Fine building model generation method based on spatial intelligence and graph semantic driving

Abstract

The application provides a fine building model generation method based on spatial intelligence and graph semantic driving, and belongs to the field of digital twinning; solves the problems of missing model details and low efficiency of artificial modeling in the building scheme deepening stage; the specific steps are as follows: obtaining a building skeleton model, converting it into a semantic graph containing multi-parent connection relationship; based on real building component data, calculating the normalized coordinates of the components relative to their host wall, and using a clustering algorithm to construct a slot library; constructing a WallCount module and an Anchor TransformerGCN module, combining the semantic graph and the slot library, predicting and determining the geometric properties and topological connection relationship of the to-be-confirmed components; selecting the to-be-confirmed components and converting the finally output model; the hybrid deep learning model of the graph neural network and the Transformer architecture is used to realize the automatic enrichment of details from the low-precision scheme model to the high-precision construction drawing model.

Description

Technical Field

[0001] This invention relates to a method for generating refined architectural models based on spatial intelligence and graph semantics, belonging to the fields of architectural design and digital twins. Background Technology

[0002] In the fields of architectural design and digital twins, existing technologies face the following problems when addressing the challenge of "how to quickly generate high-quality 3D architectural models":

[0003] Visually plausible but physically unusable: Existing mainstream generative AI (such as Stable Diffusion and Midjourney) mainly generates images based on two-dimensional pixels; although the visual effects are realistic, they lack three-dimensional spatial intelligence and cannot understand the physical attachment relationship between building components (such as windows and balconies) and walls; the generated images cannot be converted into constructible BIM models, resulting in a "semantic gap".

[0004] Lack of precision and scale: Existing 3D generation methods (such as Building-GAN) are mostly focused on the macro level of room layout or block stacking (LOD 2.0-3.0), and cannot go deep into the LOD 3.3 (refined construction drawing) level, making it difficult to automatically generate detailed structures such as window frames and grids.

[0005] Limitations of rule-based modeling: Traditional parametric modeling relies on manually writing complex geometric rule scripts. Once the architectural style changes, the rules need to be rewritten, resulting in poor versatility and low efficiency.

[0006] Lack of physical construction logic: Existing deep learning-based 3D generative models often have "physical errors" such as suspended components and overlapping components, and the generated model data is "unclean" and cannot be directly used for engineering quantity calculation or simulation. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the purpose of this invention is to provide a refined architectural model generation method based on spatial intelligence and graph semantics, aiming to solve the problem of low quality in the generation of 3D architectural models.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solution: a refined architectural model generation method based on spatial intelligence and graph semantics includes:

[0009] Obtain the LOD 3.2 skeleton model of the building, extract the geometric entities in the model, and transform it into a semantic graph containing multiple parent-level connections; based on real building component data, calculate the normalized coordinates of each component in each building relative to its host wall, and use clustering algorithms to construct a slot library;

[0010] The WallCount module is constructed to process the semantic graph and predict the number of detailed components that should be generated on each wall of the standard floor. Based on the predicted number and the initial nodes instantiated from the slot library, the Anchor TransformerGCN module is constructed to predict the geometric properties and topological connections of the components to be confirmed.

[0011] By applying physical attachment constraints and combining the geometric properties and topological connections of the components to be verified, geometric corrections and conflict eliminations are performed on the components to be verified, and the final output model is converted into a usable format for the BIM system.

[0012] Furthermore, the transformation steps of the semantic graph are as follows:

[0013] The geometric entities in the LOD 3.2 model are discretized into graph nodes, and the corresponding node definitions are as follows:

[0014] Wall node: serves as the host for generation;

[0015] Opening node: as an existing geometric constraint;

[0016] Component node: The prediction target to be generated;

[0017] The parameters corresponding to each node are standardized, and a fixed-dimensional feature vector is encoded for each node:

[0018] Type coding: One-hot coding is used to identify walls, doors, windows, or components to be identified;

[0019] Geometric center: The coordinates of the geometric center of the node; to ensure the model's adaptability to buildings of different scales, the coordinates are normalized to a local coordinate system relative to the center of its parent wall.

