A pre-training artificial intelligence method for automatically generating BIM, an electronic device and a storage medium

Abstract

The application discloses a kind of pre-training artificial intelligence methods for automatically generating BIM, electronic equipment and storage medium, belong to the field of building intelligentization.Extract five kinds of real features and spatial features of building components from BIM to construct BIM graph, after encoding by graph neural network, residual quantization is carried out on the encoding representation, the real features of the encoding representation after residual quantization are restored by reverse addition and input into residual network decoder, and the token sequence representation of the building is obtained therefrom.The token sequence of the building is input into the pre-training network model to learn the component arrangement in the building.The pre-training network model obtained generates building description sequence, and the real features decoded by the decoder are completed by BIM modeling script program.The application significantly shortens the BIM automatic generation time, is easy to expand for continuous updating and learning, the generated BIM can be displayed on multiple platforms for personnel to view and modify, and has broad application prospects.

Description

Technical Field

[0001] This invention relates to a pre-trained artificial intelligence method, electronic device, and storage medium for automatically generating BIM, belonging to the field of building intelligence. Background Technology

[0002] Building Information Modeling (BIM) technology has brought about significant changes to the entire lifecycle of the construction industry, improving design and construction efficiency and reducing construction and operation costs. However, automated BIM modeling remains a challenging problem to be solved.

[0003] Currently, traditional BIM modeling primarily relies on manual operation using specialized software (such as Revit and ArchiCAD), supplemented by design drawings, specifications, and 2D CAD software as auxiliary tools. The modeling process requires manual input of geometric shapes and attribute information, as well as integration of models from multiple disciplines. This is not only inefficient and prone to errors, but also leads to complex and cumbersome collaborative workflows. Automating these repetitive and tedious modeling processes would significantly shorten BIM modeling time, allowing engineers to focus more on other critical architectural design tasks.

[0004] Therefore, this invention aims to propose a feasible solution for automatically generating BIM (Building Information Model) to solve the problems of low efficiency, error-proneness, and complex collaboration in the traditional BIM modeling process, while providing support for intelligent building design. Summary of the Invention

[0005] Objective: This invention designs a pre-trained artificial intelligence method, electronic device, and storage medium for automatically generating BIM (Building Information Modeling). It consists of three parts: building information extraction, network model training, and BIM modeling. This invention can not only learn the layout of building components for architectural design but also automatically complete BIM modeling based on predicted building description sequences. It transforms traditional manual BIM modeling into a fully automated process. For building component features that are difficult to describe (both real and spatial features), it transforms them into trainable token vocabulary, thereby achieving the learning of architectural styles and the generation of description sequences. This method significantly improves modeling efficiency, reduces errors and time costs associated with manual operation, and provides corresponding assistance for architectural design.

[0006] Technical solution: To solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0007] Firstly, a pre-trained artificial intelligence method for automatically generating BIM images, specifically including:

[0008] Clean up the BIM model and extract relevant building component information, including the actual characteristics and spatial characteristics of the building components.

[0009] Discretize the real features of the building components to obtain the discretized real features.

[0010] Graph data is constructed based on the discretized real features and the spatial features of building components. The graph data is then input into a graph neural network, a residual quantization module, and a decoder that are connected in sequence. The output is a vocabulary of building component descriptions, as well as the trained graph neural network, residual quantization module, and decoder.

[0011] Based on relevant building component information and a glossary of building component descriptions, a descriptive sequence for the building components is constructed.

[0012] The pre-trained model is trained using the description sequence of building components to obtain the well-trained pre-trained model.

[0013] The system obtains information about the building components to be generated, and uses a trained graph neural network and residual quantization module to obtain a word sequence of the building components to be generated. The word sequence of the building components to be generated is then input into a trained pre-trained model, which outputs a description sequence of the building components to be generated.

[0014] The description sequence of the building component to be generated is quantized and deep summed, then input into the trained decoder, and the output is the real features of the building component to be generated.

