BIM-based multi-modal indoor design scheme generation method and system

Abstract

This invention relates to the field of interior architectural design technology, and discloses a method and system for generating multimodal interior design schemes based on BIM. The method includes: reading the initial spatial room BIM model, extracting boundary constraints and sampling the surface of electromechanical pipelines to construct a pipeline repulsion potential energy tensor field; acquiring natural language text and style reference images, extracting basic feature matrices and performing bidirectional cross-attention decoupling, and outputting appearance feature flow and structural feature flow; based on the above feature flow, using a graph neural network to generate a prefabricated semantic topology graph, using the pipeline repulsion potential energy tensor field as a global constraint to drive topological mutation of target edges with interference risk; inputting the prefabricated semantic topology graph into a constraint solving engine to solve for node coordinates to minimize the global cost function, and triggering a gradient back-plot pruning mechanism to update the prefabricated semantic topology graph when convergence is not achieved, until the solution converges and an interior prefabricated model is generated.

Description

Technical Field

[0001] This invention relates to the field of interior architectural design technology, specifically to a method and system for generating multimodal interior design schemes based on BIM. Background Technology

[0002] In interior architectural design scenarios, multimodal automated generation technology based on natural language and image input transforms text descriptions and visual references into three-dimensional spatial layouts. Conventional solutions typically collect multi-source input data, perform vector splicing and fusion, and then rely on feedforward neural networks to output the three-dimensional coordinates and dimensional parameters of prefabricated components to complete the initial scene construction.

[0003] Existing systems, when dealing with complex spatial physical environments, fail to incorporate existing electromechanical pipelines into a continuous avoidance mechanism. Conventional algorithms treat pipelines as static discrete collision bounding boxes, making it difficult for the generative network to perceive the interference risks in the physical space surrounding the pipelines when assigning connections and coordinates to prefabricated components. Spatial node connections directly penetrate physical pipelines, causing structural interference. Furthermore, during the multimodal data input stage, the front-end feature extraction network performs direct concatenation or summation operations on text sequences and image pixels. The material intent carried by the text sequence and the structural topological instructions contained in the image pixels undergo dimensionality aliasing within the same shared latent space. The modal fusion stage lacks independent processing branches, causing apparent attribute features to overwrite the semantics of physical assembly connections, resulting in material mismatch and loss of connection logic in the output model.

[0004] Furthermore, the forward-generated spatial layout framework lacks a closed loop for correcting underlying topology errors. When the initial component coordinates output by the algorithm result in wall out-of-bounds movement, door and window obstruction, or panel splicing gaps exceeding tolerance limits in the three-dimensional physical space, the solution engine can only output an invalid overlapping model. The underlying graph structure generation network and the end three-dimensional geometric coordinate verification module are in a one-way communication state, making it difficult for the system to backfeed the geometric deviation values ​​generated by spatial interference to the original topology network. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for generating multimodal interior design schemes based on BIM. This method solves the problems of overlapping multimodal features, lack of dynamic pipeline avoidance mechanisms, and inability to correct the underlying topology based on spatial physical errors in existing interior design schemes.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides a method for generating multimodal interior design schemes based on BIM, comprising the following steps:

[0008] Step S1: Read the initial spatial room BIM model, extract the spatial three-dimensional boundary constraints and sample the surface of electromechanical pipelines to construct the pipeline repulsion potential energy tensor field;

[0009] Step S2: Obtain natural language text and style reference image, extract their respective basic feature matrices, perform bidirectional cross-attention decoupling on the extracted basic feature matrices, and output appearance feature flow and structural feature flow respectively;

[0010] Step S3: Based on the apparent feature flow and the structural feature flow, a prefabricated semantic topology graph containing nodes, edges and their weights is generated using a graph neural network, and the pipeline repulsion potential energy tensor field is used as a global constraint to drive the topological mutation of the target edge that is determined to have interference risk.

[0011] Step S4: Input the prefabricated semantic topology graph into the constraint solving engine to solve the node coordinates in order to minimize the global cost function. If the solution does not converge, trigger the gradient back graph pruning mechanism to update the prefabricated semantic topology graph until the solution converges and the indoor assembly model is generated.

[0012] Furthermore, to transform the discrete initial spatial room BIM model into a continuous three-dimensional spatial mathematical environment, the specific implementation method for constructing the pipeline repulsion potential tensor field in step S1 includes:

[0013] A point cloud set is constructed by uniformly sampling the surface of the electromechanical pipeline and an octree spatial index is established; the shortest distance from the voxel point in the three-dimensional Euclidean space corresponding to the initial space room BIM model to the surface of the electromechanical pipeline is calculated to define an unsigned distance field function; the pipeline repulsion potential energy tensor field is constructed based on the distance field function, the preset safety avoidance threshold and the potential energy rise coefficient.

[0014] As the core innovative principle of this invention for quantifying the risk of intrusion in three-dimensional space, the mathematical relationship of the pipeline repulsion potential energy tensor field is defined as follows:

[0015] ;

[0016] The definitions and methods for obtaining the parameters in the formula are as follows: Represents the corresponding voxel point The numerical value of the pipeline repulsion potential energy tensor field at the location; Represents the coordinates of any voxel point within a three-dimensional Euclidean space; This represents the unsigned distance field function, whose value is calculated by measuring the coordinates of voxel points. The shortest Euclidean distance to the point cloud set on the surface of the electromechanical pipeline is obtained; This indicates the preset safe avoidance distance threshold, the value of which is retrieved based on historical assembly and construction experience data. This represents the potential energy rise rate coefficient, the value of which is preset based on the engineering repulsion strength requirements. Represented by natural constant An exponential function with base 0; This indicates the logic for extracting the maximum value among multiple values ​​within parentheses. Based on the above principle, when the shortest distance is greater than or equal to the safety avoidance threshold, the repulsive potential energy remains constant at zero; when the shortest distance is less than the safety avoidance threshold, the repulsive potential energy increases exponentially with the increase of intrusion depth.