[0020] Enclosing box dimensions: length, width, and height of the components;

[0021] Node degree: Represents the number of other nodes connected to this node;

[0022] Edges and multi-parent topologies:

[0023] When constructing a semantic graph, a component node is allowed to have multiple edges pointing to different wall nodes simultaneously;

[0024] The output header of the second-stage network contains a connection header that outputs a connection probability matrix, representing the connection probability between the current component and all surrounding walls. When the probability exceeds a threshold, a connection is established.

[0025] When a component is determined to be connected to two or more walls, it is determined that there is a multi-parent topological relationship between the component and the walls it is connected to.

[0026] Furthermore, the steps for constructing the slot library are as follows:

[0027] Collect component data of real buildings, extract the positional information of these components relative to the host walls of their respective buildings, clean the positional information, and then represent it in the form of normalized coordinates.

[0028] The K-Means clustering algorithm is selected to perform unsupervised learning on the processed coordinate data, and then all the coordinate data is divided into K different clusters, each represented by its centroid; where K represents the number of clusters.

[0029] After K-Means clustering is completed, the centroid coordinates of each cluster are extracted as slots. The slot coordinate information corresponding to all the determined clusters is organized and stored in a structured dataset to form a slot library.

[0030] The WallCount module is constructed to process the semantic graph and predict the number of detailed components that should be generated on each wall of the standard floor.

[0031] Furthermore, the construction steps for the WallCount module are as follows:

[0032] The network architecture of the WallCount quantity prediction module is a hybrid of Transformer and GCN.

[0033] In the GCN layer, wall nodes will detect the connection information of door and window component nodes, opening nodes, and other wall nodes that are directly connected to them;

[0034] The connection information specifically includes: the type and size of the doors and windows, and their location on the wall;

[0035] The location, shape, and size of the openings to be reserved in the wall for installing doors and windows;

[0036] The location, orientation, length, and thickness of adjacent walls;

[0037] In the Transformer layer, the self-attention mechanism is used to enable the model to obtain the four facades of each building model in the corresponding standard floor, as well as the state of all walls and doors and windows in the four facades;

[0038] During training, the model minimizes the differences in decorative features of different facades within the same standard floor while maximizing the differences between different standard floors, predicts the number of detailed components that should be generated on each wall of the standard floor, and constructs the AnchorTransformerGCN module.

[0039] Furthermore, the specific processing steps of the Transformer layer are as follows:

[0040] Find the four facades corresponding to the standard floor of each building model, encode the components such as walls, doors, windows, and openings of the four facades into high-dimensional feature vectors, and arrange them in facade order as model input;

[0041] Self-attention mechanism calculation: The model input is mapped to a Query, Key, and Value matrix through a linear transformation;

[0042] Based on the matrix Query and Key, as well as the dimensions of the matrix Key, the Softmax algorithm is used to calculate the attention weight of each facade relative to other facades;

[0043] The attention weights calculated by the Softmax algorithm are multiplied with matrix V to aggregate global information and generate features that contain global dependencies.

[0044] Style consistency constraints are set during model training: a loss function for style consistency is introduced into the feedforward network of the Transformer to force the decorative features of different facades to align in the latent space.

[0045] Furthermore, the steps to construct the Anchor TransformerGCN module are as follows:

[0046] The network architecture of the Anchor TransformerGCN attribute generation module is a parallel dual-branch network;

[0047] The GCN branch is used to handle the micro-geometric relationships between components and their host walls, as well as the doors and windows on the corresponding host walls, ensuring that components do not obstruct door and window openings or extend beyond the wall boundaries. The specific implementation process is as follows:

[0048] Based on the number of detailed components predicted in the first stage, specific door and window components are randomly pre-generated. The nodes of the pre-generated door and window components are used as initial candidate nodes. The relevant building information of the initial candidate nodes is obtained to construct the graph structure data.

[0049] Graph convolution kernels are used to perform convolution operations on graph structure data, aggregating information of each node and its neighboring nodes, and through multi-layer graph convolution, higher-level features related to components, such as semantics, geometry, physical constraints and functions, are gradually extracted.

[0050] The convolutional features are normalized and activated; physical constraints are introduced during feature processing to ensure that the generated components do not obstruct door and window openings or exceed the wall boundaries.

[0051] Transformer branch: Coordinates the size and style of all components on the four facades of the standard floor through a global attention mechanism.