[0015] Generate a BIM model based on the actual characteristics of the building components to be generated.

[0016] Optionally, the actual characteristics of the building components include: the type, orientation, size, and location of columns, beams, walls, doors, and windows.

[0017] The spatial characteristics of the building components include the adjacent relationships of columns, beams, walls, doors, and windows.

[0018] Optionally, the discretization operation specifically includes:

[0019] Using the minimum difference in the dimensions of building components as the resolution, the positions of the building components are adjusted to the positive range in the three-dimensional coordinate system. Based on the resolution and the adjusted positions, the discretized real features of the building components are constructed.

[0020] Optionally, the graph neural network includes three sequentially connected SAGEConv convolutional layers.

[0021] The quantization depth in the residual quantization module is set to 3.

[0022] The decoder comprises five layers of residual networks connected in sequence.

[0023] Optionally, the step of constructing a description sequence of building components based on relevant building component information and a building component description vocabulary specifically includes:

[0024] Retrieve the corresponding terms from the glossary of building component descriptions based on the relevant building component information.

[0025] The words are arranged in the order of pillars, beams, walls, doors, and windows. Words of the same type are arranged in ascending order of their coordinate values ​​on the Z, Y, and X axes of a three-dimensional Cartesian coordinate system.

[0026] Then, add the starting word and the ending word to the first and last positions of the sequence respectively to obtain the descriptive sequence of the building components.

[0027] Optionally, the pre-trained model adopts a Transformer-OnlyDecoder model with a 3-layers-4-heads attention mechanism.

[0028] Optionally, the step of obtaining the information of the building components to be generated, and obtaining the vocabulary sequence of the building components to be generated through a trained graph neural network and a residual quantization module, specifically includes:

[0029] The information of the building components to be generated is obtained by taking the real features of some building components, inputting the real features of some building components into the trained graph neural network and residual quantization module, and outputting the word sequence of the building components to be generated.

[0030] The information of the building components to be generated is the starting word, and the starting word is used as the word sequence of the building components to be generated.

[0031] Optionally, generating a BIM model based on the actual characteristics of the building component to be generated specifically includes:

[0032] Create IFC entities corresponding to columns, beams, walls, doors, and windows, and obtain the category, orientation, size, and location based on the actual characteristics of the building components to be generated. Generate the corresponding IFC statements for the IFC entities to define the geometric and spatial characteristics of the building components, thus completing the BIM modeling.

[0033] The Z-axis positions of columns and walls are divided and statistically analyzed to filter out intervals with more than a threshold number of data points. The number of floors and heights are then determined based on the number of intervals and the minimum value within each interval, thus obtaining information about each floor in the BIM model.

[0034] In a second aspect, a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a pre-trained artificial intelligence method for automatically generating BIM as described in any of the first aspects.

[0035] Thirdly, a computer device comprising:

[0036] Memory is used to store instructions.

[0037] A processor for executing the instructions, causing the computer device to perform operations of a pre-trained artificial intelligence method for automatically generating BIM as described in any of the first aspects.

[0038] Beneficial effects: This invention provides a pre-trained artificial intelligence method, electronic device, and storage medium for automatically generating BIM models. Through full-process automation, it significantly improves modeling efficiency and reduces errors and time costs associated with manual operation. It generates preliminary BIM designs based on predicted building description sequences, helping designers quickly optimize their designs. Furthermore, it can continuously improve and adapt to more BIM modeling needs as data accumulates and is optimized. The method of this invention is robust, concise, effective, and has broad application prospects. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating a pre-trained artificial intelligence method for automatically generating BIM according to the present invention.

[0040] Figure 2 This is a schematic diagram illustrating the BIM information extraction and result visualization of the present invention.

[0041] Figure 3 This is a schematic diagram of the spatial feature determination algorithm in this invention.

[0042] Figure 4 This is a schematic diagram of the floor division algorithm of the present invention.