[0017] Furthermore, to eliminate feature overlap contradictions during multimodal information fusion, the specific implementation process of step S2 includes: extracting the basic feature matrices of the natural language text and the style reference image using a pre-trained encoder and performing layer normalization; fusing the basic feature matrix of the natural language text as the query and the basic feature matrix of the style reference image as the key and value input using a multi-head cross-attention mechanism, and outputting the appearance feature stream by combining residual connections; and fusing the basic feature matrix of the style reference image as the query and the basic feature matrix of the natural language text as the key and value input using a multi-head cross-attention mechanism, and outputting the structural feature stream by combining residual connections.

[0018] Furthermore, in the physical parameter initialization stage of the graph node, the instantiation process for the node in step S3 includes: mapping the appearance feature flow to a node feature vector; calculating the cosine similarity between the node feature vector and the features of each real prefabricated BIM family in the BIM feature library; selecting the BIM family with the highest cosine similarity as the node instantiation object and inheriting its size and material parameters.

[0019] Furthermore, in the connection constraint stage of the graph structure, the specific process of topological mutation in step S3 includes: mapping the structural feature flow to initial semantic features; sampling the pipeline repulsion potential energy tensor field integrally along the expected connection path of the target edge in the prefabricated semantic topology graph to obtain the local maximum potential energy; when the local maximum potential energy is greater than zero, determining that the target edge has a physical interference risk, and forcibly updating the initial semantic features to the preset collision avoidance connection semantics without changing the inherent connection relationship of the graph.

[0020] Further, in step S4, the constraint solving engine establishes the global cost function using the spatial three-dimensional boundary constraints and the edges in the prefabricated semantic topology graph as constraints. The global cost function is composed of the sum of a first penalty term constraining the plate gap tolerance, a second penalty term constraining the support alignment error, a third penalty term constraining the wall boundary crossing, and a fourth penalty term constraining the door and window obstruction. The convergence criterion in step S4 is as follows: within a preset maximum number of iterations, when the calculated value of the global cost function is not greater than a preset engineering tolerance threshold, the node coordinate solution is determined to be converged, and the interior assembly model is generated based on the final node coordinate matrix and the parameters of the node instantiation.

[0021] Furthermore, to establish a closed-loop optimization channel between the underlying data structure and spatial coordinates, the specific operations of triggering the gradient back-end graph pruning mechanism in step S4 include: when the solution iteration reaches the preset maximum number of iterations and the calculated value of the global cost function is greater than the preset engineering tolerance threshold, it is determined that there is a node position conflict; a mask matrix is ​​generated, and the mask matrix elements corresponding to the nodes where the node position conflict occurs are set to one, and the mask matrix elements corresponding to the non-conflict regions are set to zero; the partial derivative matrix of the global cost function with respect to the node coordinates is calculated, and a Hadamard product operation is performed between it and the mask matrix to construct a local constraint repulsion tensor.

[0022] Furthermore, the process of updating the prefabricated semantic topology graph includes: feeding back the local constraint repulsion tensor as an additional penalty loss term to the graph neural network; the graph neural network updates the edge weights corresponding to the nodes where the node positions conflict by decaying accordingly; when the edge weights are lower than a preset truncation value, a pruning operation is triggered to reconstruct the conflicting edges, erase the connection relationships that cause size interference, generate a new prefabricated semantic topology graph, and re-input it into the constraint solving engine.

[0023] A second aspect of the present invention provides a BIM-based multimodal interior design scheme generation system, the system being configured to execute the BIM-based multimodal interior design scheme generation method described in the first aspect above, the system comprising:

[0024] The environmental analysis module is used to read the initial spatial room BIM model, extract the spatial three-dimensional boundary constraints, and sample the surface of electromechanical pipelines to construct the pipeline repulsion potential energy tensor field.

[0025] The intent decoupling module is used to acquire natural language text and style reference images, extract their respective basic feature matrices, perform bidirectional cross-attention decoupling on the extracted basic feature matrices, and output appearance feature flow and structural feature flow respectively.

[0026] The topology generation module is used to generate an assembled semantic topology graph containing nodes, edges and their weights using a graph neural network based on the apparent feature flow and the structural feature flow, and to drive the topological mutation of target edges that are determined to have interference risks using the pipeline repulsion potential energy tensor field as a global constraint.

[0027] The constraint solving and pruning module is used to input the prefabricated semantic topology graph into the constraint solving engine to solve the node coordinates in order to minimize the global cost function. When the solution fails to converge, the gradient back graph pruning mechanism is triggered to update the prefabricated semantic topology graph until the solution converges and an indoor prefabricated model is generated.

[0028] This invention provides a method and system for generating multimodal interior design schemes based on BIM. It has the following beneficial effects:

[0029] 1. This invention extracts samples from the pipeline surface to construct a repulsive potential energy tensor field. Combined with the generated prefabricated semantic topology graph, the tensor field is used as a constraint to drive the target edge to perform topological mutation, transforming the spatial avoidance boundary into a continuous field variable calculation. During the graph connection generation stage, the peeled component coincides with the pipeline and occupies the same physical coordinate condition.

[0030] 2. This invention extracts the basic feature matrices of text and images, performs bidirectional cross-attention decoupling, separates the output appearance and structural feature streams, inputs the graph neural network to generate a topology graph, strips material properties and spatial connection instructions along independent data branches, and blocks feature dimension confusion and parameter mapping overwriting that occur during the merging and operation of multimodal source data.