[0052] Furthermore, the specific implementation of the Transformer branch is as follows:

[0053] Arrange all component information on the four facades of the standard floor into a sequence according to the order in which the components belong to the facades, and use it as the input to the Transformer;

[0054] The self-attention weights between each element in the sequence and other elements are calculated using the Softmax function.

[0055] The value vectors of each element are weighted and summed according to the self-attention weights, and then new features containing global style are generated through a multi-layer self-attention mechanism and a feedforward neural network.

[0056] In the process of generating new features, a style consistency loss function is introduced to force the decorative features of different facades to align in the latent space;

[0057] Geometry Head: Uses a fully connected layer to perform regression prediction on the features generated by the Transformer to obtain the geometric properties of the component;

[0058] Connector: By using a fully connected layer and a Sigmoid activation function, the connection probability matrix between the current component and all surrounding walls is predicted, and a threshold Threshold is set. If a connection probability is greater than the threshold, it is considered that the current component has a physical connection with the corresponding wall. The topological connection relationship between the component and its corresponding wall is established, and the component to be confirmed is obtained.

[0059] Furthermore, the physical attachment constraints are as follows:

[0060] Set the physical adhesion loss function : ;

[0061] in, It is the normalized normal distance from the center of the component to be confirmed to the corresponding wall plane; this formula requires that the center of the component be along the wall normal direction and fall on the outer surface of the wall.

[0062] Using a dynamic weight training strategy:

[0063] To address the initial convergence issue in 3D generation tasks, dynamic weight adjustment is employed.

[0064] Exploration phase: The weights are extremely low, allowing the model to learn the layout rules of a two-dimensional plane first;

[0065] Adsorption period: Gradually increases during model training. The weights; after the model completes the learning of the layout rules, it adjusts the position of the components based on the calculation results of the physical attachment loss function; through continuous adjustment of the model, the normalized normal distance from the center of the component to the wall plane is kept within 1.0 to 2.0, thereby achieving accurate prediction of the component position;

[0066] Perform forced geometric calculations to eliminate tiny floating-point errors in the model's predictions during deep learning:

[0067] The model obtains the plane equation of the parent wall, projects the center of the component onto the plane based on the plane equation of the parent wall, and aligns the normal direction to complete the selection of the primary component.

[0068] Furthermore, the conflict removal process is as follows:

[0069] The 3D IoU between each primary component is calculated using the 3D NMS algorithm, and redundant components that are close to or overlap are removed. The non-maximum suppression algorithm is used to remove overlapping redundancy and retain the primary component with the highest confidence.

[0070] The semantic graph data corresponding to the primary components is parsed, and the corresponding components are generated in the building LOD 3.2 skeleton model to obtain the building model. The format of the building model is adjusted to the standard BIM format, and the corresponding parametric family is called according to the predicted Type ID of the model. The family parameters are set according to the predicted size parameters to complete the three-dimensional fine generation of the building facade.

[0071] Compared with the prior art, the beneficial effects of the present invention are:

[0072] Physical constructability: By introducing physical attachment loss ( This completely solves the common problems of component suspension and clipping in generated models, and the generated models have engineering-level physical accuracy.

[0073] Fully automated detailed design: This method can not only store city models, but also understand the planning logic behind them (FAR, coverage, architectural style), and automatically generate an infinite variety of derivative models based on these logics, realizing end-to-end automation from rough blocks to fine facades, which greatly improves the efficiency of design detailing.

[0074] Style Adaptation: Based on the prior mechanism of Slot Bank, the model can automatically learn and reproduce the architectural style rhythm implicit in the dataset without the need for manual rule writing; at the same time, this invention also supports the automatic inference and generation of complex multi-parent connection relationships such as corner windows and cross-layer components. Attached Figure Description

[0075] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0076] Figure 1 This is a schematic diagram of the method of the present invention;

[0077] Figure 2 This is a schematic diagram illustrating the construction of the slot bank of the present invention;

[0078] Figure 3 This is a schematic diagram of the process of the present invention;

[0079] Figure 4 This is a schematic diagram of the model output of the present invention. Detailed Implementation

[0080] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0081] Please see Figure 1 and Figure 3 The methods for generating refined building models based on spatial intelligence and graph semantics include:

[0082] Step S1: Obtain the LOD 3.2 skeleton model of the building, extract the geometric entities in the model, and transform it into a semantic graph containing multiple parent connections; based on real building component data, calculate the normalized coordinates of each component in each building relative to its host wall, and use a clustering algorithm to construct a slot bank.

[0083] The specific conversion steps are as follows:

[0084] The geometric entities (such as walls, openings, etc.) in the LOD 3.2 model are discretized into graph nodes, and the corresponding node definitions are as follows:

[0085] Wall node (Type=1): As the host of generation, the wall node is the basic structural unit in the spatial semantic graph. Its core function is to serve as the "host" for other components (such as windows, balconies, etc.), indicating that these components need to be attached to the wall for generation.

[0086] Opening node (Type=2): as an existing geometric constraint.

[0087] Component node (Type=3): The prediction target to be generated, such as windows, balconies, etc.;

[0088] The parameters corresponding to each node are standardized, and a fixed-dimensional feature vector is encoded for each node:

[0089] Type coding (Type, 3dims): One-hot coding is used to identify walls, doors, windows, or components to be identified;

[0090] Geometric center (Centroid, 3dims): The coordinates of the geometric center of the node; to ensure the model's adaptability to buildings of different scales, the coordinates are normalized to a local coordinate system relative to the center of its parent wall.

[0091] Bounding box dimensions (3dims): length, width, and height of the component;

[0092] Node degree (1dim): Represents the number of other nodes connected to this node;

[0093] Edges and Multi-parent Topology:

[0094] When constructing a semantic graph, a "component node" is allowed to have multiple edges pointing to different "wall nodes" at the same time;

[0095] The output head of the second-stage network in the Anchor TransformerGCN model contains a link head, which outputs a connection probability matrix (i.e., the link head does not output a single parent node ID, but instead outputs a connection probability matrix).

[0096] This connection probability matrix represents the connection probability between the current component and all surrounding walls. When the probability exceeds a threshold, a connection is established.

[0097] When a component is determined to be connected to two or more walls, it is determined that "there is a multi-parent topology relationship between the component and the walls it is connected to"; for example, if a "corner window" is determined to be "connected" to both wall A and wall B, then "corner window and" belongs to a multi-parent topology relationship.

[0098] It should be noted that traditional BIM data structures (such as IFC trees) typically only support a single parent level (a window belongs to a wall); however, in reality, corner windows or curtain wall components are often attached to multiple structures simultaneously; the graph structure design of this invention allows for multiple parent connections: in the graph data, a component node can simultaneously have edges pointing to different wall nodes. This non-tree-like mesh topology can accurately describe complex spatial attachment relationships and solve the problem of corner components being difficult to express.

[0099] Please see Figure 2Based on real building component data, the normalized coordinates of each component relative to the host wall in its respective building are calculated, and a slot bank is constructed using a clustering algorithm as a priori for spatial layout.

[0100] Among them, the slot library serves as prior knowledge, used to "guide" the prediction of the spatial layout of components on the wall;

[0101] Data collection and preprocessing:

[0102] Collect component data of real buildings, extract the positional information of these components relative to the host walls of their respective buildings, clean the positional information, and then represent it in the form of normalized coordinates.

[0103] The K-Means clustering algorithm is selected to perform unsupervised learning on the processed coordinate data (i.e., the coordinate data transformed from the location information mentioned above). Then, all the coordinate data is divided into K different clusters, each represented by its centroid. Here, K represents the number of clusters, and the value of K is determined by the Elbow Method.

[0104] After K-Means clustering is completed, the centroid coordinates are extracted from each cluster. These centroids are called "slots". The coordinate information of the "slots" corresponding to all the determined clusters is organized and stored in a structured dataset to form a "slot library".

[0105] Step S2: Construct the WallCount module to process the semantic graph and predict the number of detailed components to be generated on each wall of the standard floor; based on the predicted number and the initial nodes instantiated in the slot library, construct the Anchor TransformerGCN module to predict the geometric properties and topological connections of the components.

[0106] Phase 1: WallCount Quantity Prediction Module

[0107] Input: A LOD 3.2 semantic skeleton diagram of a standard floor building (including wall nodes and door and window opening nodes);

[0108] Task: Predict the number of detailed components that should be generated on each wall of a standard floor (classification / regression task);

[0109] Network architecture: Transformer-GCN hybrid;

[0110] GCN layer (Graph Convolutional Network): Responsible for aggregating local neighborhood features: wall nodes will detect the connection information of door and window component nodes, opening nodes and other wall nodes directly connected to them;

[0111] The connection information specifically includes: the type and size of the doors and windows (i.e., the doors and windows installed on the wall) and their location on the wall;

[0112] The location, shape, and size of the openings to be reserved in the wall for installing doors and windows;

[0113] The location, orientation, length, and thickness of adjacent walls;

[0114] Transformer layer: Responsible for capturing global dependencies of the standard layer: Utilizing a self-attention mechanism, the model acquires the state of the four facades (east, west, south, and north) within the standard layer corresponding to each building model, as well as the state of all walls and doors / windows within those four facades. The specific processing steps are as follows:

[0115] Find the four facades corresponding to the standard floor of each building model, encode the components such as walls, doors, windows, and openings of the four facades into high-dimensional feature vectors (such as normalized coordinates, type embedding, and decorative style embedding), and arrange them in facade order (east → west → south → north) as model input;

[0116] It should be noted that the term "standard floor" in this invention refers to a standard floor in architectural design. In high-rise buildings, in order to improve efficiency, floors with the same or similar functions are usually designed as "standard floors" (such as standard office floors in office buildings and typical residential floors in residential buildings). These floors are highly consistent in terms of floor plan (wall location, door and window openings), structural form (column grid, beam height), etc., with only slight differences in local details (such as decoration style and equipment layout).

[0117] Self-attention mechanism calculation: The model input is mapped to a Query(Q), Key(K), Value(V) matrix through a linear transformation;

[0118] Based on the matrices Query(Q) and Key(K) and the dimensions of the matrix Key(K), the Softmax algorithm is used to calculate the attention weight of each facade relative to other facades (e.g., the attention weight of the south facade to the north facade, and the attention weight to the east facade).

[0119] The attention weights calculated by the Softmax algorithm are multiplied by the matrix V (Value) to aggregate global information and generate features that contain global dependencies.

[0120] Style consistency constraint: In the feedforward network (FFN) of Transformer, a loss function for style consistency (such as contrastive learning loss, KL divergence) is introduced to force the decorative features of different facades (such as decoration density, color distribution) to align in the latent space ("latent space" is a low-dimensional abstract feature representation space in deep learning);

[0121] During training, the model minimizes the differences in decorative features of different facades within the same standard floor, while maximizing the differences between different standard floors (to avoid style confusion), and predicts the number of detailed components that should be generated on each wall of the standard floor.

[0122] Phase Two: Anchor TransformerGCN Property Generation Module;

[0123] Input: Initial candidate nodes instantiated in the graph using the slot library based on the predicted number of detailed components generated;

[0124] Task: Predict the precise geometric properties and topological connections of components;

[0125] Network architecture: Parallel Dual-Branch Network.

[0126] The GCN branch (Physical Attachment Branch) is used to handle the micro-geometric relationships between components and their host walls, as well as the doors and windows on the corresponding host walls, ensuring that components do not obstruct door and window openings or extend beyond the wall boundaries. The specific implementation process is as follows:

[0127] Based on the number of detailed components predicted in the first stage (WallCount quantity prediction module), specific door and window components are randomly pre-generated. The nodes of the pre-generated door and window components are used as initial candidate nodes. The relevant building information of the initial candidate nodes (such as the location and attributes of walls, doors and windows) is obtained and constructed into graph structure data.

[0128] Graph convolution kernels are used to perform convolution operations on graph structure data, aggregating information of each node and its neighboring nodes, and through multi-layer graph convolution, higher-level features related to components, such as semantics, geometry, physical constraints and functions, are gradually extracted.

[0129] The convolutional features are normalized and activated (to enhance the model's non-linear expressive power). During feature processing, physical constraints are introduced to ensure that the generated components do not obstruct door and window openings or exceed wall boundaries; for example, by setting boundary constraint functions, the position and size range of the components are limited.

[0130] Transformer branch (style unification branch): Through a global attention mechanism, the size and style of all components on the four facades of the standard layer are coordinated. The specific implementation is as follows:

[0131] All component information (including geometric attributes and decorative features) on the four facades of the standard floor are arranged into a sequence according to the order in which the components belong to the facades, and used as input for the Transformer;

[0132] The self-attention weights between each element in the sequence and other elements are calculated using the Softmax function.

[0133] The value vectors of each element are weighted and summed according to the self-attention weights, and then new features containing global style are generated through a multi-layer self-attention mechanism and a feedforward neural network.

[0134] In the process of generating new features, a style consistency loss function (such as contrastive learning loss or KL divergence) is introduced to force the decorative features of different facades to align in the latent space, ensuring that the generated style is coordinated and unified.

[0135] The output head (Multi-Head Output) includes: Geometry Head and LinkHead;

[0136] Geometry Head: Uses fully connected layers to perform regression prediction on the features generated by the Transformer to obtain the geometric properties of the component, including size, position offset, normal distance, etc.

[0137] Link Head: Through a fully connected layer and a sigmoid activation function, the connection probability matrix between the current component and all surrounding walls is predicted, and a threshold is set. If a connection probability is greater than the threshold, it is considered that the current component has a physical connection with the corresponding wall, and the topological connection relationship between the component and its corresponding wall is established to obtain the component to be confirmed.

[0138] Step S3: Apply Physical Attachment Constraint, and combine it with the geometric properties and topological connections of the component to be confirmed, to perform geometric correction and conflict elimination on the component to be confirmed, and convert the final output model into a usable format of the BIM system;

[0139] Please see Figure 4 Physical attachment constraints:

[0140] Set the physical attachment loss function. ):

[0141] Add a penalty term to the loss function of the neural network to constrain the normal distance:

[0142] ;

[0143] in It is the normalized normal distance from the center of the component to the plane of the wall corresponding to the component; this formula requires that the center of the component be along the direction of the wall normal and fall on the outer surface of the wall.

[0144] Dynamic weight training strategy (Warm-up Strategy):

[0145] To address the initial convergence issue in 3D generation tasks, dynamic weight adjustment is employed.

[0146] Early Stage: The weights are extremely low, allowing the model to first learn the layout rules of the two-dimensional plane (i.e., the positional distribution of windows on the two-dimensional plane of the wall).

[0147] Late Stage: Gradually increases during model training. The weights; after the model completes the learning of the layout rules, the model adjusts the position of the components according to the calculation results of the physical attachment loss function; through the continuous adjustment of the model, the normalized normal distance from the center of the component to the wall plane is within 1.0 to 2.0, so as to achieve accurate prediction of the position of the component;

[0148] Geometric Projection Post-treatment:

[0149] To eliminate tiny floating-point errors in predictions during deep learning (e.g., predicting a 0.5mm gap), the model performs forced geometric calculations:

[0150] The model obtains the plane equation of the parent wall, projects the center of the component onto the plane according to the plane equation of the parent wall, and aligns the normal direction to complete the selection of the primary component (that is, in the "components to be confirmed generated in step S2", select the component that meets the requirements as the primary component).

[0151] Conflict elimination and BIM conversion:

[0152] Conflict Removal (NMS): The 3D IoU between each primary component is calculated using the 3D NMS (Non-Maximum Suppression) algorithm, and redundant components that are close in position (i.e., the straight-line distance between two or more components is less than 3 cm) or overlap are removed. The non-maximum suppression algorithm is used to remove overlapping redundancy and retain the primary component with the highest confidence.

[0153] The semantic graph data corresponding to the primary components is parsed, and the corresponding components are generated in the building LOD 3.2 skeleton model to obtain the building model. The format of the building model is adjusted to the standard BIM format (such as IFC or Revit API object). The corresponding parametric family is called according to the predicted Type ID of the model. The family parameters are set according to the predicted size parameters to complete the three-dimensional detailed generation of the building facade.

[0154] The above formulas are all dimensionless calculations. The formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation. For example, there are weighting coefficients and proportional coefficients. The values ​​set are to quantify each parameter to obtain a specific value, which is convenient for subsequent comparison. The values ​​of the weighting coefficients and proportional coefficients are only required to not affect the proportional relationship between the parameters and the quantified values.

[0155] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for generating refined architectural models based on spatial intelligence and graph semantics, characterized in that, The method is as follows: Obtain the building skeleton model, extract the geometric entities in the model, and transform it into a semantic graph containing multiple parent-level connections; based on real building component data, calculate the normalized coordinates of each component in each building relative to its host wall, and use clustering algorithms to construct a slot library; The steps to build a slot library are as follows: Collect component data of real buildings, extract the positional information of these components relative to the host walls of their respective buildings, clean the positional information, and then represent it in the form of normalized coordinates. The K-Means clustering algorithm is selected to perform unsupervised learning on the processed coordinate data, and then all the coordinate data is divided into K different clusters, each represented by its centroid; where K represents the number of clusters. After K-Means clustering is completed, the centroid coordinates of each cluster are extracted as slots. The slot coordinate information corresponding to all the determined clusters is organized and stored in a structured dataset to form a slot library. The WallCount module is constructed to process the semantic graph and predict the number of detailed components that should be generated on each wall of the standard floor. Based on the predicted number and the initial nodes instantiated from the slot library, the Anchor TransformerGCN module is constructed to predict the geometric properties and topological connections of the components to be confirmed. The steps to build the WallCount module are as follows: The network architecture of the WallCount quantity prediction module is a hybrid of Transformer and GCN. In the GCN layer, wall nodes will detect the connection information of door and window component nodes, opening nodes, and other wall nodes that are directly connected to them; The connection information specifically includes: the type and size of the doors and windows, and their location on the wall; The location, shape, and size of the openings to be reserved in the wall for installing doors and windows; The location, orientation, length, and thickness of adjacent walls; In the Transformer layer, the self-attention mechanism is used to enable the model to obtain the four facades of each building model in the corresponding standard floor, as well as the state of all walls and doors and windows in the four facades; During training, the differences in decorative features of different facades within the same standard floor are minimized while the differences between different standard floors are maximized. The number of detailed components that should be generated on each wall of the standard floor is predicted, and the Anchor TransformerGCN module is constructed. The steps to build the Anchor TransformerGCN module are as follows: The network architecture of the Anchor TransformerGCN attribute generation module is a parallel dual-branch network; The GCN branch is used to handle the micro-geometric relationships between components and their host walls, as well as the doors and windows on the corresponding host walls, ensuring that components do not obstruct door and window openings or extend beyond the wall boundaries. The specific implementation process is as follows: Based on the number of detailed components predicted in the first stage, specific door and window components are randomly pre-generated. The nodes of the pre-generated door and window components are used as initial candidate nodes. The relevant building information of the initial candidate nodes is obtained to construct the graph structure data. Graph convolution kernels are used to perform convolution operations on graph structure data, aggregating information of each node and its neighboring nodes, and through multi-layer graph convolution, higher-level features related to components, such as semantics, geometry, physical constraints and functions, are gradually extracted. The convolutional features are normalized and activated; physical constraints are introduced during feature processing to ensure that the generated components do not obstruct door and window openings or exceed the wall boundaries. Transformer branch: Coordinates the size and style of all components on the four facades of the standard floor through a global attention mechanism; By applying physical attachment constraints and combining the geometric properties and topological connections of the components to be verified, geometric corrections and conflict eliminations are performed on the components to be verified, and the final output model is converted into a usable format for the BIM system.

2. The method for generating refined architectural models based on spatial intelligence and graph semantics as described in claim 1, characterized in that, The steps for transforming a semantic graph are as follows: The geometric entities in the model are discretized into graph nodes, and the corresponding node definitions are as follows: Wall node: serves as the host for generation; Opening node: as an existing geometric constraint; Component node: The prediction target to be generated; The parameters corresponding to each node are standardized, and a fixed-dimensional feature vector is encoded for each node: Type coding: One-hot coding is used to identify walls, doors, windows, or components to be identified; Geometric center: The coordinates of the geometric center of the node; to ensure the model's adaptability to buildings of different scales, the coordinates are normalized to a local coordinate system relative to the center of its parent wall. Enclosing box dimensions: length, width, and height of the components; Node degree: Represents the number of other nodes connected to this node; Edges and multi-parent topologies: When constructing a semantic graph, a component node is allowed to have multiple edges pointing to different wall nodes simultaneously; The output header of the second-stage network contains a connection header that outputs a connection probability matrix, representing the connection probability between the current component and all surrounding walls. When the probability exceeds a threshold, a connection is established. When a component is determined to be connected to two or more walls, it is determined that there is a multi-parent topological relationship between the component and the walls it is connected to.

