BIM model intelligent creation method and system based on AI algorithm

Abstract

The invention discloses a BIM model intelligent creation method and system based on an AI algorithm, and relates to the technical field of building information models.The method comprises the steps that RGB images are generated on the basis of component pair coordinates, Shapley is used for detecting the RGB images to generate mark conflicts and missing labels, RGB image feature vectors are extracted on the basis of a convolutional neural network model, image features serve as node features, and a BIM model is created; node features are updated through a graph neural network model, conflict and missing probabilities are predicted by using a feedforward neural network, a graph structure is constructed based on image feature vectors, and an environment function is calculated according to the graph structure and the predicted conflict and missing probabilities. A convolutional neural network and a graph neural network are fused, efficient recognition of component conflicts and missing is achieved, the model self-adaption and generalization ability is improved, probability prediction is conducted in combination with a feedforward neural network, the complex space and function relation between components is dynamically captured, and the intelligence of the BIM modeling process is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of building information modeling, and in particular to a method and system for intelligently creating a BIM model based on an AI algorithm. Background Art

[0002] Building Information Modeling (BIM) technology, as a core tool for the digital transformation of the construction industry, has been widely used in the design, construction, and operation and maintenance stages of buildings. However, existing technologies still have significant deficiencies in terms of automation, component feature modeling granularity, and intelligent prediction capabilities. Existing BIM component clash detection methods mainly rely on geometric Boolean operations or rule-based semantic inspection tools. For example, the Clash Detection module is widely used in software such as Revit and Navisworks. However, such methods have the following shortcomings: First, they cannot effectively handle the analysis of high-dimensional spatial overlap and missingness under large-scale component relationships, resulting in slow processing speed and accuracy that is easily affected by modeling errors. Second, they lack end-to-end learning and generalization capabilities and cannot adapt to dynamically changing layout plans. Third, their ability to automatically understand and discriminate spatial functional constraints is weak, making it difficult to automatically identify areas with incomplete functions according to architectural design specifications. Fourth, they lack an intelligent prediction mechanism integrated with the graph structure and cannot understand structural conflicts in the overall layout at the component level. To address the above problems, the use of deep learning methods based on RGB image features can improve the ability to extract geometric features, the introduction of graph neural networks can realize the structured expression of complex component relationships, and the combination of reinforcement learning strategies can make global optimal adjustments to component positions and functional configurations, making up for the shortcomings of existing methods in terms of intelligence and automation. Existing BIM model construction methods generally have problems such as low automation, inaccurate conflict and missing identification, and difficulty in intelligent optimization. We have invented a BIM model intelligent creation method based on AI algorithm, which can effectively solve the problems of component space conflict identification, function missing detection and position adjustment optimization, and improve the intelligence level of BIM modeling. This invention belongs to the interdisciplinary field of building information modeling, artificial intelligence and computer vision, and has strong practicality and technology promotion value. Summary of the Invention

[0003] In view of the above existing problems, the present invention is proposed.

[0004] Therefore, the present invention provides a method and system for intelligent creation of BIM models based on AI algorithm to solve the problems of component space conflict identification, function loss detection and position adjustment optimization.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] In a first aspect, the present invention provides an AI-based BIM model intelligent creation method, comprising: parsing CAD drawings based on the Revit API to extract initial component coordinates, generating particle coordinates based on the initial component coordinates, calculating a multi-objective cost for each particle based on the Evaluate Cost function of the AI ​​algorithm, and outputting an optimal position coordinate set based on the multi-objective cost;

[0007] Use the optimal position coordinate set to generate component pair coordinates, generate an RGB image based on the component pair coordinates, use Shapely to detect the RGB image and generate conflict and missing labels, extract the RGB image feature vector based on the convolutional neural network model, use the image features as node features, update the node features through the graph neural network model, use the feedforward neural network to predict conflict and missing probabilities, construct a graph structure based on the image feature vectors, and calculate the environment function based on the graph structure and the predicted conflict and missing probabilities;

[0008] Based on the environment function, the PPO algorithm is used to output the optimized component coordinates and eigenvectors, and the BIM model structure is generated according to the optimized data and written into the IFC file;

[0009] Build a real-time data streaming platform to realize structured storage of BIM data, and perform real-time optimization and automatic update integration of BIM data.

[0010] As a preferred solution of the AI ​​algorithm-based BIM model intelligent creation method described in the present invention, wherein: the extraction of component initial coordinates from CAD drawings based on Revit API, and the generation of particle coordinates based on the initial coordinates, the calculation of multi-objective costs for each particle based on the Evaluate Cost function of the AI ​​algorithm, and the output of the optimal position coordinate set based on the multi-objective costs refer to the use of Revit API to parse CAD drawings, extract component geometric information, material properties and function labels, the extracted component initial position coordinates are the initial component layout in the CAD drawings, and the use of PythoNumPy method to generate a feature vector for each component initial position coordinate. i :O i =[x,y,z,l,w,h,m1,…,m k ,f1,…,f n ,], the initial feature vector set {O1,O2,...,O i}, stored as a NumPy array, and a fixed number of particle coordinates are initialized using the Python DEAP framework for each component coordinate And set the initial velocity for the particles Use the Evaluate Cost function to calculate the multi-objective cost A for each particle and set the multi-objective cost threshold C. If the current cost A is lower than the set threshold C, the particle position is updated through the particle swarm algorithm. Use the clamping method to limit the new position, repeat the cost evaluation and particle update iteration to the set number of times, and output the optimal coordinate position corresponding to each component The optimal position coordinates form a coordinate set

[0011] As a preferred solution of the AI ​​algorithm-based BIM model intelligent creation method described in the present invention, the method comprises the following steps: using the optimal position coordinate set to generate component pair coordinates, generating an RGB image based on the component pair coordinates, detecting the RGB image using Shapely to generate conflict and missing labels, extracting RGB image feature vectors based on a convolutional neural network model, using image features as node features, updating node features through a graph neural network model, predicting conflict and missing probabilities using a feedforward neural network, constructing a graph structure based on the image feature vectors, and calculating an environment function based on the graph structure and the predicted conflict and missing probabilities. The method comprises the following steps: using Python to import the optimal position coordinate set into a Pandas DataFrame, using Pandas indexing and nested loops to generate component pairs (q, j), and generating corresponding RGB images F for each component pair (q, j) based on Python Pillow. qj ;