[0031] 3. In this invention, when the engine fails to converge in minimizing the global cost function, the gradient back-end graph pruning mechanism is triggered to update the assembled semantic topology graph. The partial derivatives are converted into penalty terms to reduce the weights of the node connections. The closed loop of the underlying connected network node graph relationship is established by cutting off the cross-boundary interference values ​​according to the three-dimensional spatial coordinates and reshaping them. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating the overall process of the BIM-based multimodal interior design scheme generation method of the present invention.

[0033] Figure 2 This is a block diagram of the architecture of the BIM-based multimodal interior design scheme generation system of the present invention;

[0034] Figure 3 This is a schematic diagram illustrating the principle of constructing the pipeline repulsion potential energy tensor field in this invention.

[0035] Figure 4 This is a schematic diagram of the bidirectional cross-attention decoupling network structure in this invention;

[0036] Figure 5This is a logical diagram illustrating the generation of prefabricated semantic topology graphs and topological mutations in this invention.

[0037] Figure 6 This is a flowchart of the gradient back-plot pruning and constraint solving closed-loop mechanism in this invention. Detailed Implementation

[0038] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Please see the appendix Figure 1 To be continued Figure 6 This invention provides a method and system for generating multimodal interior design schemes based on BIM. The method is applied to computer equipment with data processing capabilities, including a local computing workstation or a cloud computing server with a multi-core processor and a graphics processing unit.

[0040] The BIM-based multimodal interior design scheme generation method may include the following steps:

[0041] Step S1: Read the initial spatial room BIM model, extract the spatial three-dimensional boundary constraints and sample the surface of electromechanical pipelines to construct the pipeline repulsion potential energy tensor field;

[0042] Step S2: Obtain natural language text and style reference image, extract their respective basic feature matrices, perform bidirectional cross-attention decoupling on the extracted basic feature matrices, and output appearance feature flow and structural feature flow respectively;

[0043] Step S3: Based on the apparent feature flow and structural feature flow, a graph neural network is used to generate an assembled semantic topology graph containing nodes, edges and their weights. The pipeline repulsion potential energy tensor field is used as a global constraint to drive the topological mutation of the target edge that is determined to have interference risk.

[0044] Step S4: Input the prefabricated semantic topology graph into the constraint solving engine to solve the node coordinates in order to minimize the global cost function. If the solution does not converge, trigger the gradient back graph pruning mechanism to update the prefabricated semantic topology graph until the solution converges and the indoor prefabricated model is generated.

[0045] The computer equipment establishes a data exchange area in memory to receive initial data input by the user. This initial data includes an initial BIM model of the room space, natural language text, and style reference images. The natural language text carries data indicating structural design requirements and construction process constraints, while the style reference images carry data indicating color schemes and material textures. The computer equipment loads this initial data into the processor via an internal data bus to perform calculations, establishing a data mapping relationship between the unstructured design intent and the three-dimensional solid model.

[0046] Computer equipment reads the initial spatial room BIM model and extracts the set of spatial three-dimensional boundary constraints. And electromechanical pipeline model data. Spatial three-dimensional boundary constraint set. The three-dimensional geometric parameters, including wall boundaries, floor slab boundaries, and ceiling boundaries, are extracted from the original building's civil engineering survey data. Computer equipment uses algorithms to transform the discrete electromechanical pipeline model data into a continuous three-dimensional spatial mathematical environment, generating a pipeline repulsion potential energy tensor field. This tensor field provides a quantitative calculation basis for subsequent steps to assess the risk of any coordinate point in three-dimensional space intruding into the avoidance zone of concealed pipeline works.

[0047] The computer equipment invokes a multimodal pre-trained model to extract the basic feature matrices from the natural language text and style reference images, respectively. The multimodal pre-trained model executes a bidirectional cross-attention algorithm to separate the feature vectors that interfere with each other between modal data, outputting an appearance feature stream and a structural feature stream. The appearance feature stream carries information on the materials and colors of the interior decoration, while the structural feature stream carries information on the physical connection process of the components and the gaps between the seams.

[0048] The computer device inputs the apparent feature stream and structural feature stream into the graph neural network to construct an assembly-type semantic topology graph. The set of nodes in the graph For each solid BIM component in the corresponding indoor assembly environment, edge assembly This corresponds to the assembly connection logic between each entity's BIM components. At this stage, the computer equipment reads the pipeline repulsion potential energy tensor field as prior environmental data constraint input, along the edge set... The expected spatial connection paths of each edge are integrally sampled. When a collision interference risk is determined in a path region, the computer device forcibly performs a topological mutation operation, updates the connection semantic features of the corresponding edge, and avoids spatial intersection conflicts between physical entities in the assembly model from the graph structure data source.

[0049] The computer device will generate a complete assembly semantic topology diagram. Import into the constraint solving engine. The constraint solving engine constructs a global cost function for the 3D spatial location of all nodes. In this embodiment, the constraint solving engine is equipped with the Adam optimizer. Before initiating gradient descent, to prevent the optimization process from getting stuck in local optima, the computer device extracts the geometric center coordinates of each voxel corresponding to each node in the aforementioned three-dimensional voxel mesh, and uses them as a three-dimensional coordinate matrix. Initial iteration zero point The coordinate initialization is then completed. Subsequently, the computer device starts the Adam optimizer to execute the gradient descent optimization algorithm, solving for the partial derivative of the global cost function with respect to the node coordinates in each iteration, and strictly following... The reverse update rule adjusts the node positions until the rate of change of the global cost function loss value is lower than the preset convergence tolerance or the maximum number of iterations is reached, at which point the iteration terminates, and finally outputs the three-dimensional coordinate matrix of each node in three-dimensional Euclidean space without interference contradictions. When the constraint solving engine reaches the set maximum number of iterations and the global cost function fails to meet the engineering tolerance threshold, the computer determines that there is a contradiction in the underlying physical dimensions of the current prefabricated semantic topology graph.