3. The method for generating refined architectural models based on spatial intelligence and graph semantics as described in claim 1, characterized in that, The specific processing steps of the Transformer layer are as follows: Find the four facades corresponding to the standard floor of each building model, encode the walls, doors, windows, and openings of the four facades into high-dimensional feature vectors, and arrange them in facade order as model input; Self-attention mechanism calculation: Mapping the model input to a Query, Key, and Value matrix through linear transformation; Based on the matrix Query and Key, as well as the dimensions of the matrix Key, the Softmax algorithm is used to calculate the attention weight of each facade relative to other facades; The attention weights calculated by the Softmax algorithm are multiplied with matrix V to aggregate global information and generate features that contain global dependencies. Style consistency constraints are set during model training: a loss function for style consistency is introduced into the feedforward network of the Transformer to force the decorative features of different facades to align in the latent space.

4. The method for generating refined architectural models based on spatial intelligence and graph semantics as described in claim 1, characterized in that, The specific implementation of the Transformer branch is as follows: Arrange all component information on the four facades of the standard floor into a sequence according to the order in which the components belong to the facades, and use it as the input to the Transformer; The self-attention weights between each element in the sequence and other elements are calculated using the Softmax function. The value vectors of each element are weighted and summed according to the self-attention weights, and then new features containing global style are generated through a multi-layer self-attention mechanism and a feedforward neural network. In the process of generating new features, a style consistency loss function is introduced to force the decorative features of different facades to align in the latent space; Geometry Head: Uses a fully connected layer to perform regression prediction on the features generated by the Transformer to obtain the geometric properties of the component; Connector: By using a fully connected layer and a Sigmoid activation function, the connection probability matrix between the current component and all surrounding walls is predicted, and a threshold Threshold is set. If a connection probability is greater than the threshold, it is considered that the current component has a physical connection with the corresponding wall. The topological connection relationship between the component and its corresponding wall is established, and the component to be confirmed is obtained.

5. The method for generating refined architectural models based on spatial intelligence and graph semantics as described in claim 1, characterized in that, The physical attachment constraints are as follows: Set the physical adhesion loss function : ; in, It is the normalized normal distance from the center of the component to be confirmed to the corresponding wall plane; this formula requires that the center of the component be along the wall normal direction and fall on the outer surface of the wall. Using a dynamic weight training strategy: To address the initial convergence issue in 3D generation tasks, dynamic weight adjustment is employed. Exploration phase: The weights are extremely low, allowing the model to learn the layout rules of a two-dimensional plane first; Adsorption period: Gradually increases during model training. The weights; after the model completes the learning of the layout rules, it adjusts the position of the components based on the calculation results of the physical attachment loss function; through continuous adjustment of the model, the normalized normal distance from the center of the component to the wall plane is kept within 1.0 to 2.0, thereby achieving accurate prediction of the component position; Perform forced geometric calculations to eliminate tiny floating-point errors in the model's predictions during deep learning: The model obtains the plane equation of the parent wall, projects the center of the component onto the plane based on the plane equation of the parent wall, and aligns the normal direction to complete the selection of the primary component.

6. The method for generating refined architectural models based on spatial intelligence and graph semantics as described in claim 1, characterized in that, The steps for conflict removal are as follows: The 3D IoU between each primary component is calculated using the 3D NMS algorithm, and redundant components with overlapping positions are removed. The non-maximum suppression algorithm is used to remove overlapping redundancy and retain the primary component with the highest confidence. The semantic graph data corresponding to the primary components is parsed, and the corresponding components are generated in the building skeleton model to obtain the building model; The building model is formatted to the standard BIM format. The corresponding parametric family is called according to the predicted Type ID of the model. The family parameters are set according to the predicted size parameters to complete the three-dimensional detailed generation of the building facade.

Images