[0043] Figure 5 This is a diagram showing the input patterns of the pre-trained model for two generation tasks in this invention.

[0044] Figure 6 This is a schematic diagram of the BIM model generated by the present invention. Detailed Implementation

[0045] The technical solutions 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 protection scope of the present invention.

[0046] The present invention will be further described below with reference to specific embodiments.

[0047] Example 1:

[0048] This embodiment introduces a pre-trained artificial intelligence method for automatically generating BIM (Building Information Modeling). This method encompasses BIM information processing, network model training, and an automated BIM modeling program, integrating these three elements into a continuously learning automated BIM modeling system. The method is as follows:

[0049] Step 1: Clean up the acquired BIM model and extract relevant building component information, which includes the actual characteristics (category, orientation, size, location) and spatial characteristics (adjacent relationships between components) of the building components.

[0050] This step is used to obtain the five types of building components that this method focuses on, which helps to obtain a more concise IFC file and improve the efficiency of building component information extraction.

[0051] Furthermore, step 1 specifically includes:

[0052] Step 1.1: Obtain the true characteristics of building components:

[0053] Clean up the building components that are not considered in the BIM model to obtain the cleaned BIM model.

[0054] The cleaned BIM model is exported as the corresponding IFC (Industry Foundation Classes) file format, and the actual features of the components are extracted according to the semantic rules of IFC.

[0055] The step of extracting the true features of components according to the semantic rules of IFC includes accessing the IfcColumn, IfcBeam, IfcWallStandardCase, IfcDoor, and IfcWindow entities in the file through a designed true feature extraction program, and obtaining relevant information about columns, beams, walls, doors, and windows respectively. For columns and walls, dimensions are obtained by traversing their dimensioning statements (such as IfcRectangleProfileDef, IfcArbitraryClosedProfileDef, IfcPolyLine, IfcClosedShell), and the position and orientation are determined based on the IfcAxis2Placement3D statement. For beams, dimensions are obtained through dimensioning statements (such as IfcRectangleProfileDef, IfcConnectedFaceSet), and the position and orientation are derived in reverse using the IfcAxis2Placement3D statement. For doors and windows, dimensions are obtained through OverallWidth, OverallHeight, IfcPolyLine, IfcClosedShell statements, and the position and orientation are determined using the IfcAxis2Placement3D statement, resulting in a CSV file of the actual features of the building components.

[0056] Step 1.2: Obtain the spatial characteristics between building components:

[0057] Access BIM model information via API and implement external command (IExternalCommand) to obtain the boundary box of building components and add boundary box expansion value. ,in, ( (This is a manually set bounding box expansion value) to expand the bounding box, and the overlap of the expanded bounding box is used to determine whether... (Volumn is the amount of overlap between the bounding boxes of expanded building components) This determines whether building components are adjacent and generates the corresponding CSV file, such as... Figure 3 As shown.

[0058] Step 2: Divide and analyze the distribution of the actual features of the building components from Step 1, calculate the minimum difference in the dimensions of the building components, and define the BIM modeling spatial resolution. The extracted overall building location coordinates are then adjusted to the positive range in the three-dimensional coordinate system. That is, the BIM modeling space is defined based on the maximum values ​​of the building components' dimensions and location attributes. The dimensions and positional features of the building components after displacement and discretization are used as the true features after discretization.

[0059] This step preserves the characteristics of BIM itself to the greatest extent while making the decoder's prediction of the true features in step 3.3 more accurate.

[0060] Step 3: The discrete real features are encoded by a graph neural network encoder, and then fed into the decoder network after passing through the residual quantization module to obtain the predicted real features of the building components. These predicted real features are used as modeling parameters of the building components and are used for BIM modeling in step 6.