[0012] Retrieve the coordinates and dimensions for each pair of components (q, j) from the DataFrame using Pandas' index. Use the coordinates and dimensions to represent each component as a 2D rectangular component. Use Shapely's geometric intersection test to determine if the rectangles overlap. If they overlap, mark them with a conflict label {y qj}, using a binary label (1 for overlap, 0 for no overlap) for each artifact pair, filter the DataFrame to select the conflicting labels {y qj}, For each space s, create a 2D polygon representing the boundary using the coordinates and dimensions. The polygon is defined by four vertices ((x s ,y s )、(x s +l s ,y s )、(x s +l s ,y s +w s )、(x s ,y s +w s )) , parse component function labels from the Revit API and extract space design specifications through PandasDataFrame, generate a list of necessary component types for each space s, use Shapely geometry inclusion test method to check whether each space s polygon contains all necessary component types. If any component type required by the specification is missing, mark the space as missing label Us =1, otherwise marked as U s = 0, the missing label {U s} and the conflicting label {y qj} and image {F qj}stored in the labeled dataset;

[0013] Based on the convolutional neural network (CNN) model, the ResNet-18 architecture is used as the backbone network to qj Extract its geometric features, use the graph neural network (GNN) model, take the component feature vector as the node feature, and initialize it as the node feature For each graph neural network layer, node features are updated and conflicts are predicted using a feedforward neural network. and missing probability

[0014] Based on the feature vector set, each building component (such as wall, beam, etc.) is defined as a node constituting a node set V, and a unique identifier is used as the index of the Pandas DataFrame. The Euclidean distance is calculated according to the position of the component in three-dimensional space, and the distance matrix is ​​implemented using NumPy. The edge set E is generated based on the set spatial proximity threshold. NetworkX is used to construct an undirected graph data structure G = (V, E) based on V and E. The undirected graph data structure G is connected to the collision probability and missing probability Encapsulate it into JSON for integration, use Python to export the JSON file, create a Gym environment with action and observation space, and use the state encoding of the graph based on the Gym environment to encode the conflict probability and missing probability Convert to a state vector, encode the state vector through ArrayFlattening, and extract the encoded state vector and Calculate the environment function R based on the undirected graph data structure G and the conflict and missing probabilities t .

[0015] As a preferred solution of the BIM model intelligent creation method based on AI algorithm described in the present invention, wherein: the component coordinates and feature vectors optimized by the PPO algorithm based on the environment function are outputted, which means initializing the Gym environment based on the custom Gym environment registration method, training the strategy network (MLP) in the PPO learning model in the Gym environment, determining the action according to the current state vector, optimizing the dynamic weight calculation, updating the PPO strategy network, and based on the environment function R t , output the optimized component coordinates and component feature vectors.

[0016] As a preferred solution of the AI-based BIM model intelligent creation method described in the present invention, the generating of the BIM model structure according to the optimized data and writing the BIM model structure into the IFC file refers to creating an empty file based on the IFC4 standard, defining the project, site and building in sequence, setting the basic structure of the file, creating an IfcBuildingElement for each component by traversing and optimizing the component coordinates and feature vectors, and creating an IfcSpace entity for each space s, defining the boundary based on the spatial polygon data, using IfcPolyLoop to represent the polygon outline, storing the geometric boundary of the space, creating an IfcRelConnectsElements entity based on the edge set E of the undirected graph data structure G using ifcopenshell, writing the components, spaces and relationships into the IFC file using ifcopenshell, using ifcopenshell.validate to check the syntax and semantic integrity of the file to ensure compliance with the IFC4 standard, using the ifcopenshell.file.write method to save it as an optimized_bim.ifc file, and directly generating a BIM model containing the optimized data.

[0017] As a preferred solution of the AI-based BIM model intelligent creation method described in the present invention, the construction of a real-time data stream platform to achieve structured storage of BIM data refers to using Apache Kafka to deploy a real-time data stream platform, configuring the topic bim data for transmitting BIM data and status. In the system architecture, Kafka acts as a circulation hub for data updates, and continuously receives component position change information from sensors and construction progress monitoring real-time sources to implement an event-driven mechanism for data streams, detect conflicts and missing status tags, and a change from 0 to 1 is considered a potential problem. The component feature optimization method is called to automatically adjust the size and material properties according to the component category, and parameter optimization is performed under the premise of meeting the specification constraints. MongoDB is initialized, and the collections bim models and graph data and the optimization results of component sizes and material properties are created for data storage. The IFC file is parsed through ifcopenshell, and the components, spaces and relationships are extracted, converted into JSON format and stored in the bim_models collection.

[0018] As a preferred solution of the AI-based BIM model intelligent creation method described in the present invention, the real-time optimization and automatic update integration of BIM data refers to using Kafka consumers to subscribe to the bim-data topic, receiving real-time update data, parsing the new data into IFC format through ifcopenshell, updating the component coordinates and spatial data in the MongoDB collection bim models, using the AE-HNN model to infer the current coordinates and the optimized graph, combining NumPy to calculate the Euclidean distance to update the edge set and adjacency matrix of the graph, obtaining the current conflict probability and missing probability, using the pre-trained LSTM model to predict the conflict and missing probability of the next time step based on the current probability and GNN node features, calculating the current reward, and triggering PPO optimization if it is lower than the statistical threshold to obtain the optimized component coordinates and feature vectors, updating the IFC file through ifcopenshell, writing the optimized coordinates into the component position, updating the neighboring relationship, and saving it as updated bim.ifc after verification. In the real-time visualization dashboard built through the Dash framework, the conflict and missing labels are displayed as key diagnostic information to assist users in judging the changing trend of the conflict area, and the updated Import bim.ifc into Autodesk Revit to update the existing BIM model data in Revit.