[0050] When a discrepancy occurs in the underlying physical dimensions, the computer device obtains the cost function for the three-dimensional coordinate matrix. The partial derivative matrix is ​​used to generate a repulsive gradient tensor based on the location of conflicting nodes, and this tensor is then back-propagated to the aforementioned graph neural network. The graph neural network, based on the weight parameters of edges causing spatial conflicts due to the reduction of the repulsive gradient tensor, triggers a graph pruning mechanism to remove corresponding conflicting edges, generating an updated assembly semantic topology graph. This updated graph is then re-input into the constraint solving engine for iterative closed-loop optimization calculations. After the constraint solving engine confirms that the calculated global cost function meets the convergence requirements, the computer device uses the final three-dimensional coordinate matrix... The system uses the parameter information of each node to call the application programming interface of the BIM modeling software to automatically instantiate all node components, generate complete interior assembly model data, and output and save it.

[0051] In this embodiment, in order to convert the discrete initial spatial room BIM model into a continuous three-dimensional spatial mathematical environment, the computer device analyzes the underlying geometric data structure of the initial spatial room BIM model and separates the spatial three-dimensional boundary constraint set and electromechanical pipeline geometric entities from it.

[0052] The computer equipment performs a mesh subdivision operation on the outer surface of the electromechanical pipeline, obtaining a mesh surface containing multiple triangular facets. Then, it performs uniform point sampling on each triangular facet according to the equidistant sampling rule to generate a three-dimensional point cloud set representing the physical contour of the electromechanical pipeline. To accelerate the subsequent spatial traversal calculation process, the computer device establishes an initial bounding box containing the entire 3D point cloud in the 3D coordinate system, and recursively divides the bounding box space according to the set maximum depth level to generate an octree spatial index structure; the octree spatial index records the correspondence between each spatial block and the coordinates of the sampling points falling into it.

[0053] Computer equipment in the initial space room BIM model corresponding to the three-dimensional Euclidean space The interior is divided into a uniform three-dimensional voxel mesh with a resolution of 50 mm. This is for three-dimensional Euclidean space. coordinates of any voxel point within The computer device traverses the aforementioned octree spatial index to find the coordinates of the voxel point. Calculate the voxel coordinates of the nearest spatially nearest point cloud subset. The shortest Euclidean distance to the surface of electromechanical pipelines is used to define a continuous unsigned distance field function. The corresponding calculation formula is:

[0054] ;

[0055] In the formula, Represents a set of three-dimensional point clouds The coordinates of a single sampling point in the data; This represents the solution for the L2 norm between two three-dimensional coordinate points; This indicates finding the minimum value among the traversal calculation results. This unsigned distance field function maps the absolute physical distance from the coordinates of any point in the space to the nearest pipeline surface.

[0056] The computer equipment retrieves preset safe avoidance distance thresholds based on historical assembly and construction experience data. (In this embodiment, the range is set to 150mm to 300mm) and the potential energy rise rate coefficient. (In this embodiment, it is set between 0.5 and 1.2). Based on the unsigned distance field function Safe avoidance distance threshold and the coefficient of potential energy rise rate Computer equipment construction pipeline repulsive potential energy tensor field Its mathematical relationship is:

[0057] ;

[0058] In the formula, Represented by natural constant An exponential function with base 0; This indicates the logic for extracting the maximum value from multiple values ​​within parentheses.

[0059] The computer device establishes a globally continuous field tensor based on this relational operation, when the unsigned distance field function... The shortest distance obtained is greater than or equal to the safe avoidance distance threshold. When, it represents the coordinates of the corresponding voxel point. In the interference-free region, at this time The extraction result is zero, making the calculated repulsive potential energy constant at zero; when the shortest distance is less than the safe avoidance distance threshold. When, it represents the coordinates of the corresponding voxel point. Intrusion into the pipeline safety zone, at this time This produces a difference greater than zero, which is related to the potential energy rise rate coefficient. The product is then used as an input to the exponential term, causing the repulsive potential energy to increase exponentially with the increase of penetration depth.

[0060] In this embodiment, the computer device reads natural language text. and style reference images The computer device accesses the internally stored text pre-trained encoder based on the RoBERTa architecture. Processing natural language text The output is a basic feature matrix with a feature dimension of 768. Simultaneously, it invokes an image pre-trained encoder based on the ResNet-50 architecture. Processing style reference image After being flattened by the internal global average pooling layer, the output is a basic feature matrix with the same 768 dimensions. The formula for their relationship is:

[0061] ;

[0062] ;

[0063] To separate the material properties of entities and the spatial connectivity and process attributes that coexist in multimodal data, computer equipment extracts the basic feature matrix. With the fundamental feature matrix Perform bidirectional cross-attention decoupling operations. The computer device calls the layer normalization function. For the fundamental feature matrix With the fundamental feature matrix Perform mean and variance calculations to standardize the feature distribution and avoid gradient divergence in the calculation.