[0061] This step uses an encoder-decoder training method to input the real features and spatial features of the building components from step 2 into a graph data form. The residual quantization module then re-describes the high-dimensional features of the encoded building components, and finally, the graph is fed into the decoder to obtain the predicted values ​​of the real features of the building components. Cross-loss entropy is used to train the parameters of the encoder-decoder neural network. After training, a token list describing the building components is obtained from the residual quantization module.

[0062] Furthermore, step 3 specifically includes:

[0063] Step 3.1: Encode the true features of the building components to obtain high-dimensional features. Provide a higher-dimensional description of the building components:

[0064] The graph neural network consists of 3 SAGEConv convolutional layers; a depth of 3 residual quantization depths; and a 5-layer residual network that forms the decoder. Specifically, the first SAGEConv convolutional layer maps the input feature dimension from 15 to 32, the second layer maps the feature dimension from 32 to 64, and the third layer maps the feature dimension from 64 to 128.

[0065] Step 3.2: The quantization depth in the residual quantization module is set to 3. Residual quantization is performed on the high-dimensional features encoded by the graph neural network to obtain the words (tokens) in the vocabulary. The words (tokens) in the vocabulary are updated by the average exponential moving average method.

[0066]

[0067]

[0068] in, Represents a vocabulary list. This indicates the residual quantization operation. This represents the quantization feature obtained after the q-th residual quantization. This represents the distance in the vocabulary after residual quantization. Recent vocabulary index. This means retrieving the corresponding word embedding based on the word index, and then using the residual features obtained in the previous step ( Subtract the distance Recent word embeddings Quantization features obtained after obtaining the q-th residual .

[0069]

[0070] Finally, the redescribed architectural component features are obtained by summing the residual depths. .

[0071] Step 3.3: The decoder consists of a 5-layer residual network, which converts the redescribed building component features obtained in Step 3.2 into... The data is fed into the decoder, where each residual network layer predicts the true features of the building components through residual connections and convolution operations. The decoder's prediction of the true features of the building components is based on the output range and accuracy of the predicted values ​​within the BIM modeling space and spatial resolution limitations defined in step 2.

[0072] Step 4: Obtain the description sequence of the building components corresponding to the BIM-related building component information extracted in Step 1 according to the vocabulary obtained in Step 3. Arrange the (vocabulary) tokens in the description sequence of each building component in a specific order: for different building components, arrange them in the order of column -> beam -> wall -> door -> window. For the same component, arrange them in the order of priority from small to large based on the three-dimensional Cartesian coordinate system ZYX. Add the start word (SOS) and end word (EOS) to the first and last positions of the sequence, respectively. Use this as the training data for the pre-training model.

[0073] Step 5: Use a Transformer-Only Decoder model with a 3-layers-4-heads attention mechanism as the pre-trained model. Its input is a sequence of building descriptions obtained by comparing the building component descriptions with the vocabulary list acquired in Step 3.2 and sorting them according to the rules described in Step 4 (a self-supervised learning approach—serving as both input corpus and real building description sequence). Based on the prior distribution of the input description sequence, predict the content of subsequent building component description sequences to output a complete predicted building description sequence. Specifically:

[0074] Training: The building description sequence is fed into the pre-trained model to learn the layout of building components. The model parameters are updated based on the cross loss entropy between the predicted building sequence and the real building description sequence as the error backpropagation function.

[0075] The loss function of the pre-trained model is as follows:

[0076]

[0077]

[0078] in: Indicates the total number of building components. This indicates the number of words in the vocabulary list. Indicates the residual depth. This represents the probability distribution of predicted words in the vocabulary. Indicates the first The building component in the Layer quantization depth prediction vocabulary , This represents the smoothed probability distribution tensor obtained by smoothing the one-hot tensors of words in the real sequence. Indicates a smooth operation. Represents the first in the real sequence The building component in the Layered quantization depth of vocabulary indexing.

[0079] Generation: Set two task generation modes: incomplete BIM completion and random generation.