[0019] In a second aspect, the present invention provides an AI-based BIM model intelligent creation system, including a CAD data parsing and particle swarm optimization module for extracting component information to generate initial component coordinates and feature vectors, and calculating multi-objective costs to output an optimal position coordinate set.

[0020] Feature extraction and graph construction module, used to extract image feature vectors, update node features, and predict conflict and missing probabilities;

[0021] Reinforcement learning optimization module, used for the PPO strategy network to output component coordinates and feature vectors;

[0022] BIM model generation module, used to write optimized coordinates and feature vectors into IFC files;

[0023] Real-time data streaming and dynamic optimization modules are used to transmit BIM data and status, and update BIM models through Revit API integration.

[0024] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the method for intelligently creating a BIM model based on an AI algorithm as described in the first aspect of the present invention is implemented.

[0025] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the method for intelligently creating a BIM model based on an AI algorithm as described in the first aspect of the present invention.

[0026] The beneficial effects of the present invention are: integrating image processing with convolutional neural networks (CNN) and graph neural networks (GNN) to achieve efficient recognition of component conflicts and missing components, and improve the model's adaptability and generalization capabilities. By constructing a graph structure with image features as node features and combining it with a feedforward neural network for probability prediction, the system can dynamically capture the complex spatial and functional relationships between components, introduce a reinforcement learning mechanism, and build a Gym-based reinforcement learning environment to define the action space and observation space. The conflict and missing probabilities are introduced as state vectors. Combined with dynamic adjustment strategies, the reinforcement learning agent is used to adjust and optimize the component position and type. It can iteratively learn the optimal layout plan within multiple time steps, significantly improving the rationality and constructability of the BIM model configuration, making up for the inefficiency of "static verification and manual adjustment" in the existing modeling process, and significantly improving the automation, intelligence and global optimality of the BIM modeling process. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0028] Figure 1 This is a flowchart of the AI-based BIM model intelligent creation method in Example 1.

[0029] Figure 2 This is a structural diagram of the AI-based BIM model intelligent creation system in Example 1.

[0030] Figure 3 This is a flow chart of feature extraction and conflict detection for the AI-based BIM model intelligent creation method in Example 1.

[0031] Figure 4 This is a real-time optimization and update flowchart of the AI ​​algorithm-based BIM model intelligent creation method in Example 1. DETAILED DESCRIPTION

[0032] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0033] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0034] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0035] Example 1, with reference to Figures 1 to 4 , which is the first embodiment of the present invention, provides a method for intelligently creating a BIM model based on an AI algorithm, comprising the following steps:

[0036] S1. Analyze CAD drawings based on the Revit API to extract the initial coordinates of components, generate particle coordinates based on the initial coordinates of components, calculate the multi-objective cost for each particle based on the Evaluate Cost function of the AI ​​algorithm, and output the optimal position coordinate set based on the multi-objective cost;

[0037] Use the optimal position coordinate set to generate component pair coordinates, generate an RGB image based on the component pair coordinates, use Shapely to detect the RGB image and generate conflict and missing labels, extract the RGB image feature vector based on the convolutional neural network model, use the image features as node features, update the node features through the graph neural network model, use the feedforward neural network to predict conflict and missing probabilities, construct a graph structure based on the image feature vectors, and calculate the environment function based on the graph structure and the predicted conflict and missing probabilities;

[0038] Specifically, the initial coordinates of the components are extracted by parsing the CAD drawings based on the Revit API, and the particle coordinates are generated according to the initial coordinates. The Evaluate Cost function based on the AI ​​algorithm calculates the multi-objective cost for each particle, and outputs the optimal position coordinate set according to the multi-objective cost. It refers to using the Revit API to parse the CAD drawings, extracting the component geometry information (the coordinates and dimensions of the physical components in the building information model (BIM), such as walls, pipes, beams), material properties (concrete, steel, etc.) and function labels (load-bearing, non-load-bearing). The extracted initial position coordinates of the components are the initial component layout in the CAD drawings, and the PythoNumPy method is used to generate a feature vector O for the initial position coordinates of each component. i :O i =[x,y,z,l,w,h,m1,…,m k ,f1,…,f n,], where x, y, z are the coordinates of the centroid of the component (the equilibrium point of the component geometry, calculated by the component geometry), l, w, h are the length, width, and height of the component, extracted through the Revit API, m1,…,m k , is the material one-hot encoding (e.g. m1 = 1 for concrete), obtained from the material list through Python script, f1,…,f n The functional attributes (e.g. f1 = 1 for load-bearing) are obtained from the design specification code through Python scripts, and the initial feature vector set {O1, O2, ..., O i}, stored as a NumPy array;

[0039] For each component coordinate, a fixed number of particle coordinates are initialized using the Python DEAP framework. (By applying random perturbations (Perturb method) to the initial coordinates, the initial positions of 100 particles are generated). Particle swarm optimization (PSO) is a two-dimensional space process, using x, y coordinates to represent the position of components in the plane layout for ease of calculation and optimization. The z coordinate and size information are incorporated into the cost evaluation through the eigenvector to simplify the algorithm complexity and set the initial velocity for the particles. Obtained through the NumPy library;

[0040] Use the Evaluate Cost function to calculate the multi-objective cost A for each particle. This calculation includes a comprehensive score for space utilization (evaluating the proportion of the area occupied by the components in the particle within the overall building site, measuring the effective use of space), streamline efficiency (assessing the efficiency of equipment movement between different components by analyzing the connectivity and path distance between components), component applicability (determining whether the components belong to categories that have a practical impact on streamline efficiency, such as aisles, walkways, and pipes), and violation count (counting the number of particles that do not meet layout specifications, such as the close distance between components (less than 0.5 meters)). By weighting these indicators, a layout quality score for each particle is obtained, thereby quantifying its multi-objective cost A.

[0041] The multi-objective cost threshold C is set by statistical analysis. If the current cost A is lower than the set threshold C, the particle position is updated by the particle swarm algorithm. Use the clamping method to constrain the new position:

[0042]

[0043] in, is the update speed of the i-th particle at time t+1, is the velocity at the current time t, from the previous iteration (DEAP storage), d1 and d2 are learning factors, standard APSO values ​​(hard-coded constants in Python), r1 and r2∈[0,1] are random factors obtained through Python, x min and x max The coordinate boundaries defined by the design specifications (such as the minimum spacing of 0.5 meters) and Clamp are clamp calculation methods used to limit the new position to the predefined boundaries. but like but Otherwise, keep is the current position of the i-th particle at time t G best is the individual optimal position, obtained by particle swarm optimization individual optimal selection method, g best is the global optimal position, obtained by the particle swarm optimization global optimal selection method, W t The dynamic inertia weight during iteration controls the influence of the previous velocity on the new velocity in the APSO algorithm and is calculated using the linear interpolation method.

[0044] Repeat the cost evaluation and particle update iterations to the set number of times, and output the optimal coordinate position corresponding to each component The optimal position coordinates form a coordinate set

[0045] By parsing CAD drawings based on the Revit API, extracting the geometric information, material properties, and functional labels of components, generating a particle swarm with initial position coordinates, and building a particle swarm optimization model with the help of the Python DEAP framework, this method effectively solves the problem that component layout optimization relies on empirical adjustment and is difficult to balance multi-objective costs in existing technologies. The Evaluate Cost function is used to comprehensively consider multi-objective costs such as space occupancy, component conflict, and layout rationality. The particle position is iteratively updated through the particle swarm algorithm, and the cost threshold and position clamping method are combined to ensure the convergence and feasibility of the optimization results. Finally, the globally optimal component layout coordinate set is generated. This method improves the refinement and automation level of component layout, overcomes the inefficiency and limitations of traditional layout methods, and is suitable for the intelligent creation and optimization of complex building information models (BIM).

[0046] Furthermore, generate component pair coordinates using the optimal position coordinate set, generate an RGB image based on the component pair coordinates, use Shapely to detect the RGB image to generate label conflicts and missing labels, extract the RGB image feature vector based on the convolutional neural network model, use the image features as node features, update the node features through the graph neural network model, use the feed-forward neural network to predict the conflict and missing probabilities, construct a graph structure based on the image feature vector, and calculate the environment function according to the graph structure and the predicted conflict and missing probabilities. Use Python to import the optimal position coordinate set into a Pandas DataFrame, and use the Pandas index and nested loops to generate component pairs (q, j) (referring to the pairing of any two components (such as pipes, walls, ventilation devices, etc.) in the Building Information Model (BIM). To calculate the relationships (such as Euclidean distance and cosine similarity) between all component pairs, the Pandas index ensures that each component has a unique identifier, and the nested loops generate all component pairs (q < j to avoid duplicates)), and generate the corresponding RGB image F for each component pair (q, j) based on Python Pillow (a library for processing images). qj :

[0047] F qj = Python Pillow(O q ,O j ,)

[0048] Retrieve the coordinates and dimensions (referring to the length (l) and width (w) of the component) for each pair of components (q, j) from the DataFrame using the Pandas index. Represent each component as a 2D rectangular component using the coordinates and dimensions (for 2D plane overlap detection, the z coordinate and height can be ignored). Use the geometric intersection test of Shapely to determine whether the rectangles overlap. If they overlap, mark it as a conflict label {y qj}}. Each component pair uses a binary label (1 indicates overlap, 0 indicates no overlap). For example, for 1000 components, 500 pairs are marked (e.g., y 12 = 1 indicates that pipe 1 overlaps with pipe 2). Filter the DataFrame to select the conflict label {y qj}}. For each space s (space in the building layout (such as a room, office, corridor)), create a 2D polygon representing the boundary using the coordinates and dimensions. The polygon is defined by four vertices ((x s ,y s ), (x s + l s ,y s s), (x s + l s ,y s s + w s ), (x s,y s +w s ), vertex x s ,y s is the coordinate of the lower left corner of the space, l s , w s Extract the length and width of the space from the DataFrame, parse the component function tags from the Revit API through PandasDataFrame and extract the space design specifications (such as offices must contain load-bearing walls and pipes), generate a list of necessary component types for each space s, and use Shapely's geometric inclusion test method to check whether each space s polygon (defined by four vertices) contains all necessary component types (based on the inclusion relationship between the component centroid coordinates and the polygon). If any component type required by the specification is missing, mark the space as missing label U s =1, otherwise marked as U s = 0, the missing label {U s} and the conflicting label {y qj} and image {F qj}stored in the labeled dataset;

[0049] Based on the convolutional neural network (CNN) model, the ResNet-18 architecture is used as the backbone network to qj Extract its geometric features:

[0050]

[0051] Among them, θ CNN The weights of ResNet-18 are pre-trained on ImageNet. The image is convolved and mapped layer by layer through the multi-layer residual structure of ResNet-18, and the final output is a 512-dimensional image feature.

[0052] Using the graph neural network (GNN) model, the component feature vector is used as the node feature and initialized as the node feature For each graph neural network layer, update the node features:

[0053]

[0054] Among them, p qj is an element of the adjacency matrix, which refers to the connection relationship between component node q and node j, indicating whether there is an edge in the graph structure. The adjacency matrix of NetworkX is generated. (L) is the weight matrix of the Lth layer, obtained by back propagation using Adam Optimizer, B(l) is the weight matrix of the layer, which transforms the node’s own features and is obtained by back propagation using Adam Optimizer, αqj The attention coefficient that determines the importance of neighbors is obtained by the transformer attention calculation method, σ is the activation function, activated by ReLU, is the input feature of the current layer (referring to the representation after aggregation of the previous layer);

[0055] Using Feedforward Neural Network (FNN) to predict conflicts and missing probability

[0056]

[0057] Based on the feature vector set, each building component (such as wall, beam, etc.) is defined as a node constituting a node set V, and a unique identifier is used as the index of the Pandas DataFrame. The Euclidean distance is calculated according to the position of the component in three-dimensional space, and the distance matrix is ​​implemented using NumPy. The edge set E is generated based on the set spatial proximity threshold. NetworkX is used to construct an undirected graph data structure G = (V, E) based on V and E. The undirected graph data structure G is connected to the collision probability and missing probability Encapsulate as JSON for integration, use Python to export JSON files, and integrate with BIM software;

[0058] Create a Gym environment with action and observation space, (the action space includes continuous position adjustment, scaled to [-0.5, 0.5] meters, binary decision to add ventilation equipment to a specific space, the total dimension is 4N (where N is the number of components), the observation space includes normalized coordinates, conflict probability and missing probability, the dimension is 3N+E+N, capturing layout dynamics, and achieving fine optimization through continuous adjustment, simplified component addition and comprehensive state representation. Each time an environment interaction is performed, a new time step t is entered. The agent (refers to the decision-making subject in reinforcement learning, which selects actions (such as adjusting coordinates) according to the current state vector and learns the optimal strategy through environmental feedback (reward)) takes actions according to the current state vector and obtains prediction results, making wise decisions in the DRL process). The state encoding of the Gym environment is used to convert the conflict probability into a state representation. and missing probability Convert to a state vector, encode the state vector through Array Flattening, and extract the encoded state vector and Calculate the environment function R based on the undirected graph data structure G and the conflict and missing probabilities t :

[0059]

[0060] Among them, w c and w m To balance the dynamic weights of conflict and missing goals, the weighted calculation based on sigmoid is obtained in [0, 1], t is the time step, which is obtained by the increment of Gym during the agent interaction, and e is the base of the natural logarithm, that is, a mathematical constant;

[0061] Through the combination of CNN and GNN, the geometric features and spatial relationships of components are effectively captured, the accuracy of conflict detection and missing judgment is greatly improved, missed detection and false detection are reduced, component pairs are automatically generated and conflicts and missing are detected without manual intervention, and the component layout optimization process is fully automated. The generated JSON data is easy to integrate with other BIM software, supports the collaborative work of multiple architectural design and optimization tools, supports arbitrarily complex BIM layouts, and adapts to architectural design needs of different scales and diversity. Based on the reinforcement learning framework, through the dynamic adjustment of the Gym environment, the layout plan can be quickly iterated and optimized, saving a lot of time and cost, supporting arbitrarily complex BIM layouts, and adapting to architectural design needs of different scales and diversity. Through the deep integration of algorithms and tool chains, the problems of low detection accuracy, poor optimization efficiency, and insufficient decision support in existing technologies are solved. It has significant technological breakthroughs and broad practical application value in the field of building information modeling.

[0062] S2. Output the optimized component coordinates and eigenvectors through the PPO algorithm based on the environmental function, generate the BIM model structure according to the optimized data, and write it into the IFC file;

[0063] Specifically, outputting optimized component coordinates and feature vectors through the PPO algorithm based on the environment function means initializing the Gym environment based on the custom Gym environment registration method. In the Gym environment, the policy network (MLP) in the PPO learning model is trained to determine actions (such as component coordinate adjustment) based on the current state vector. The policy network is configured as a two-layer 128-unit MLP, the learning rate is set to 0.0003, the batch size is 64, and the CUDA environment (NVIDIA RTX 3090) is loaded to achieve training acceleration and optimize dynamic weight calculation:

[0064]

[0065] Among them, Sigmoid is the weighted normalization method, μ is the conflict weight, and β is the adjustment coefficient of the missing weight, which is obtained by the grid search hyperparameter tuning method;

[0066] Update the PPO policy network based on the environment function R t, output the optimized component coordinates and component feature vectors, (run 1000 steps in the Gym environment, obtain the current state vector (dimensions include normalized component coordinates, conflict probability and missing probability), input the state vector into the PPO MLP strategy network to calculate the action distribution (mean and standard deviation), sample actions from the action distribution (Gaussian distribution), pass the sampled actions to the Gym environment, perform component coordinate adjustment, and the environment calculates the reward value based on the output action (integrating the weight calculation of conflict and missing). The PPO algorithm maximizes the cumulative reward and continuously updates the parameters of the strategy network based on gradient optimization, thereby improving the selection of the optimal action. If the reward change is lower than the set threshold L within 100 consecutive steps, the early stopping mechanism is triggered, and finally the optimized component coordinates and component feature vectors are output for subsequent BIM integration).

[0067] Continuous iterative training of the policy network through the PPO algorithm can achieve global optimization in large-scale complex layouts and avoid falling into local optimality. The environment function introduces dynamic conflict weights and missing weight mechanisms, allowing the optimization process to flexibly weigh multiple objectives according to actual conditions, improving the ability to adapt to different project needs. Based on the state vector in the Gym environment, the PPO policy network can intelligently perceive the current layout quality, automatically learn the optimal adjustment strategy, and improve the layout rationality and compliance with regulations. The action space design supports continuous position adjustment and component addition operations, so that the model can not only optimize the position of existing components, but also meet the needs of spatial function completion and improve the overall design integrity.

[0068] Furthermore, generating a BIM model structure based on the optimized data and writing it into an IFC file means creating an empty file based on the IFC4 standard, defining the project, site, and building in sequence, setting the basic structure of the file, and creating an IfcBuildingElement (an entity used in the IFC standard to express design constraints, such as minimum / maximum size restrictions, representing any physical element in a building, such as a wall, beam, or column) for each component by traversing and optimizing the component coordinates and eigenvectors. The specific settings include: i ,y i ,z i ) is converted to IfcCartesianPoint (an entity used to define the location of a point in the IFC standard), the size (l i ,w i ,h i ) is defined as IfcRectangleProfileDef (2D profile) and IfcExtrudedAreaSolid (3D extrusion), material m i Mapped to IfcMaterial (an entity used to define material properties in the IFC standard), function tag f iAdded as IfcPropertySet (a collection containing a set of properties (such as name, description, etc.) to describe the specific characteristics of the component, create an IfcSpace entity for each space s, define the boundary based on the spatial polygon data (four vertices), use IfcPolyLoop to represent the polygon outline, and store the geometric boundary of the space;

[0069] Based on the edge set E of the undirected graph data structure G, use ifcopenshell to create IfcRelConnectsElements entities to represent the spatial proximity relationship between components (such as the proximity between a pipe and a wall). E corresponds to one IfcRelConnectsElement, which records the connection relationship between components q and j and reflects the spatial topology structure in the BIM model.

[0070] Use ifcopenshell to write components, spaces, and relationships (spatial proximity relationships between components, such as the proximity between pipes and walls, obtained from the edge set E of the undirected graph data structure G) into an IFC file. Use ifcopenshell.validate to check the syntax and semantic integrity of the file to ensure compliance with the IFC4 standard. Use the ifcopenshell.file.write method to save it as an optimized_bim.ifc file to directly generate a BIM model containing optimized data (component geometry, materials, function labels, space boundaries, and ventilation devices).

[0071] Based on the optimized component coordinates and eigenvectors, IfcBuildingElement and IfcSpace entities are automatically created, reducing manual operations and significantly improving modeling efficiency and accuracy. IfcPolyLoop is used to define spatial polygon boundaries to ensure that the spatial geometric structure is accurate and complete and complies with design specifications. IfcRelConnectsElements relationship entities are generated through the edge set of the graph data structure G to ensure clear connection logic between components and realize the expression of model topology semantics. Optimized data directly drives model generation, avoiding repeated modeling and data conversion, forming an integrated process from design optimization to modeling output, and improving the automation and intelligence level of BIM projects. Files are created and verified strictly in accordance with the IFC4 standard, and ifcopenshell.validate is used to ensure syntax and semantic compliance, enhancing the compatibility and interactivity of the model between various BIM platforms.

[0072] S3. Build a real-time data streaming platform to achieve structured storage of BIM data, and perform real-time optimization and automatic update integration of BIM data;

[0073] Specifically, building a real-time data stream platform and realizing structured storage of BIM data means using Apache Kafka (version 3.5) to deploy a real-time data stream platform, configuring the topic bim data (implemented using the Kafka topic creation method to represent BIM-related data) for transmitting BIM data and status (dynamic information of the current building layout, including component coordinates, conflict probability and missing probability). In the system architecture, Kafka acts as a circulation hub for data updates, and realizes the event-driven mechanism of data stream by continuously receiving component position change information from sensors and real-time sources of construction progress monitoring, detecting conflict and missing status labels, and changing from 0 to 1, which is considered a potential problem. The component feature optimization method is called to locate the changed component and its surrounding affected area, extract relevant parameters and functional requirements, and analyze the adaptability of the component in the current environment and usage scenario, including structural safety, thermal performance and construction feasibility. According to the component category, the size (such as wall thickness, door and window size) and Material properties (such as strength, weight, and thermal conductivity) are optimized while meeting regulatory constraints. (For example, in the middle area of ​​the first floor of an office building, a door component shifts and conflicts with a wall. The system detects that after the layout change, the center coordinates of a door component overlap with the adjacent load-bearing wall. The status label changes from 0 to 1, indicating a conflict. The component feature optimization method is invoked to first lock the door component and its location, extract its current size, material, functional label, and surrounding wall structure information (200mm thickness, load-bearing, concrete material). The structural compatibility of the door in its new location with the wall is evaluated. It is found that the door width does not match the wall thickness, and the wooden door material cannot meet the fire rating requirements of the area. Under the premise of meeting regulatory constraints, the door width is adjusted and the material is changed from wood to fire-resistant steel to achieve structural adaptation and functional compliance.)

[0074] Initialize MongoDB (version 6.0) and create the collections bim_models (which stores BIM data parsed from IFC files, including components, spaces, and relationships, stored in JSON format) and graph_data (which stores data on the optimized graph G, including nodes V (components), edges E (adjacent relationships), and associated conflict and missing probabilities, stored in MongoDB document format). The optimized results of component sizes and material properties (such as new sizes and material properties) are then stored. Use ifc_openshell to parse the IFC file, extract components (including coordinates and sizes), spaces (including numbers and polygon boundaries), and relationships (including source component numbers and target component numbers), convert them to JSON format, and store them in the bim_models collection.

[0075] Kafka, as a high-throughput streaming media hub, supports continuous updates of component locations and status, breaking through the bottleneck of static BIM models that cannot respond to site dynamics, and achieving efficient linkage between models and construction progress. By monitoring changes in status tags, it can realize real-time detection and response to potential problems, effectively supporting risk warnings and intelligent adjustments during the construction process. Through MongoDB, components, spaces, relationships and optimization results are managed in a centralized manner, supporting fast retrieval and historical tracking, enhancing data integrity and availability, using ifcopenshell to parse standard IFC models, converting BIM information into structured JSON document storage, maintaining data semantic consistency and facilitating docking between systems, effectively improving the dynamic adaptability, data controllability and decision-making intelligence of the BIM system, and significantly outperforming existing passive and static data management methods.

[0076] Furthermore, real-time optimization and automatic update integration of BIM data refers to using Kafka consumers (confluent-kafka library) to subscribe to the bim_data topic, receive real-time update data (such as component position changes and construction progress detected by sensors), parse the new data into IFC format through ifcopenshell, update the component coordinates and spatial data in the MongoDB collection bim models, use the AE-HNN model (ResNet-18 to extract image features, GAT to update node features) to infer the current coordinates and optimized graph, combine NumPy to calculate the Euclidean distance to update the edge set and adjacency matrix of the graph, and obtain the current conflict probability and missing probability. Use the pre-trained LSTM model (two layers of 128 units, batch size 32, CUDA acceleration) to predict the conflict and missing probability of the next time step based on the current probability and GNN node features, calculate the current reward, and if it is lower than the statistical threshold, trigger PPO optimization to obtain the optimized component coordinates and feature vectors, update the IFC file through ifcopenshell, write the optimized coordinates to the component position, update the neighboring relationship, and save it as updated after verification. bim.ifc, a real-time visualization dashboard built with the Dash framework, displays conflict and missing tags as key diagnostic information, helping users identify changing trends in conflict areas. Updated bim.ifc can be imported into Autodesk Revit through the Revit API to update existing BIM model data in Revit.

[0077] Using ifcopenshell to parse and write real-time data ensures the standardization and scalability of data exchange, enhances the system's compatibility across multiple platforms, optimizes and infers the current graph structure and node status, and predicts the conflict and loss probabilities of the next time step, enabling early perception and preventive optimization of model quality changes. Combining NumPy and MongoDB to update graph structure edge sets ensures the consistency of the graph structure and the coherence of the optimization context after each optimization. The optimized component coordinates and proximity relationships are written to the IFC file through ifcopenshell and saved as updated_bim.ifc after verification, ensuring that the model always remains in its optimal state. This significantly improves the real-time, intelligent, and platform integration capabilities of the BIM system, breaking through the bottlenecks of existing technologies in response speed, prediction depth, and collaborative capabilities.

[0078] This embodiment also provides an AI-based BIM model intelligent creation system, including: CAD data parsing and particle swarm optimization modules for extracting component information to generate initial component coordinates and feature vectors, calculating multi-objective costs and outputting an optimal position coordinate set;

[0079] Feature extraction and graph construction module, used to extract image feature vectors, update node features, and predict conflict and missing probabilities;

[0080] Reinforcement learning optimization module, used for the PPO strategy network to output component coordinates and feature vectors;

[0081] BIM model generation module, used to write optimized coordinates and feature vectors into IFC files;

[0082] Real-time data streaming and dynamic optimization modules are used to transmit BIM data and status, and update BIM models through Revit API integration.

[0083] This embodiment also provides a computer device, which is suitable for a BIM model intelligent creation method based on an AI algorithm, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement a BIM model intelligent creation method based on an AI algorithm as proposed in the above embodiment.

[0084] The computer device may be a terminal, comprising a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0085] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements a method and system for intelligently creating a BIM model based on an AI algorithm as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0086] In summary, the present invention achieves efficient recognition of component conflicts and missing components by integrating image processing with convolutional neural networks (CNN) and graph neural networks (GNN), improves the model's adaptability and generalization capabilities, and dynamically captures the complex spatial and functional relationships between components by constructing a graph structure with image features as node features and combining it with a feedforward neural network for probability prediction. The system introduces a reinforcement learning mechanism, defines the action space and observation space by constructing a Gym-based reinforcement learning environment, introduces conflict and missing probabilities as state vectors, combines dynamic adjustment strategies, and utilizes reinforcement learning agents to adjust and optimize component positions and types. It can iteratively learn the optimal layout plan within multiple time steps, significantly improving the rationality and constructability of the BIM model configuration, making up for the inefficiency of "static verification and manual adjustment" in the existing modeling process, and significantly improving the automation, intelligence, and global optimality of the BIM modeling process.

[0087] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for intelligently creating a BIM model based on an AI algorithm, characterized by: include, The Revit API is used to parse CAD drawings to extract the initial coordinates of components. Particle coordinates are generated based on the initial coordinates of the components. The Evaluate Cost function based on the AI ​​algorithm calculates the multi-objective cost for each particle and outputs the optimal position coordinate set based on the multi-objective cost. Use the optimal position coordinate set to generate component pair coordinates, generate an RGB image based on the component pair coordinates, use Shapely to detect the RGB image and generate conflict and missing labels, extract the RGB image feature vector based on the convolutional neural network model, use the image features as node features, update the node features through the graph neural network model, use the feedforward neural network to predict conflict and missing probabilities, construct a graph structure based on the image feature vectors, and calculate the environment function based on the graph structure and the predicted conflict and missing probabilities; Based on the environment function, the PPO algorithm is used to output the optimized component coordinates and eigenvectors, and the BIM model structure is generated according to the optimized data and written into the IFC file; Build a real-time data streaming platform to realize structured storage of BIM data, and perform real-time optimization and automatic update integration of BIM data.

2. The method for intelligently creating a BIM model based on an AI algorithm according to claim 1, wherein: The method of extracting component initial coordinates by parsing CAD drawings based on Revit API, generating particle coordinates based on the initial coordinates, calculating multi-objective costs for each particle based on the Evaluate Cost function of the AI ​​algorithm, and outputting the optimal position coordinate set based on the multi-objective costs refers to using Revit API to parse CAD drawings, extracting component geometric information, material properties and function labels, the extracted component initial position coordinates are the initial component layout in the CAD drawings, and using the PythoNumPy method to generate a feature vector O for each component initial position coordinate. i :O i =[x,y,z,l,w,h,m1,…,m k ,f1,…,f n ,], the initial feature vector set {O1,O2,...,O i }, stored as a NumPy array, and a fixed number of particle coordinates are initialized using the Python DEAP framework for each component coordinate And set the initial velocity for the particles Use the Evaluate Cost function to calculate the multi-objective cost A for each particle and set the multi-objective cost threshold C. If the current cost A is lower than the set threshold C, the particle position is updated through the particle swarm algorithm. Use the clamping method to limit the new position, repeat the cost evaluation and particle update iteration to the set number of times, and output the optimal coordinate position corresponding to each component The optimal position coordinates form a coordinate set 3. The method for intelligently creating a BIM model based on an AI algorithm according to claim 2, wherein: The method uses the optimal position coordinate set to generate component pair coordinates, generates an RGB image based on the component pair coordinates, uses Shapely to detect the RGB image and generate conflict and missing labels, extracts RGB image feature vectors based on a convolutional neural network model, uses image features as node features, updates node features through a graph neural network model, predicts conflict and missing probabilities using a feedforward neural network, constructs a graph structure based on the image feature vector, and calculates the environment function based on the graph structure and the predicted conflict and missing probabilities. The method uses Python to import the optimal position coordinate set into a Pandas DataFrame, uses Pandas indexing and nested loops to generate component pairs (q, j), and generates a corresponding RGB image F for each component pair (q, j) based on Python Pillow. qj ; Retrieve the coordinates and dimensions for each pair of components (q, j) from the DataFrame using Pandas' index. Use the coordinates and dimensions to represent each component as a 2D rectangular component. Use Shapely's geometric intersection test to determine if the rectangles overlap. If they overlap, mark them with a conflict label {y qj }, using a binary label (1 for overlap, 0 for no overlap) for each artifact pair, filter the DataFrame to select the conflicting labels {y qj }, For each space s, create a 2D polygon representing the boundary using the coordinates and dimensions. The polygon is defined by four vertices ((x s ,y s )、(x s +l s ,y s )、(x s +l s ,y s +w s )、(x s ,y s +w s )) , parse component function labels from the Revit API and extract space design specifications through PandasDataFrame, generate a list of necessary component types for each space s, use Shapely geometry inclusion test method to check whether each space s polygon contains all necessary component types. If any component type required by the specification is missing, mark the space as missing label U s =1, otherwise marked as U s = 0, the missing label {U s } and the conflicting label {y qj } and image {F qj }stored in the labeled dataset; Based on the convolutional neural network (CNN) model, the ResNet-18 architecture is used as the backbone network to qj Extract its geometric features, use the graph neural network (GNN) model, take the component feature vector as the node feature, and initialize it as the node feature For each graph neural network layer, node features are updated and conflicts are predicted using a feedforward neural network. and missing probability Based on the feature vector set, each building component (such as wall, beam, etc.) is defined as a node constituting a node set V, and a unique identifier is used as the index of the Pandas DataFrame. The Euclidean distance is calculated according to the position of the component in three-dimensional space, and the distance matrix is ​​implemented using NumPy. The edge set E is generated based on the set spatial proximity threshold. NetworkX is used to construct an undirected graph data structure G = (V, E) based on V and E. The undirected graph data structure G is connected to the collision probability and missing probability Encapsulate it into JSON for integration, use Python to export the JSON file, create a Gym environment with action and observation space, and use the state encoding of the graph based on the Gym environment to encode the conflict probability and missing probability Convert to a state vector, encode the state vector through ArrayFlattening, and extract the encoded state vector and Calculate the environment function R based on the undirected graph data structure G and the conflict and missing probabilities t .

4. The method for intelligently creating a BIM model based on an AI algorithm according to claim 3, wherein: The component coordinates and feature vectors optimized by the PPO algorithm based on the environment function are initialized based on the custom Gym environment registration method, and the strategy network (MLP) in the PPO learning model is trained in the Gym environment, the action is determined according to the current state vector, the dynamic weight calculation is optimized, and the PPO strategy network is updated based on the environment function R. t , output the optimized component coordinates and component feature vectors.

5. The method for intelligently creating a BIM model based on an AI algorithm according to claim 4, wherein: Generating a BIM model structure based on optimized data and writing it into an IFC file means creating an empty file based on the IFC4 standard, defining the project, site, and building in sequence, setting the basic structure of the file, creating an IfcBuildingElement for each component by traversing the optimized component coordinates and feature vectors, and creating an IfcSpace entity for each space s. Boundaries are defined based on spatial polygon data, and polygon outlines are represented using IfcPolyLoop to store the geometric boundaries of the space. Based on the edge set E of the undirected graph data structure G, IfcRelConnectsElements entities are created using ifcopenshell. Components, spaces, and relationships are written to the IFC file using ifcopenshell. Ifcopenshell.validate is used to check the syntax and semantic integrity of the file to ensure compliance with the IFC4 standard. The file is saved as an optimized_bim.ifc file using the ifcopenshell.file.write method to directly generate a BIM model containing optimized data.

6. The method for intelligently creating a BIM model based on an AI algorithm according to claim 5, wherein: The construction of a real-time data stream platform and the realization of structured storage of BIM data refer to deploying a real-time data stream platform using Apache Kafka, configuring the topic bim data for transmitting BIM data and status. In the system architecture, Kafka acts as a circulation hub for data updates, and realizes an event-driven mechanism for data streams by continuously receiving component position change information from sensors and real-time sources of construction progress monitoring, detecting conflicts and missing status tags. When a status tag changes from 0 to 1, it is considered a potential problem. The component feature optimization method is called to automatically adjust the size and material properties according to the component category, and optimize the parameters under the premise of meeting the specification constraints. MongoDB is initialized, and the collection bim models and graph data and the optimization results of component size and material properties are created for data storage. The IFC file is parsed through ifcopenshell, and the components, spaces and relationships are extracted, converted into JSON format and stored in the bim_models collection.

7. The method for intelligently creating a BIM model based on an AI algorithm according to claim 6, wherein: The real-time optimization and automatic update integration of BIM data involves using a Kafka consumer to subscribe to the bim-data topic to receive real-time updated data, parsing the new data into IFC format using ifcopenshell, updating the component coordinates and spatial data in the MongoDB collection of bim models, using the AE-HNN model to infer the current coordinates and the optimized graph, and using NumPy to calculate the Euclidean distance to update the edge set and adjacency matrix of the graph to obtain the current conflict probability and missing probability. Using a pre-trained LSTM model to predict the conflict and missing probabilities for the next time step based on the current probabilities and GNN node features, the current reward is calculated. If it is below a statistical threshold, PPO optimization is triggered to obtain optimized component coordinates and feature vectors. The IFC file is updated using ifcopenshell, the optimized coordinates are written to the component locations, the neighbor relationships are updated, and after verification, the file is saved as updated bim.ifc. The conflict and missing labels are displayed as key diagnostic information in a real-time visualization dashboard built using the Dash framework to assist users in determining the changing trends of conflict areas. The updated bim.ifc is imported into Autodesk Revit via the Revit API to update the existing BIM model data in Revit.

8. A BIM model intelligent creation system based on an AI algorithm, based on the BIM model intelligent creation method based on an AI algorithm according to any one of claims 1 to 7, characterized in that: include, CAD data parsing and particle swarm optimization modules are used to extract component information, generate initial component coordinates and feature vectors, calculate multi-objective costs, and output the optimal position coordinate set; Feature extraction and graph construction module, used to extract image feature vectors, update node features, and predict conflict and missing probabilities; Reinforcement learning optimization module, used for the PPO strategy network to output component coordinates and feature vectors; BIM model generation module, used to write optimized coordinates and feature vectors into IFC files; Real-time data streaming and dynamic optimization modules are used to transmit BIM data and status, and update BIM models through Revit API integration.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for intelligently creating a BIM model based on an AI algorithm as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for intelligently creating a BIM model based on an AI algorithm according to any one of claims 1 to 7 are implemented.

Images