[0064] The computer equipment is configured with a first computational branch to extract the apparent feature flow. This is achieved by calling the internally initialized first set of learnable linear projection weight matrices, which then transforms the basic feature matrix... Mapped to a multi-head cross-attention mechanism query matrix and the basic feature matrix These are respectively mapped to the key matrix of this multi-head cross-attention mechanism. Sum matrix The computer device will query the matrix. Key matrix Sum matrix The multi-head cross-attention mechanism performs dot product and weight allocation operations, combined with the basic feature matrix. The residual connection output apparent feature stream The corresponding calculation formula is:

[0065] ;

[0066] In the formula, This represents the preset scaling weight for the appearance attention operation (initially 0.5 in this embodiment). The apparent residual connection scaling weights are predefined (initial value is 0.5 in this embodiment), representing the apparent feature flow. Visual material and color attribute data of the solid components used in interior assembly.

[0067] The computer equipment is configured with a second computational branch to extract structural feature flow. This is achieved by calling an internally initialized second set of learnable linear projection weight matrices, which then transforms the basic feature matrix... Mapped to a multi-head cross-attention mechanism query matrix and the basic feature matrix Mapped to the corresponding key matrices Sum matrix The computer equipment uses the same characteristic fusion operation logic to set the query matrix for the current computing branch. Key matrix Sum matrix The input is processed by a multi-head cross-attention mechanism, combined with the basic feature matrix. residual connection output structure feature flow The corresponding calculation formula is:

[0068] ;

[0069] In the formula, This represents the preset scaling weight for the structural attention operation (in this embodiment, the initial value is 0.7). The structural residual connection scaling weights (initial value is 0.3 in this embodiment) represent the structural feature flow. Physical connection technology and splicing gap constraint data for supporting indoor assembled solid components.

[0070] In this embodiment, the computer device outputs the structural feature stream based on the aforementioned steps. The number of independent entities in the initial assembly semantic topology graph. Node set For the initial set of edges between nodes. The computer equipment first uses the aforementioned extracted set of three-dimensional spatial boundary constraints. The bounding box of the effective space is used to assign three-dimensional coordinates to each node within the effective space using a uniform random sampling algorithm, which are then used as the initial estimated coordinates of the nodes. Subsequently, the computer equipment employs... The K-Nearest Neighbor algorithm calculates the Euclidean distance between nodes in a 3D Euclidean space based on their initial estimated coordinates, and establishes a nearest neighbor relationship for each node. 1 node (in this embodiment) Let the initial undirected connected edges be 5). After the initial graph topology is established, the computer device calls the internally deployed graph neural network for processing. In this embodiment, the graph neural network specifically adopts a 3-layer graph attention network. The graph attention network pre-deploys a multilayer perceptron unit, which is specifically composed of three sequentially connected linear fully connected layers and alternating ReLU activation functions. This is for the assembled semantic topology graph. any node within The computer device reduces the apparent feature stream input to the multilayer perceptron unit by dimensionality reduction and mapping it into a 256-dimensional node feature vector. Subsequently, in each layer of the graph attention network, the computer device uses attention coefficients to perform weighted summation and message passing on the feature vectors of neighboring nodes, thereby fusing local topological information.

[0071] The computer device accesses the BIM feature library data stored in memory. This BIM feature library has been pre-processed by an offline autoencoder with the same structure as the aforementioned multimodal pre-trained encoder, uniformly compressing and encoding the parameters of the real physical model, including doors, windows, wall panels, and pipelines, to generate and store real prefabricated BIM family features with multiple dimensions strictly locked at 256×1. The corresponding mathematical formula is:

[0072] ;

[0073] In the formula, The numerator represents the cosine similarity function; the denominator represents the result of the vector dot product operation; The result of the L2 norm operation on the representative vector. Based on the same feature dimension, the computer device establishes a 256-dimensional node feature vector. Features of each real prefabricated BIM family The logic for the cosine similarity dot product operation is as follows: The computer device iterates through all cosine similarity values ​​and outputs the feature of the actual prefabricated BIM family with the largest value. Established as the current node The instantiation target is defined. The computer device parses the metadata dictionary bound to the instantiated target, extracting its spatial dimensions (length, width, and thickness), material texture encoding, and physical density value. Then, it writes these parameters into the node using a key-value pair matching method. The parameters are assigned in the internal attribute table, thus completing the physical parameter initialization process of the graph node.

[0074] Computer device processing edge set Edges connecting each node in the graph Computer equipment will stream structural features. Mapped to edges initial semantic features Then, the pipeline repulsion potential energy tensor field established by the aforementioned steps is retrieved. As global spatial constraint data. Regarding edges... The computer device extracts the center coordinates of the two pointed-to node entities and calls the Bresenham line generation algorithm in three-dimensional Euclidean space. Internally generate a set of discrete 3D coordinate points connecting the two central coordinates, thereby establishing the expected connection path between them. The computer device follows the expected connection path. Pipeline repulsive potential tensor field Perform discrete point integral sampling to extract all potential energy values ​​within the path range, output and cache the local maximum potential energy. .

[0075] Computer devices load dynamic topology mutation function The logic for handling spatial interference is as follows. In this invention, semantic features are specifically defined as a joint concatenated vector containing the component's physical connection type, spatial relative position vector, and interference collision penalty weight; feature mutation aims to forcibly overwrite the aforementioned joint concatenated vector on connected edges without changing the inherent connection state of graph nodes. For edges... Its initial edge semantic features It is a column vector with dimensions 256×1. When the local maximum potential energy... When the value is greater than zero, the computer device determines the expected connection path. Intrusion into the mechanical and electrical pipeline avoidance area, target edge There is a risk of physical assembly collision interference. The computer device's memory pre-constructs an avoidance semantic mapping dictionary, which classifies interference risk levels based on potential energy ranges and stores a space offset transformation matrix. The matrix is ​​a non-singular square matrix with dimensions of 256×256, thus ensuring dimension alignment for matrix multiplication: when the local maximum potential energy is in the first interval (0,10], the dictionary mapping corresponds to a spatial offset transformation matrix with a 50mm normal offset; when it is in the second interval (10,50], the dictionary mapping corresponds to a spatial offset transformation matrix with a 150mm lateral deviation. The computer device extracts the local maximum potential energy from the aforementioned calculation output, compares it with the above numerical intervals, and extracts the corresponding spatial offset transformation matrix from the dictionary. Then, a matrix multiplication operation is performed between this and the initial edge semantic features, outputting updated edge semantic features of the same 256×1 dimension. The corresponding formula is:

[0076] ;

[0077] By performing this mutation calculation, the computer equipment forcibly injects pre-defined collision avoidance spatial transformation data (such as increasing the distance of suspended supports or performing lateral offset detours) into the feature vector dimension of the connection semantics containing interference risk data, eliminating physical interference contradictions between BIM components from the underlying data structure stage. This updates the edge semantic features. It carries spatial assembly constraint rules such as floating supports or offset detours. Based on these rules, computer equipment eliminates physical interference inconsistencies between BIM components at the underlying data structure stage, generating an assembly semantic topology map with anti-collision constraints. It is then output to the subsequent constraint solution data channel.

[0078] In this embodiment, the computer device inputs the aforementioned generated prefabricated semantic topology graph into an internally deployed constraint solving engine, and establishes a global cost function using spatial three-dimensional boundary constraints and edges in the prefabricated semantic topology graph as constraints. The computer device then uses this constraint solving engine to calculate the three-dimensional coordinate matrix of each entity BIM component within the graph node set. ,in Representing the The coordinate vector data of each node in three-dimensional Euclidean space.

[0079] Global cost function set by computer equipment It consists of the sum of four penalty terms, and the corresponding mathematical formula is:

[0080] ;

[0081] In the formula, Represents each associated edge in the assembly semantic topology graph; This represents each node in the assembled semantic topology graph. The computer device defines each penalty term according to the following mathematical logic: The first penalty term constraining the tolerance of the gap between sheet metal sections is calculated using the following formula: ,in This is a preset standard splicing distance vector for adjacent components in the BIM family library. This item is used to quantify the Euclidean distance squared error between the actual assembly coordinates and the ideal coordinates. The second penalty term for constraining the alignment error of the support component is calculated using the vector connecting the coordinates of the two nodes and the spatial standard axial vector. The cross product 2-norm is calculated using the following formula: This is used to quantify the axial deviation of the spatial support skeleton; To constrain the third penalty term for wall encroachment, a smooth ReLU function is used to calculate the depth of node coordinates intruding into the wall boundary, as shown in the formula: ,in To preset the safe thickness from the wall, coordinates The shortest Euclidean distance to the nearest wall boundary; To constrain the fourth penalty item for obstructing doors and windows, the center coordinates of the reserved door and window openings are used. and repulsion radius Construct the Gaussian penalty formula This is to prevent the component coordinates from falling into the reserved opening area.

[0082] The constraint solving engine iteratively updates the 3D coordinate matrix using the gradient descent algorithm. To find the minimum global cost function The coordinate solution at time. The computer equipment retrieves the preset engineering tolerance threshold based on the construction tolerance standards. (In this embodiment, the value is 5.0 mm) and the maximum number of iterations. (In this embodiment, the number of iterations is set to 5000). During the iterative calculation process, when the current iteration step is not greater than the maximum number of iterations... Furthermore, the calculated value of the global cost function is not greater than the engineering tolerance threshold. At this point, the computer determines that the node coordinate solution process has reached a convergence state. The computer extracts the final node coordinate matrix at this time, combines it with the material and dimension parameters of the instantiated entities of each node, and calls the program interface of the BIM application software to generate an interior assembly model in the spatial coordinate system.

[0083] When the number of iterations reaches the maximum number of iterations. And the global cost function The calculated value is still greater than the engineering tolerance threshold. At this point, the computer device determines that there is a node position conflict in the underlying data of the current prefabricated semantic topology graph. The computer device triggers the gradient backpropagation graph pruning mechanism, extracts the node coordinate pairs that cause the global cost function to fail to converge, and constructs a dimension-1 in memory. (in A mask matrix (total number of nodes) This is done to match the data structure of the three-dimensional coordinate matrix. The computer device sets the values ​​of the row and column intersection matrix elements corresponding to the node pairs with conflicting node positions to one, while setting the values ​​of the matrix elements in the remaining non-interference areas to zero.

[0084] Computer devices for global cost function Solve for the three-dimensional coordinate matrix The partial derivatives of the coordinates of each node are used to obtain the directional gradient of each node. This is used to match the mask matrix. of Dimension, any two nodes in a computer device computation graph. and The relative repulsive gradient magnitudes between them are constructed with dimension 1. Pairwise gradient matrices , where matrix elements Computer devices will pair gradient matrices. The same as the dimensions established above. mask matrix The input tensor computation logic unit performs the Hadamard product operation to construct a local constraint repulsion tensor. Its mathematical relationship is:

[0085] ;

[0086] In the formula, The symbol represents the Hadamard product logic, which involves element-wise multiplication of corresponding matrices. This local constraint excludes the tensor. Filter out gradient data in conflict-free regions and extract and retain only the absolute gradient strength between specific node pairs that cause interference conflicts.

[0087] The computer device excludes the local constraint tensor. The additional penalty loss term is fed back to the graph neural network that generates the assembled semantic topology graph. The graph neural network receives the additional penalty loss term and locates the specific graph structure region where node position conflicts occur. The computer device then extracts the target edge within the conflict region at the current iteration step. The edge weight parameters below And based on local constraints, exclude tensors The gradient value at the corresponding coordinate position is calculated using a preset optimization learning rate. (In this embodiment, the value is 0.01) Perform gradient descent decay update operation, and the calculation formula is as follows:

[0088] ;

[0089] In the formula, Represents the edge weight for the next iteration. This indicates that a norm is taken for the corresponding elements of the repulsion tensor to extract the absolute gradient strength. The computer continuously evaluates the updated associated edge weights, and when determining the weight of a certain edge... Edges below the preset threshold have a cutoff value. When the value is 0.15 in this embodiment, a pruning operation is triggered to remove the edge from the graph data structure, thereby breaking the connection between the original graph nodes that have size interference and reconstructing the topology. The computer device generates a new assembled semantic topology graph based on the pruned and reconstructed node and edge data, and re-inputs it into the constraint solving engine to initiate a new round of coordinate solving iterations, eliminating the data barrier between the topology and the three-dimensional coordinates, until the global cost function meets the convergence criteria.

[0090] In this embodiment, a BIM-based multimodal interior design scheme generation system is provided. A computer device runs this system to execute the aforementioned logical steps. The system's logical data architecture is divided into an environment parsing module, an intent decoupling module, a topology generation module, and a constraint solving and pruning module.

[0091] The environment analysis module is used to read the initial spatial room BIM model and extract spatial three-dimensional boundary constraint data. This module performs uniform point sampling on the surface of electromechanical pipelines, calculates and generates a spatial point cloud set and an octree spatial index, and constructs a pipeline repulsion potential energy tensor field by combining an unsigned distance field function and a preset safety avoidance threshold, outputting spatial collision interference prior constraint data covering three-dimensional Euclidean space.

[0092] The intent decoupling module acquires natural language text and style reference images, and calls a pre-trained network structure to extract the basic feature matrices of both. This module performs layer normalization and bidirectional cross-attention fusion operations on these basic feature matrices, outputting appearance feature streams and structural feature streams, thus removing feature overlap and inconsistencies during multimodal information fusion from the data's underlying layers.

[0093] The topology generation module is used to establish the mapping relationship between the apparent feature flow and the node vector, and outputs the assembly semantic topology graph using a graph neural network. This module reads the repulsive potential energy tensor field of the aforementioned pipeline, performs spatial integral sampling along the expected connection path of the component assembly to obtain the local maximum potential energy; when the local maximum potential energy is greater than zero, this module forcibly updates the initial semantic features between nodes, drives the interference target edge to complete the topological mutation operation to generate collision-avoiding connection semantics.

[0094] The constraint solving and pruning module is used to establish a global cost function containing four independent penalty terms. It performs gradient descent optimization based on the prefabricated semantic topology graph data to obtain node coordinates. When the number of iterations reaches the maximum set value and the global cost function has not converged, this module generates a Hadamard product tensor of the partial derivative matrix and the mask matrix. This tensor is then backpropagated to attenuate the edge weights corresponding to spatial conflicts, triggering a graph pruning mechanism to reconstruct the topology graph. After the data iterations have fully converged, the underlying interface is called to generate the interior assembly model.

[0095] In this invention, the hardware computer device supporting the aforementioned system operation includes a central processing unit (CPU) for main control logic scheduling, a graphics processing unit (GPU) specifically for accelerating tensor operations in large-scale graph neural networks, random access memory (RAM), and a high-speed network communication interface. The CPU and GPU communicate via a PCI-E high-speed bus to exchange internal electrical signals and transmit underlying feature matrices. The RAM not only stores computer program instructions but also pre-loads complete BIM feature library entity information and local constraint exclusion tensor matrices generated in each iterative calculation step. The network communication interface receives initial multimodal design intent data via data streams and simultaneously outputs the final parsed model file to an external BIM application platform.

[0096] As the execution entity of a computer device, the processor reads computer program instructions stored in memory. By executing these program instructions, the processor sequentially schedules the operation of the aforementioned functional modules, handling processes such as spatial 3D boundary resolution, multi-head cross-attention decoupling calculation, graph neural network topology inference, and tensor gradient backpropagation calculation, thus connecting the data conversion link from unstructured design intent to 3D solid component models.

[0097] This embodiment also relates to a non-volatile computer-readable storage medium, which may specifically include conventional storage devices such as read-only memory, programmable read-only memory, electrically erasable programmable read-only memory, flash memory, or solid-state drive. The computer-readable storage medium contains computer program instructions. When the processor in the aforementioned computer device reads and executes these computer program instructions, it completes all the computational processing steps in the above-described BIM-based multimodal interior design scheme generation method, solving the technical problem of the difficulty in automatically eliminating physical pipeline conflicts and cumulative dimensional errors in traditional prefabricated modeling operations from the underlying data architecture level.

Claims

1. A BIM-based multimodal interior design scheme generation method, characterized in that, include: Step S1: Read the initial spatial room BIM model, extract the spatial three-dimensional boundary constraints and sample the surface of electromechanical pipelines to construct the pipeline repulsion potential energy tensor field; Step S2: Obtain natural language text and style reference image, extract their respective basic feature matrices, perform bidirectional cross-attention decoupling on the extracted basic feature matrices, and output appearance feature flow and structural feature flow respectively; Step S3: Based on the apparent feature flow and the structural feature flow, a prefabricated semantic topology graph containing nodes, edges and their weights is generated using a graph neural network, and the pipeline repulsion potential energy tensor field is used as a global constraint to drive the topological mutation of the target edge that is determined to have interference risk. Step S4: Input the prefabricated semantic topology graph into the constraint solving engine to solve the node coordinates in order to minimize the global cost function. If the solution does not converge, trigger the gradient back graph pruning mechanism to update the prefabricated semantic topology graph until the solution converges and the indoor assembly model is generated.

2. The BIM-based multimodal interior design scheme generation method according to claim 1, characterized in that, The construction of the pipeline repulsive potential energy tensor field in step S1 includes: The surface of the electromechanical pipeline is uniformly sampled to construct a point cloud set and an octree spatial index is established; Calculate the shortest distance from the voxel point in the three-dimensional Euclidean space corresponding to the initial space room BIM model to the surface of the electromechanical pipeline, so as to define the unsigned distance field function; The pipeline repulsion potential energy tensor field is constructed based on the distance field function, the preset safety avoidance threshold, and the potential energy rise coefficient; wherein, when the shortest distance is greater than or equal to the safety avoidance threshold, the repulsion potential energy is zero, and when the shortest distance is less than the safety avoidance threshold, the repulsion potential energy rises exponentially.

3. The method for generating multimodal interior design schemes based on BIM according to claim 1, characterized in that, The specific process of step S2 includes: The basic feature matrices of the natural language text and the style reference image are extracted using a pre-trained encoder and then layer normalized. The system uses the basic feature matrix of the natural language text as the query and the basic feature matrix of the style reference image as the key and value input to fuse the multi-head cross-attention mechanism, and combines residual connections to output the appearance feature stream. The structural feature stream is output by fusing the basic feature matrix of the style reference image as the query and the basic feature matrix of the natural language text as the key and value input using a multi-head cross-attention mechanism, combined with residual connections.

4. The method for generating multimodal interior design schemes based on BIM according to claim 1, characterized in that, The instantiation process for the node in step S3 includes: The apparent feature flow is mapped to a node feature vector; the cosine similarity between the node feature vector and the features of each real prefabricated BIM family in the BIM feature library is calculated; the BIM family with the highest cosine similarity is selected as the node instantiation object and its size and material parameters are inherited.

5. The method for generating multimodal interior design schemes based on BIM according to claim 1, characterized in that, The specific process of topological mutation in step S3 includes: The structural feature flow is mapped to initial semantic features; the pipeline repulsion potential energy tensor field is integrally sampled along the expected connection path of the target edge in the prefabricated semantic topology graph to obtain the local maximum potential energy; when the local maximum potential energy is greater than zero, it is determined that the target edge has an interference risk, and the initial semantic features are forcibly updated to the preset collision avoidance connection semantics.

6. The method for generating multimodal interior design schemes based on BIM according to claim 1, characterized in that, In step S4, the constraint solving engine establishes the global cost function using the spatial three-dimensional boundary constraints and the edges in the prefabricated semantic topology graph as constraints. The global cost function is composed of the sum of the first penalty term constraining the plate gap tolerance, the second penalty term constraining the support alignment error, the third penalty term constraining the wall overstepping the boundary, and the fourth penalty term constraining the door and window occlusion.

7. The BIM-based multimodal interior design scheme generation method according to claim 6, characterized in that, The convergence criterion in step S4 is as follows: Within the preset maximum number of iterations, when the calculated value of the global cost function is not greater than the preset engineering tolerance threshold, the node coordinate solution is determined to be converged, and the indoor assembly model is generated based on the final node coordinate matrix and the parameters of the node instantiation.

8. The method for generating multimodal interior design schemes based on BIM according to claim 1, characterized in that, The step S4 that triggers the gradient back graph pruning mechanism includes: When the solution iteration reaches the preset maximum number of iterations and the calculated value of the global cost function is greater than the preset engineering tolerance threshold, it is determined that there is a node position conflict. Generate a mask matrix, set the mask matrix elements corresponding to nodes that have conflicting node positions to one, and set the mask matrix elements corresponding to non-conflicting regions to zero. Calculate the partial derivative matrix of the global cost function with respect to the node coordinates, and perform a Hadamard product operation with the mask matrix to construct a local constraint repulsion tensor.

9. The method for generating multimodal interior design schemes based on BIM according to claim 8, characterized in that, The feedback update of the assembled semantic topology graph includes: The local constraint exclusion tensor is fed back into the graph neural network as an additional penalty loss term. The graph neural network updates the edge weights corresponding to nodes where node positions conflict by decaying them; when the edge weights are lower than a preset truncation value, a pruning operation is triggered to reconstruct the conflicting edges, generate a new assembled semantic topology graph, and re-input it into the constraint solving engine.

10. A BIM-based multimodal interior design scheme generation system, characterized in that, The system is configured to perform the BIM-based multimodal interior design scheme generation method as described in any one of claims 1 to 9, the system comprising: The environmental analysis module is used to read the initial spatial room BIM model, extract the spatial three-dimensional boundary constraints, and sample the surface of electromechanical pipelines to construct the pipeline repulsion potential energy tensor field. The intent decoupling module is used to acquire natural language text and style reference images, extract their respective basic feature matrices, perform bidirectional cross-attention decoupling on the extracted basic feature matrices, and output appearance feature flow and structural feature flow respectively. The topology generation module is used to generate an assembled semantic topology graph containing nodes, edges and their weights using a graph neural network based on the apparent feature flow and the structural feature flow, and to drive the topological mutation of target edges that are determined to have interference risks using the pipeline repulsion potential energy tensor field as a global constraint. The constraint solving and pruning module is used to input the prefabricated semantic topology graph into the constraint solving engine to solve the node coordinates in order to minimize the global cost function. When the solution fails to converge, the gradient back graph pruning mechanism is triggered to update the prefabricated semantic topology graph until the solution converges and an indoor prefabricated model is generated.

Images