[0080] The incomplete BIM completion generation involves taking the real features of some building components as input, passing them through the encoder and quantization module mentioned above to obtain a partial building description sequence of words, and then inputting them into a pre-trained model to obtain a complete building description sequence.

[0081] Random generation: Given a start word (SOS) as input to a pre-trained model, a random sequence of complete building descriptions is obtained.

[0082] Step 6: The building description sequence generated by the pre-trained model is summed according to the quantization depth and then input into the decoder to decode the real features of the building components. BIM modeling is then completed based on these decoded real features. The specific BIM modeling process includes: creating a building project, defining building units and model views, modeling building components, and dividing floor information.

[0083] As a preferred embodiment, step 6 specifically includes:

[0084] Step 6.1: Modeling Building Components:

[0085] The modeling method for building components includes five types: columns, beams, walls, doors, and windows. Modeling of each component begins with creating a corresponding `ifc_entity` (solid) and specifying its category. Subsequently, combined with decoded dimension, position, and orientation information, corresponding IFC statements are generated to define the component's geometric and spatial characteristics. Specifically, the modeling of columns, beams, and walls mainly involves the `IfcRectangleProfileDef` (rectangular profile), `IfcExtrudedAreaSolid` (extruded body), and `IfcAxis2Placement3D` (3D coordinates) statements, whose dimension parameters (such as XDim, YDim, Depth, and OuterCurve) are directly taken from the decoded component dimension values. Modeling doors and windows additionally includes the statements IfcArbitraryClosedProfileDef (polygon profile definition), IfcDoorPanelProperties (door panel properties), IfcDoorLiningProperties (door frame properties), IfcWindowPanelProperties (window surface properties), and IfcWindowLiningProperties (window frame properties). Their size parameters are also derived from decoded values, while style parameters (such as Operation_type) are set to specific styles based on the component length (doors: left-opening single sliding door SINGLE_SWING_LEFT, double sliding door DOUBLE_DOOR_SLIDING; windows: vertical double-sided window DOUBLE_PANEL_VERTICAL, vertical three-sided window TRIPLE_PANEL_VERTICAL, single-sided window SINGLE_PANEL). The orientation of door and window components is checked, a preliminary matching and potential wall pairing list is generated, and then the Euclidean distance between the potential wall pairing list and the door is calculated. The expression is as follows:

[0086]

[0087] in Indicates the location of door and window components. The location of the wall is indicated, and the minimum Euclidean distance between the door and window components and each wall is calculated. First, define the walls to which the door and window components belong. Then, based on their size and location, create an ifc_entity and specify its category as IfcOpeningElement (opening space). Create an IfcRelVoidsElement (space belonging) statement to describe the association between the door and its wall.

[0088] For the location and orientation attributes of all the above building components, generate the Location and RefDirection parameter values ​​in the IfcAxis2Placement3D statement based on the decoded coordinates and orientation attributes.

[0089] Step 6.2: Divide building floor information, such as Figure 4 As shown:

[0090] Following step 6.1, after modeling the components, the Z-axis positions of the subsequent column and wall building components are divided and statistically analyzed into intervals. Intervals with more than 9% of the total number of building components are selected. Based on the number of intervals and the minimum value within each interval, the number and height of floors are determined. This generates the ObjectPlacement (location), Representation (attribute representation), and LongName (floor naming) parameter values ​​in the IfcBuildingStorey (floor definition) statement. Finally, the IfcRelContainedInSpatialStructure (space) statement is generated to associate the building components with the corresponding floors. Through these steps, a complete BIM modeling process from building component modeling to floor information division can be achieved.

[0091] Example 2:

[0092] This embodiment describes a computer-readable storage medium storing a computer program that, when executed by a processor, implements a pre-trained artificial intelligence framework for automatically generating BIM (Building Information Model) as described in any of Embodiment 1.

[0093] Example 3:

[0094] This embodiment describes a computer device, including:

[0095] Memory is used to store instructions.

[0096] A processor is configured to execute the instructions, causing the computer device to perform operations of a pre-trained artificial intelligence framework for automatically generating BIM (Building Information Model) as described in any of Embodiment 1.

[0097] Example 4:

[0098] The specific implementation process of the method of the present invention is described in this embodiment as follows: Figure 1 As shown:

[0099] Clean up the acquired BIM (remove components that are not considered and will interfere with information extraction, such as curtain walls, non-standard walls, circular columns, etc.).

[0100] First, based on the written .NET development program, the spatial features are obtained by determining the overlap of bounding boxes in BIM components. Then, the corresponding IFC file for the BIM is exported. An information extraction script is used to extract the actual features (category, orientation, size, location) of columns, beams, walls, doors, and windows, generating corresponding CSV files. Taking the information extraction of a single BIM building component as an example, the extraction results are as follows: Figure 2 As shown.

[0101] Pseudocode for extracting real features of building components (Pyhton):

[0102] Step 1: Initialize path and filename

[0103] Step 1.1: Set the IFC file path and name.

[0104] Step 1.2: Set the CSV file path and name.

[0105] Step 2: Define auxiliary functions

[0106] Step 2.1: Define the direction vector normalization function: `RegulateAndRespond_direct`.

[0107] Step 2.2: Define the size update function:

[0108] Step 2.2.1: Define `update_dimensions_with_Outercurve`.

[0109] Step 2.2.2: Define `update_dimensions`.

[0110] Step 2.3: Define the geometric representation processing function: `calculate_dimensions`.

[0111] Step 3: Process building components (columns, beams, walls, doors, windows)

[0112] Step 3.1: For each component type:

[0113] Step 3.1.1: Initialize the relevant lists (name, direction, position, size, category).

[0114] Step 3.1.2: Obtain the list of components of this type.

[0115] Step 3.1.3: Traverse each component:

[0116] Step 3.1.3.1: Calculate the dimensions and position based on the geometric representation type (Clipping, MappedRepresentation, Brep, etc.).

[0117] Step 3.1.3.2: Call the auxiliary function to process the direction vector and size.

[0118] Step 3.1.3.3: Add the results to the corresponding list.

[0119] Step 4: Recombination Characteristics

[0120] Step 4.1: Merge the characteristics of all components into a single list: `building_list`.

[0121] Step 4.2: Sort by x, y, z coordinates: `building_list`.

[0122] Step 4.3: Move the coordinates back to the origin.

[0123] Step 5: Generate a CSV file

[0124] Step 5.1: Set the CSV file header.

[0125] Step 5.2: Open the CSV file and write the header.

[0126] Step 5.3: Iterate through `building_list` and write the characteristics of each component into a CSV file.

[0127] Pseudocode for extracting spatial features of building components (C#):

[0128] Step 1: Initialize variables

[0129] Step 1.1: Define the filter: OrFilter.

[0130] Step 1.2: Filter elements: selectedElements.

[0131] Step 1.3: Initialize the list: adjacencyList.

[0132] Step 2: Iterate through selectedElements

[0133] Step 2.1: For each element:

[0134] Step 2.1.1: Get the bounding box: box1.

[0135] Step 2.1.2: Expand the bounding box: Box1 (call the ExpandBoundingBox method).

[0136] Step 2.1.3: Iterate through selectedElements again:

[0137] Step 2.1.3.1: If the current element is not the same as the previous element:

[0138] Step 2.1.3.1.1: Get the bounding box: box2.

[0139] Step 2.1.3.1.2: Expand the bounding box: Box2 (call the ExpandBoundingBox method).

[0140] Step 2.1.3.1.3: Check if the bounding boxes overlap (call the DoBoundingBoxesOverlap method):

[0141] Step 2.1.3.1.3.1: If the bounding boxes overlap:

[0142] Step 2.1.3.1.3.1.1: Calculate the overlap volume. If the volume meets the condition: add the result to adjacencyList.

[0143] