporary facilities based on the generation conditions and integrates them with the fixed facilities to form candidate layout data. The layout schemes in the training dataset are used as real samples, and the candidate layout data are used as generated samples. The discriminator uses a classifier to determine whether the input layout data is a real sample or a generated sample and outputs the real probability value. At the same time, an auxiliary self-supervised decoder is used to reconstruct the bounding boxes, coordinates and category labels of the facilities in the layout data. The discriminator's loss function includes the classification loss of real samples and generated samples, as well as the reconstruction error. The generator's loss function is to maximize the true probability value output by the discriminator for the generated samples. Adversarial training is performed by alternately optimizing the loss functions of the generator and discriminator. The generator continuously improves the generation quality so that the candidate layout data is closer to the distribution characteristics of the training dataset. The discriminator continuously improves its discrimination ability to distinguish between real samples and generated samples. When the discriminator can no longer effectively distinguish between real samples and generated samples, the training is completed, and the trained adversarial generative network is obtained.

[0011] Preferably, the process of generating candidate layout data for temporary facilities in step S3 is as follows: Convert the site constraints and fixed facility bounding boxes, coordinates, and categories of the current project into binary mask representations and input them into the trained shape encoder; The shape encoder uses a convolutional neural network to downsample and compress the binary mask to obtain a low-dimensional encoding vector containing current project constraints and fixed facility information as the encoding feature. A random latent vector is obtained by sampling from the standard normal distribution N(0,1), and the dimension of the random latent vector corresponds to the number of temporary facilities; A high-dimensional feature representation is obtained by inputting a random latent vector into a multilayer perceptron and performing a nonlinear transformation. The high-dimensional feature representation and the encoded features are fed into the conditional Transformer block, which uses the encoded features as conditional input to learn the spatial relationships between temporary facilities through a self-attention mechanism. The output of the conditional Transformer block is decoded by a fully connected layer to generate the center coordinates of the temporary facility. , bounding box width and height And category labels, where k is the temporary facility index; The generated temporary facility data is integrated with the fixed facility data to form candidate layout data.

[0012] Preferably, the method by which the discriminator judges and standardizes the candidate layout data and provides feedback for adjustment in step S3 is as follows: Input the candidate layout data into the trained discriminator; The classifier of the discriminator encodes the bounding boxes, coordinates and category labels of facilities in the candidate layout data into attribute vectors, and encodes site constraints into constraint vectors; Input the attribute vector and constraint vector along with the category tag into the conditional Transformer block. The category tag is used to aggregate global layout information. The Conditional Transformer block extracts the distribution features of candidate layout data through a self-attention mechanism. The output of the classification label is passed through a fully connected layer to output the probability value of the candidate layout data conforming to the distribution of the training dataset as the judgment result. The auxiliary self-supervised decoder reconstructs the bounding boxes, coordinates and category labels of facilities in the candidate layout data and calculates the reconstruction error as the standard verification result. If the probability value is lower than the set threshold or the reconstruction error is higher than the set threshold, the decision fails. The generator adjusts the random latent vector to regenerate candidate layout data and inputs it into the discriminator for decision again. The process iterates until the probability value is higher than the set threshold and the reconstruction error is lower than the set threshold to complete the decision and specification verification. The candidate layout data that has passed the verification is then output.

[0013] Preferably, the method by which the optimizer iteratively eliminates facility overlap and satisfies spacing requirements in the latent space in step S3 is as follows: The augmented Lagrange method is used to transform the 5m minimum spacing constraint into a non-overlapping constraint for the facility boundary box; Define the overlap loss function , where Z is the latent space code, L is the set of facility category labels, D represents the probability value of the trained discriminator that the generated layout conforms to the distribution of the training dataset, and G represents the layout data generated by the trained generator based on the latent space code Z and the category labels L. Define overlapping constraint functions , where O is the overlap judgment function that calculates the sum of the overlapping areas between all facility pairs in the generated layout; Construct the augmented Lagrangian objective function: ; in For Lagrange multipliers, For penalty parameters; By iteratively updating the latent space encoding Z and the Lagrange multipliers and penalty parameters The optimal latent space encoding that minimizes the objective function and satisfies the tolerance threshold requirement of the overlap constraint function is searched in the latent space. The optimal latent space encoding is input into the generator to obtain optimized layout data that eliminates overlap and meets the 5m spacing requirement.

[0014] Preferably, the modeling and registration method in step S4 is as follows: Extract the center coordinates, bounding boxes, and category labels of each facility from the optimized layout data. Import the initial site model into the BIM platform as the base model. Call the corresponding 3D model from the 3D component model library according to the facility category label. Determine the spatial position of the 3D model in the unified coordinate system of the initial site model according to the center coordinates of the facility. Scale the 3D model according to the width and height of the facility's bounding box to match the bounding box size. Place the scaled 3D model in the determined spatial position to complete the 3D modeling of the facility. Merge the 3D models of all facilities with the initial site model in the unified coordinate system to generate a complete construction site layout model that includes the site environment and facility layout.

[0015] The present invention has the following advantages over the prior art: (1) The point cloud mesh model of the construction site is obtained by using multi-view image three-dimensional reconstruction technology. Geometric parameters and spatial coordinates are extracted to construct the initial site model. The spatial information such as the topography and boundary shape of the real construction environment is accurately restored, providing a precise spatial base for the subsequent site layout plan. This avoids the errors that may be caused by manual measurement and hand-drawn drawings in traditional methods, and realizes the digital expression of construction site information. (2) An initial dataset was constructed by collecting historical prefabricated building construction site layout drawings. A genetic algorithm was used to iteratively optimize the dataset based on multiple objective functions such as safety, noise pollution, crane construction efficiency and road impact, generating a training dataset containing Pareto optimal solutions, thus overcoming the problem of inconsistent historical data quality. (3) The automatic generation of construction site layout scheme is achieved by using conditional generative adversarial network. The site constraints and fixed facilities are encoded by shape encoder. The generator generates layout data of temporary facilities based on the encoded features and random latent vectors. The discriminator performs dataset distribution matching judgment and construction specification verification on the generated layout through classifier and auxiliary self-supervised decoder. This realizes intelligent layout generation that combines data-driven and specification constraints, and improves the rationality and standardization of site layout scheme. (4) The optimizer uses the augmented Lagrange method to iteratively optimize in the latent space, transforming the facility spacing constraint into a non-overlapping constraint. By constructing an objective function that includes overlap loss and overlap constraint, iteratively updates the latent space encoding and Lagrange multipliers, automatically eliminating facility overlap and meeting the minimum spacing requirement. This avoids the tedious process of manually adjusting the layout repeatedly in traditional methods and ensures that the generated scheme meets construction safety specifications. (5) The optimized layout data is modeled and registered with the initial site model on the BIM platform. The spatial positioning and size matching are completed by calling the three-dimensional component model library according to the facility coordinates and boundary boxes, and a complete three-dimensional model containing the site environment and facility layout is generated. This realizes the automatic conversion from two-dimensional layout scheme to three-dimensional visualization model, which makes it easier for construction personnel to intuitively understand the site layout scheme. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the genetic algorithm of the present invention; Figure 3 This is a schematic diagram of the architecture of the adversarial generative network of the present invention; Figure 4 This is a diagram of the shape encoder network structure of the present invention; Figure 5 This is a diagram of the generator network structure of the present invention; Figure 6 This is a diagram of the discriminator network structure of the present invention; Figure 7 This is a flowchart of the optimizer iteration process of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1As shown, this invention provides a method for prefabricated building construction site layout based on adversarial generative networks, including: S1, acquiring multi-view images of the construction site and performing 3D reconstruction to obtain a point cloud mesh model, extracting geometric parameters and spatial coordinates to construct an initial site model; S2, collecting historical construction site layout drawings, simplifying facilities by rectangularization and labeling center coordinates, bounding boxes, and categories to form an initial dataset, using a genetic algorithm based on a multi-objective function to iteratively optimize the initial dataset to generate a training dataset, and training the adversarial generative network based on the training dataset; S3, obtaining the site constraints and the bounding boxes, coordinates, and categories of fixed facilities for the current project. The binary mask is converted and input into the trained Generative Adversarial Network (GAN). The GAN uses a shape encoder to encode and decode the binary mask to obtain coded features containing constraints and fixed facility information. The generator generates candidate layout data for temporary facilities based on the coded features and sampling vectors. The discriminator judges and standardizes the complete layout containing fixed and temporary facilities and provides feedback for adjustment. The optimizer iteratively eliminates facility overlaps in the latent space and meets the spacing requirements, outputting optimized layout data. S4: Extract the coordinates, bounding boxes, and categories of each facility in the optimized layout data, and perform modeling and registration based on the initial site model to generate a construction site layout model.

[0020] In one embodiment of the present invention, step S1 includes: Multi-view images of the construction site are collected around the perimeter. Adjacent images have continuity and overlapping areas. During the collection, drones or handheld cameras are used to take pictures around the construction site at different angles and heights to ensure that the images cover the entire boundary and topographic features of the construction site. The overlap rate between adjacent images is not less than 60% to ensure the accuracy of subsequent feature point matching. The image is matched to extract common feature points. Feature detection algorithms such as SIFT or ORB are used to extract feature points in each image. The feature descriptors are matched to find the corresponding common feature points in different images. The RANSAC algorithm is used to remove mismatched points. The relative pose of the camera is calculated based on the matched feature points. Dense point cloud data is generated based on common feature points to obtain 3D point cloud data. The multi-view stereo vision MVS algorithm or motion recovery structure SfM algorithm is used to calculate the 3D spatial coordinates of each pixel in the scene according to the camera pose and the correspondence between feature points, generating dense point cloud data containing hundreds of thousands to millions of points. Each point contains 3D coordinates (x, y, z) and RGB color information. The 3D point cloud data is meshed to connect discrete point clouds into continuous triangular mesh surfaces. Poisson surface reconstruction or Delaunay triangulation algorithm is used to construct the triangular mesh topology based on the spatial distribution and normal vector information of the point cloud, thus connecting discrete point clouds into continuous mesh surfaces. By mapping the image texture onto the grid surface, a point cloud grid model containing geometric parameters and texture information is obtained. Based on the spatial position of the grid surface and the relationship between the camera pose and the image texture, the texture information in the original image is projected onto the corresponding grid surface to generate a 3D model with realistic visual effects. The geometric parameters of the model are extracted, including the coordinates of the site boundary contour line, terrain elevation data, slope information, etc. The point cloud grid model is exported as IFC format and imported into the unified coordinate system of the BIM platform to form an initial site model containing site geometry, spatial position and texture information as the spatial base for subsequent construction site layout. This implementation method obtains the real spatial information of the construction site through multi-view image 3D reconstruction technology. Compared with traditional manual measurement and CAD drawing methods, it improves the accuracy and efficiency of site information acquisition and provides accurate spatial constraints for construction site layout.

[0021] In one embodiment of the present invention, step S2 includes: Collect historical construction site layout drawings. Collect CAD drawings of prefabricated building construction site layouts from CAD drawing cloud libraries, project archives and architectural design websites. Filter drawings that include prefabricated building features such as storage areas for precast concrete components and tower cranes. Delete traditional cast-in-place building construction site layout drawings that lack precast concrete components to ensure that the collected drawings are representative of prefabricated building construction. Establish a coordinate system by selecting the boundary of the construction site. The origin of the coordinate system is located at one corner of the boundary of the construction site and covers the entire construction site. The X-axis and Y-axis are parallel to the main boundary direction of the site. Establish the mapping relationship between the coordinate system and the pixels of the drawing based on the scale and pixel coordinates of the CAD drawing. Each facility is simplified into a rectangular region with its edges parallel to the coordinate axes. The i-th facility is defined by its central coordinates. ,width and height Description, in which This indicates the horizontal length of the facility's boundary frame parallel to the X-axis. The vertical length of the facility's bounding box parallel to the Y-axis is represented. An initial dataset is formed by labeling the center coordinates, bounding box, and category label of each facility. Facility categories include fixed facilities and temporary facilities. Fixed facilities include buildings and construction site gates, while temporary facilities include tower cranes, construction roads, steel processing sheds and storage yards, temporary material storage yards, precast component storage areas, offices, dormitories, large machinery parking areas, tool and equipment sheds, and power distribution rooms, totaling 10 categories. LabelMe annotation software is used to manually annotate CAD drawings. The annotation data for each facility includes center coordinates, bounding box width and height, and category label. The annotation data is stored in JSON format. The initial dataset contains annotation data from 200 to 500 historical construction site layout drawings. like Figure 2 As shown, a genetic algorithm is used to iteratively optimize the initial dataset based on a multi-objective function to generate a training dataset. The specific process is as follows: Each layout scheme in the initial dataset is encoded as a chromosome. A chromosome consists of N gene segments, and each gene segment corresponds to a facility, containing the center coordinates of that facility. , bounding box width and height And category labels, forming an initial population with a population size of M, which is 100 in this embodiment; Calculate the multi-objective function for each layout scheme in the initial population. ,in For the safety objective function, the minimum spacing between all facility pairs is calculated. The formula is: The goal is to maximize the spacing between facilities to improve safety; The objective function for noise pollution is obtained by calculating the weighted distance from the noise source facility to the sensitive area, as shown in the formula: ,in The noise source facilities include tower cranes and steel processing sheds. The sensitive area includes offices and dormitories. Let be the noise intensity coefficient of facility i, and the goal is to minimize noise pollution; The objective function for crane construction efficiency is obtained by calculating the average distance from the crane to the precast component storage area and the building under construction, as shown in the formula. Where C represents the tower crane assembly, Let i be the distance from the tower crane to the precast component storage area. Let i be the distance from the tower crane to the building under construction. The goal is to minimize the transportation distance to improve construction efficiency. The objective function for the impact on surrounding roads is obtained by calculating the minimum distance from the construction facilities to the road at the site boundary, as shown in the formula. ,in Let i be the distance from the site boundary road, and the goal is to maximize this distance to minimize the impact on the surrounding roads; Based on Pareto dominance, non-dominated ranking of placement schemes in the population is performed. If placement scheme A is not inferior to placement scheme B on all objective functions and is superior to placement scheme B on at least one objective function, then A dominates B. For the security objective function... and the objective function of the impact on surrounding roads The goal is to maximize, therefore A being superior to B in these objectives means and For the noise pollution objective function Objective function for crane construction efficiency The goal is to minimize, therefore A being superior to B in these objectives means and All non-dominated individuals in the population are divided into the first non-dominated level. Non-dominated individuals are then identified from the remaining individuals and divided into the second non-dominated level, and so on, until all individuals are assigned to a certain level. Parent individuals are selected based on non-dominated hierarchy and crowding distance. Individuals with higher non-dominated hierarchies are more likely to be selected. Within the same hierarchy, individuals with larger crowding distances are more likely to be selected. Crowding distance is calculated by taking into account the distances between an individual and its neighbors in the target space, using the following formula: ,in and Let be the target values ​​of individual i and its neighboring individuals on target k. and These are the maximum and minimum values ​​of the target, respectively. Perform a crossover operation on the selected parent individuals, and randomly select two parent individuals. and K facilities are randomly selected as intersection points, where K represents 30% to 50% of the total number of facilities. The center coordinates of corresponding facilities in two parent individuals are swapped to generate two child individuals. Specifically... inherit The coordinates of the first K facilities and The coordinates of the last NK facilities, inherit The coordinates of the first K facilities and The coordinates of the last NK facilities are set to have an intersection probability of 0.8. Mutation is performed on offspring individuals. Facilities are randomly selected from the offspring individuals according to a set mutation probability. The center coordinates of the selected facilities are then assigned a Gaussian distribution. Random disturbances, disturbance amplitude The variation should not exceed 10% of the site dimensions, and the resulting coordinates are: and ,in and The perturbation is Gaussian random, and the mutation probability is set to 0.1. Perform boundary constraint checks on offspring individuals; if the facility center coordinates are plus or minus half the bounding box width... and half-high If the facility extends beyond the site boundary, the center coordinates will be adjusted until the facility is completely within the site boundary. The adjustment rules are as follows: , ,in and These represent the lengths of the site along the X and Y axes, respectively. Add offspring individuals to the population, recalculate the multi-objective function values ​​of all individuals and perform non-dominated sorting, retain the M individuals with higher non-dominated levels as the new population; Determine whether the maximum number of iterations has been reached or whether the Pareto front has not changed significantly for several consecutive generations. If the conditions are met, the iteration is terminated; otherwise, the iteration continues. The maximum number of iterations is set to 500 generations. If the change rate of the Pareto front is less than 1% for 50 consecutive generations, it is considered that there has been no significant change. All layout schemes with the first non-dominated level are extracted from the final population as the Pareto optimal solution set. The Pareto optimal solution set is added to the initial dataset to expand it into a training dataset, which contains 1,000 to 2,000 layout schemes. Through multi-objective optimization using genetic algorithms, the layout schemes in the training dataset achieve Pareto optimality in terms of safety, noise pollution control, construction efficiency, and impact on surrounding roads. This overcomes the problem of inconsistent historical data quality and provides high-quality training samples for generative adversarial networks.

[0022] like Figure 3 As shown, in one embodiment of the present invention, the process of training an adversarial generative network based on a training dataset in step S2 includes: The layout schemes in the training dataset are divided into fixed facilities and temporary facilities using a conditional generation strategy. The fixed facilities include buildings and construction site gates, while the temporary facilities include tower cranes, construction roads, steel processing sheds and storage yards, turnover material storage yards, prefabricated component storage areas, offices, dormitories, large machinery parking areas, tool and equipment sheds, and power distribution rooms. The boundary boxes, coordinates, and categories of site constraints and fixed facilities are converted into binary mask representations. On a two-dimensional grid plane, grid cells within the construction site boundary are marked as 1, and grid cells outside the boundary are marked as 0 to form a boundary constraint mask. Grid cells occupied by fixed facilities are marked with corresponding category codes to form a fixed facility mask. Category codes are assigned to different categories of fixed facilities starting from 2 and incrementing. The boundary constraint mask and the fixed facility mask are merged to form a binary mask as the generation condition. The grid resolution is set to 1% of the site size to ensure sufficient spatial accuracy. like Figure 4 As shown, the shape encoder consists of an encoder and a decoder based on a convolutional neural network. The encoder uses 5 convolutional layers for downsampling, with each convolutional kernel having a size of [size missing]. With a step size of 2, ReLU is used as the activation function. The binary mask is downsampled and the features are compressed to obtain a low-dimensional encoding vector. The typical dimension of the encoding vector is 128 to 256. The decoder uses 5 transposed convolutional layers for upsampling, with each convolutional kernel having a size of [size missing]. The step size is 2, the activation function is ReLU, and the last layer uses the Sigmoid activation function to output the reconstructed mask. The value of each grid cell of the reconstructed mask is a probability value between 0 and 1. The decoder upsamples the encoded vector and restores the features to reconstruct the output that matches the initial binary mask. The shape encoder is trained by minimizing the reconstruction error, which is then lost using a binary cross-entropy loss function. Where M is the input binary mask, The binary mask for reconstruction is defined by H and W, which are the height and width of the mask, respectively. The Adam optimizer is used with a learning rate of 0.0001, and the training is performed for 100 epochs until the reconstruction error converges. like Figure 5 As shown, the generator consists of a multilayer perceptron and a conditional Transformer block. The multilayer perceptron contains three fully connected layers with 256, 512, and 1024 neurons in each layer, respectively. The activation function is LeakyReLU. A random latent vector Z is sampled from a standard normal distribution N(0,1). The dimension of the random latent vector corresponds to the number of temporary facilities. In this embodiment, the number of temporary facilities is 10, therefore the dimension of Z is... ,in For each facility, the potential vector dimension is set to 64. Z is input into a multilayer perceptron and subjected to nonlinear transformation to obtain a high-dimensional feature representation. The Conditional Transformer block contains 6 Transformer layers, each including a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism has 8 heads, and the feedforward neural network has a hidden layer dimension of 2048. The high-dimensional feature representation is concatenated with the encoded features output by the shape encoder and then input into the Conditional Transformer block. The encoded features are used as conditional input through a cross-attention mechanism. The Transformer block learns the spatial relationships between temporary structures through the self-attention mechanism, and the output is decoded by a fully connected layer to generate the center coordinates of the temporary structures. , bounding box width and height And the logits of category labels, where k is the temporary facility index. The Softmax function is applied to the logits of category labels to obtain the category probability distribution, and the category labels are sampled. The generated temporary facility data is integrated with the fixed facility data to form candidate layout data. like Figure 6As shown, the discriminator consists of a classifier and an auxiliary self-supervised decoder. The classifier determines whether the input layout data is a real sample from the training dataset or a generated sample from the generator. The auxiliary self-supervised decoder constrains the discriminator to learn construction specifications by reconstructing the bounding boxes, coordinates, and class labels of facilities in the layout data. The classifier encodes the bounding boxes, coordinates, and class labels of facilities in the layout data into attribute vectors and the site constraints into constraint vectors. The classifier uses an 8-layer conditional Transformer block, each layer containing a multi-head self-attention mechanism and a feedforward neural network. The number of attention heads is 8, and the hidden layer dimension is 2048. The attribute vector of each facility includes the center coordinates. , bounding box width and height And one-hot encoding of category labels, mapping attribute vectors to linear layers. 3D embedding vector, The dimension is set to 1024, and a classification label cls token with the same dimension of 1024 is added to aggregate global layout information. The embedding vectors of all facilities and the cls token are input into the conditional Transformer block. The conditional Transformer block extracts the distribution features of the candidate layout data through a self-attention mechanism. The output of the cls token is passed through a fully connected layer to output the probability value of the candidate layout data conforming to the distribution of the training dataset. This probability value is used as the output of the classifier; The auxiliary self-supervised decoder works in parallel with the classifier to reconstruct the bounding boxes, coordinates, and class labels of facilities in the candidate layout data. The decoder uses three fully connected layers with 512, 256, and 128 neurons per layer, respectively. The input is the output features of the penultimate layer of the conditional Transformer block, and the output is the reconstructed coordinates of each facility. Reconstruct the bounding box In addition, the logits of reconstructed category labels, and the reconstruction error including the mean square error of coordinates and bounding boxes. and cross-entropy loss of category labels Where N is the number of facilities and C is the number of categories. For the true label of facility category c, Let be the reconstructed class probability, and the total reconstruction error of the auxiliary self-supervised decoder be . ; The layout schemes in the training dataset are used as real samples, and the candidate layout data are used as generated samples. The discriminator's loss function includes the classification loss of real samples and generated samples, as well as the reconstruction error. The classification loss adopts the binary cross-entropy loss. ,in For real samples, The candidate layout data generated by the generator based on the latent vector z and the condition c. This represents the true probability value output by the discriminator, and the total loss function of the discriminator is: ,in The auxiliary loss weighting coefficient is set to 0.5; The generator's loss function is to maximize the discriminator's true probability value for the generated sample output, as shown in the formula: Adversarial training is performed by alternately optimizing the loss functions of the generator and discriminator. In each iteration, the generator parameters are fixed and the discriminator parameters are updated first, and then the discriminator parameters are fixed and the generator parameters are updated. The Adam optimizer is used, the generator learning rate is set to 0.0002, the discriminator learning rate is set to 0.0001, the batch size is 32, and the training lasts for 300 epochs. The generator continuously improves the generation quality to make the candidate layout data closer to the distribution characteristics of the training dataset, while the discriminator continuously improves its discrimination ability to distinguish between real samples and generated samples. During the training process, the classification accuracy of the discriminator on the validation set is monitored. When the classification accuracy of the discriminator for both real samples and generated samples is close to 50%, it indicates that the discriminator cannot effectively distinguish between real samples and generated samples. At this point, the adversarial training reaches Nash equilibrium, and the training is completed to obtain the trained adversarial generative network. This training process employs a conditional generation strategy, using fixed facilities as generation conditions. The generator only needs to generate the layout of temporary facilities, reducing the complexity of the generation task. The auxiliary self-supervised decoder, by reconstructing the bounding boxes and category labels of the facilities, forces the discriminator to learn the reasonable size and category distribution of facilities in the construction specifications, thereby improving the standardization of the generated layout.

[0023] In one embodiment of the present invention, step S3 includes: Obtain the site constraints and the bounding boxes, coordinates, and categories of fixed facilities for the current project. Site constraints include boundary constraints, spacing constraints, and crane coverage constraints. Boundary constraints are described by the coordinates of the polygon vertices of the construction site boundary, requiring that the bounding boxes of all facilities must not exceed the construction site boundary. The spacing constraint requires that the minimum distance between facilities be no less than 5m. This constraint is determined based on the facility spacing requirements in the construction safety code. The crane coverage constraint requires that the tower crane's operating radius completely cover the area of ​​the building under construction. The tower crane's operating radius is determined according to the tower crane model. Fixed facilities include building and construction site gates; the boundary of a fixed facility is defined by its central coordinates. ,width ,high And category label description, where j is the fixed facility index, the location and size of the building are determined by the architectural design drawings, and the location of the site gate is determined by the site entrance and exit location; The site constraints and the bounding boxes, coordinates, and categories of fixed facilities in the current project are converted into binary mask representations. On the two-dimensional grid plane, the grid cells within the construction site boundary are marked as 1 and the grid cells outside the boundary are marked as 0 to form a boundary constraint mask. The grid cells occupied by the fixed facilities are marked with the corresponding category codes to form a fixed facility mask. The category code of the building is 2 and the category code of the construction site gate is 3. The above masks are combined and input into the trained shape encoder. The encoder part of the shape encoder uses a convolutional neural network to downsample and compress the binary mask. It gradually reduces the spatial resolution and increases the channel dimension through convolutional layers to obtain a low-dimensional encoded vector containing the current project constraints and fixed facility information as the encoded features. It should be noted that the decoder part of the shape encoder is not used in application. Random latent vectors are sampled from a standard normal distribution N(0,1). The dimension of the random latent vectors corresponds to the number of temporary facilities. These temporary facilities include 10 categories: tower cranes, construction roads, steel processing sheds and storage yards, temporary material storage yards, precast component storage areas, offices, dormitories, large machinery parking areas, tool and equipment sheds, and power distribution rooms. Therefore, the dimension of the random latent vectors is... ; A high-dimensional feature representation is obtained by inputting a random latent vector into a multilayer perceptron and performing a nonlinear transformation. The high-dimensional feature representation and the encoded features are input together into the conditional Transformer block. The conditional Transformer block uses the encoded features as conditional input and fuses them with the high-dimensional features through a cross-attention mechanism. In the cross-attention mechanism, the query vector comes from the high-dimensional feature representation, and the key vector and value vector come from the encoded features. Attention weights are obtained by calculating the dot product of the query vector and the key vector. The attention weights are then weighted and summed with the value vectors to obtain the fused feature representation. The conditional Transformer block learns the spatial relationships between temporary facilities through a self-attention mechanism. In the self-attention mechanism, the query vector, key vector, and value vector all come from the fused feature representation. Attention weights are obtained by calculating the dot product of the query vector and the key vector. The attention weights reflect the strength of the association between different temporary facilities. The output of the conditional Transformer block is decoded by a fully connected layer to generate the center coordinates of the temporary facility. , bounding box width and height And category labels, where k is the temporary facility index; the fully connected layer contains 2 layers, the first layer has 256 neurons and the activation function is ReLU, the second layer outputs a 5-dimensional vector for each temporary facility, the first 4 dimensions are the center coordinates and bounding box size, and the 5th dimension is the probability distribution of the category label; The generated temporary facility data is integrated with the fixed facility data to form candidate layout data, which contains complete information on 12 facilities. The candidate layout data is input into the trained discriminator; the discriminator's classifier encodes the bounding boxes, coordinates, and class labels of facilities in the candidate layout data into attribute vectors, each facility's attribute vector containing... One-hot encoding of categories encodes site constraints into constraint vectors, which contain information such as normalized coordinates of site boundaries, minimum spacing requirements, and crane operating radius. Input the attribute vector and constraint vector along with the classification token (cls token) into the conditional Transformer block. The classification token is used to aggregate global layout information. The Conditional Transformer block extracts the distribution features of candidate layout data through a self-attention mechanism, calculates the spatial relationship between each facility and other facilities, and the similarity between the overall layout and the distribution of the training dataset. The output of the classification label is passed through a fully connected layer to output the probability value of the candidate layout data conforming to the distribution of the training dataset as the judgment result. The probability value ranges from [0,1]. The higher the probability value, the more the candidate layout data conforms to the distribution characteristics of the training dataset. The auxiliary self-supervised decoder reconstructs the bounding boxes, coordinates, and category labels of facilities in the candidate layout data and calculates the reconstruction error as the standard verification result. The reconstruction error includes the mean square error of the coordinates and bounding boxes and the cross-entropy loss of the category labels. The smaller the reconstruction error, the more the candidate layout data conforms to the reasonable size and category distribution of facilities in the construction specifications. The probability threshold is set to 0.8, and the reconstruction error threshold is determined based on the training set statistics as the mean reconstruction error of the training set plus one standard deviation. When the probability value is lower than the set threshold or the reconstruction error is higher than the set threshold, the decision fails. The generator adjusts the random latent vector to regenerate candidate layout data and inputs it into the discriminator for judgment again. The adjustment method is to add a small random perturbation to the random latent vector or resample some dimensions. Iteration continues until the probability value is higher than the set threshold and the reconstruction error is lower than the set threshold to complete the judgment and specification verification. The maximum number of iterations is set to 20, and the candidate layout data that passes the verification is output. The augmented Lagrangian method is used to iteratively eliminate facility overlaps in the latent space while satisfying the spacing requirements. The 5m minimum spacing constraint is transformed into a facility boundary box non-overlap constraint, and an overlap loss function is defined. , where Z is the latent space code, L is the set of facility category labels, D represents the probability value of the trained discriminator that the generated layout conforms to the distribution of the training dataset, and G represents the layout data generated by the trained generator based on the latent space code Z and the category labels L. Define overlapping constraint functions Where O is the sum of the overlapping areas between all pairs of facilities in the generated layout, calculated by the overlap judgment function. For two facilities i and j, their overlapping area is calculated as follows: The total overlapping area is ; Constructing the augmented Lagrangian objective function ,in For Lagrange multipliers, For penalty parameters; like Figure 7 As shown, the iterative optimization process begins with initialization. and Fixed in the nth iteration and Solving the objective function for the latent space encoding Z The gradient of Z is calculated using automatic differentiation. The latent space encoding that minimizes the objective function is obtained by updating Z using gradient descent. The update step size is set to 0.01, and the gradient descent is iterated for 50 steps or until the gradient norm is less than 0.001; Update Lagrange multipliers ; Update penalty parameters in As a penalty factor; Calculate the constraint function value and tolerance threshold Compare, Set to 0.01m 2 ,like Then the constraint conditions are met, and The input generator obtains optimized layout data that eliminates overlap and meets the 5m spacing requirement; otherwise, it continues to iterate until the constraint is met or the maximum number of iterations of 30 is reached, and then outputs optimized layout data. In this embodiment, the discriminator's judgment and standard verification ensure that the generated layout scheme conforms to the distribution characteristics of the training dataset and the construction specifications. The optimizer iteratively eliminates facility overlap in the latent space, avoiding the problem that directly adjusting the facility position in the layout space may destroy the discriminator's verification results. The latent space optimization maintains the data distribution characteristics and standardization of the layout, while satisfying the no-overlap constraint.

[0024] In one embodiment of the present invention, step S4 includes: Extract the center coordinates, bounding boxes, and category labels of each facility from the optimized layout data. The optimized layout data includes the center coordinates of all facilities. , bounding box width and height and category labels; Import the initial site model into the BIM platform as the base model. The initial site model is converted from the point cloud mesh model generated in step S1. It includes the geometry, boundary contour, terrain elevation and texture information of the construction site. Establish a unified coordinate system in the BIM platform. The origin of the coordinate system is consistent with the origin of the coordinate system of the optimized layout data. The X-axis and Y-axis directions are consistent, and the Z-axis is the vertical upward direction. The corresponding 3D model is retrieved from the 3D component model library based on the facility category label. The 3D component model library contains standard 3D models of facilities such as tower cranes, construction roads, steel processing sheds, turnover material storage yards, prefabricated component storage areas, offices, dormitories, large machinery parking areas, tool and equipment sheds, power distribution rooms, buildings, and construction site gates. Each 3D model contains geometric shape, size parameters, material information, and component attributes. The spatial position of the 3D model is determined in the unified coordinate system of the initial site model based on the center coordinates of the facility. Mapped to three-dimensional coordinates in the BIM platform coordinate system ,in The height of the facility's base above the ground is determined based on the topographic elevation data of the initial site model. (For ground-level facilities...) This refers to the ground elevation at that location, for tall structures such as tower cranes. The elevation of the bottom surface of the tower crane foundation; The 3D model is scaled to match the bounding box dimensions based on the width and height of the facility. The original dimensions of the 3D model are denoted as follows: ,in For model width, For model length, The model is scaled along the X and Y axes to its height, with scaling ratios of [missing values]. and The Z-axis height remains the original height or is scaled according to the facility height parameters; The scaled 3D model is placed in a defined spatial location to complete the 3D modeling of the facility. For linear facilities such as construction roads, the length and width of the road are determined based on the width and height of its bounding box, and a 3D model of the road is generated and placed in the corresponding position. The 3D models of all facilities are fused with the initial site model in a unified coordinate system to generate a complete construction site layout model that includes the site environment and facility layout. The fusion process includes checking the spatial relationships between models, adjusting the relative positions of models to ensure no overlap or suspension, and adding materials and textures to the models to improve the visualization effect. The generated construction site layout model can be 3D walked around, simulated during construction, and collided during collision detection on the BIM platform. It can be exported to IFC format for data exchange with other BIM software, providing 3D visualization support for subsequent construction management and progress control. This implementation method converts optimized layout data into a 3D BIM model through modeling and registration, realizing the automatic conversion from a 2D layout scheme to a 3D visualization model. This facilitates construction personnel's intuitive understanding of the site layout scheme and improves the visualization level and practicality of the construction site layout.

[0025] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for arranging construction site for prefabricated buildings based on generative adversarial networks, characterized in that, include: S1. Collect multi-view images of the construction site and perform three-dimensional reconstruction to obtain a point cloud mesh model. Extract geometric parameters and spatial coordinates to construct an initial site model. S2. Collect historical construction site layout drawings, simplify the facilities by rectangularization, and mark the center coordinates, bounding boxes and categories to form an initial dataset. Use a genetic algorithm based on a multi-objective function to iteratively optimize the initial dataset to generate a training dataset. Train an adversarial generative network based on the training dataset. S3. Obtain the site constraints and bounding boxes, coordinates, and categories of fixed facilities for the current project, convert them into binary mask representations, and input them into the trained Generative Adversarial Network (GAN). The GAN uses a shape encoder to encode and decode the binary mask to obtain encoded features containing constraint and fixed facility information. The generator generates candidate layout data for temporary facilities based on the encoded features and sampling vectors. The discriminator judges and standardizes the complete layout containing fixed and temporary facilities and provides feedback for adjustment. The optimizer iteratively eliminates facility overlaps and meets spacing requirements in the latent space, and outputs optimized layout data. S4. Extract the coordinates, bounding boxes, and categories of each facility from the optimized layout data, and perform modeling and registration based on the initial site model to generate a construction site layout model.

2. The prefabricated building construction site layout method based on adversarial generative networks according to claim 1, characterized in that, The process of obtaining the point cloud mesh model in step S1 is as follows: Multi-view images of the construction site were collected around the entire perimeter, with continuity and overlapping areas between adjacent images; Feature point matching is performed on the image to extract common feature points, and dense point cloud is generated based on the common feature points to obtain three-dimensional point cloud data; The discrete point cloud data is connected into a continuous triangular mesh surface by meshing the 3D point cloud data. The image texture is then projected onto the mesh surface through texture mapping to obtain a point cloud mesh model containing geometric parameters and texture information.

3. The prefabricated building construction site layout method based on adversarial generative networks according to claim 1, characterized in that, The method for simplifying the facility into a rectangular shape in step S2 is as follows: Establish a coordinate system by selecting the boundary of the construction site. The origin of the coordinate system is located at one corner of the boundary of the construction site and covers the entire construction site. Each facility is simplified into a rectangular region with its edges parallel to the coordinate axes. The i-th facility is defined by its central coordinates. ,width and height Description, in which This indicates the horizontal length of the facility's boundary frame parallel to the X-axis. The vertical length of the facility's bounding box parallel to the Y-axis is represented by the initial dataset, which is formed by labeling the center coordinates, bounding box, and category label of each facility.

4. The prefabricated building construction site layout method based on adversarial generative networks according to claim 1, characterized in that, The iterative optimization process using a genetic algorithm in step S2 includes: Step 1: Encode each layout scheme in the initial dataset as a chromosome. A chromosome consists of N gene segments, each corresponding to a facility and containing the center coordinates of that facility. , bounding box width and height And category labels, forming an initial population with a population size of M; Step 2: Calculate the multi-objective function for each layout scheme in the initial population. ,in The safety objective function is obtained by calculating the minimum distance between facilities; The objective function for noise pollution is obtained by calculating the weighted distance from the noise source facilities to the sensitive area. The objective function for crane construction efficiency is obtained by calculating the average distance from the crane to the precast component storage area and the building under construction. The objective function for the impact on surrounding roads is obtained by calculating the minimum distance from the construction facilities to the road at the site boundary. Step 3: Based on the Pareto dominance relationship, perform non-dominated ranking of the layout schemes in the population. If layout scheme A is not inferior to layout scheme B on all objective functions and is superior to layout scheme B on at least one objective function, then A dominates B, and the population is divided into multiple non-dominated levels. Step 4: Select parent individuals based on non-dominated hierarchy and crowding distance. Individuals with higher non-dominated hierarchy are more likely to be selected, and individuals with larger crowding distance within the same hierarchy are more likely to be selected. Crowding distance is obtained by calculating the distance between an individual and its neighboring individuals in the target space. Step 5: Perform a crossover operation on the selected parent individuals. Randomly select two parent individuals and K facilities as crossover points. Swap the center coordinates of the corresponding facilities in the two parent individuals to generate two child individuals. Step 6: Perform a mutation operation on the offspring individuals. Randomly select facilities from the offspring individuals according to the set mutation probability, and add a random perturbation that follows a Gaussian distribution to the center coordinates of the selected facilities. The perturbation amplitude shall not exceed 10% of the site size. Step 7: Perform boundary constraint checks on the offspring individuals. If the facility's center coordinates exceed the site boundary after adding or subtracting the half-width and half-height of the boundary frame, adjust the center coordinates until the facility is completely within the site boundary. Step 8: Add the offspring individuals to the population, recalculate the multi-objective function values ​​of all individuals and perform non-dominated sorting, retain the M individuals with higher non-dominated levels as the new population; Step 9: Determine whether the maximum number of iterations has been reached or whether the Pareto front has not changed significantly for several consecutive generations. If so, terminate the iteration; otherwise, return to step 4. Step 10: Extract all layout schemes with the first non-dominated level from the final population as the Pareto optimal solution set, and add the Pareto optimal solution set to the initial dataset to expand and form the training dataset.

5. The prefabricated building construction site layout method based on adversarial generative networks according to claim 1, characterized in that, The site constraints in step S3 include boundary constraints, spacing constraints, and crane coverage constraints. Boundary constraints require that the boundary frames of all facilities must not exceed the construction site boundaries. Spacing constraints require that the minimum distance between facilities be no less than 5m. Crane coverage constraints require that the operating radius of the tower crane completely covers the area of ​​the building under construction. Fixed facilities include building and construction site gates; the boundary of a fixed facility is defined by its central coordinates. ,width ,high And category label description, where j is the fixed facility index.

6. The prefabricated building construction site layout method based on adversarial generative networks according to claim 1, characterized in that, The training method for the generative adversarial network in step S2 is as follows: A conditional generation strategy is used to divide the layout schemes in the training dataset into fixed facility parts and temporary facility parts; The site constraints and fixed facilities are input into the generator as generation conditions. The generator generates the bounding box, coordinates and category labels of the temporary facilities based on the generation conditions and integrates them with the fixed facilities to form candidate layout data. The layout schemes in the training dataset are used as real samples, and the candidate layout data are used as generated samples. The discriminator uses a classifier to determine whether the input layout data is a real sample or a generated sample and outputs the real probability value. At the same time, an auxiliary self-supervised decoder is used to reconstruct the bounding boxes, coordinates and category labels of the facilities in the layout data. The discriminator's loss function includes the classification loss of real samples and generated samples, as well as the reconstruction error. The generator's loss function is to maximize the true probability value output by the discriminator for the generated samples. Adversarial training is performed by alternately optimizing the loss functions of the generator and discriminator. The generator continuously improves the generation quality so that the candidate layout data is closer to the distribution characteristics of the training dataset. The discriminator continuously improves its discrimination ability to distinguish between real samples and generated samples. When the discriminator can no longer effectively distinguish between real samples and generated samples, the training is completed, and the trained adversarial generative network is obtained.

7. The method for arranging prefabricated building construction sites based on adversarial generative networks according to claim 1, characterized in that, The process of generating candidate layout data for temporary facilities in step S3 is as follows: Convert the site constraints and fixed facility bounding boxes, coordinates, and categories of the current project into binary mask representations and input them into the trained shape encoder; The shape encoder uses a convolutional neural network to downsample and compress the binary mask to obtain a low-dimensional encoding vector containing current project constraints and fixed facility information as the encoding feature. A random latent vector is obtained by sampling from the standard normal distribution N(0,1), and the dimension of the random latent vector corresponds to the number of temporary facilities; A high-dimensional feature representation is obtained by inputting a random latent vector into a multilayer perceptron and performing a nonlinear transformation. The high-dimensional feature representation and the encoded features are fed into the conditional Transformer block, which uses the encoded features as conditional input to learn the spatial relationships between temporary facilities through a self-attention mechanism. The output of the conditional Transformer block is decoded by a fully connected layer to generate the center coordinates of the temporary facility. , bounding box width and height And category labels, where k is the temporary facility index; The generated temporary facility data is integrated with the fixed facility data to form candidate layout data.

8. The method for arranging prefabricated building construction sites based on adversarial generative networks according to claim 1, characterized in that, The method by which the discriminator judges and standardizes the candidate layout data and provides feedback for adjustment in step S3 is as follows: Input the candidate layout data into the trained discriminator; The classifier of the discriminator encodes the bounding boxes, coordinates and category labels of facilities in the candidate layout data into attribute vectors, and encodes site constraints into constraint vectors; Input the attribute vector and constraint vector along with the category tag into the conditional Transformer block. The category tag is used to aggregate global layout information. The Conditional Transformer block extracts the distribution features of candidate layout data through a self-attention mechanism. The output of the classification label is passed through a fully connected layer to output the probability value of the candidate layout data conforming to the distribution of the training dataset as the judgment result. The auxiliary self-supervised decoder reconstructs the bounding boxes, coordinates and category labels of facilities in the candidate layout data and calculates the reconstruction error as the standard verification result. If the probability value is lower than the set threshold or the reconstruction error is higher than the set threshold, the decision fails. The generator adjusts the random latent vector to regenerate candidate layout data and inputs it into the discriminator for decision again. The process iterates until the probability value is higher than the set threshold and the reconstruction error is lower than the set threshold to complete the decision and specification verification. The candidate layout data that has passed the verification is then output.

9. A method for arranging prefabricated building construction sites based on generative adversarial networks according to claim 5, characterized in that, In step S3, the optimizer iteratively eliminates facility overlaps and satisfies spacing requirements in the latent space as follows: The augmented Lagrange method is used to transform the 5m minimum spacing constraint into a non-overlapping constraint for the facility boundary box; Define the overlap loss function , where Z is the latent space code, L is the set of facility category labels, D represents the probability value of the trained discriminator that the generated layout conforms to the distribution of the training dataset, and G represents the layout data generated by the trained generator based on the latent space code Z and the category labels L. Define overlapping constraint functions , where O is the overlap judgment function that calculates the sum of the overlapping areas between all facility pairs in the generated layout; Construct the augmented Lagrangian objective function: ； in For Lagrange multipliers, For penalty parameters; By iteratively updating the latent space encoding Z and the Lagrange multipliers and penalty parameters The optimal latent space encoding that minimizes the objective function and satisfies the tolerance threshold requirement of the overlap constraint function is searched in the latent space. The optimal latent space encoding is input into the generator to obtain optimized layout data that eliminates overlap and meets the 5m spacing requirement.

10. A method for arranging prefabricated building construction sites based on adversarial generative networks according to claim 1, characterized in that, The modeling and registration method in step S4 is as follows: Extract the center coordinates, bounding boxes, and category labels of each facility from the optimized layout data. Import the initial site model into the BIM platform as the base model. Call the corresponding 3D model from the 3D component model library according to the facility category label. Determine the spatial position of the 3D model in the unified coordinate system of the initial site model according to the center coordinates of the facility. Scale the 3D model according to the width and height of the facility's bounding box to match the bounding box size. Place the scaled 3D model in the determined spatial position to complete the 3D modeling of the facility. Merge the 3D models of all facilities with the initial site model in the unified coordinate system to generate a complete construction site layout model that includes the site environment and facility layout.

Images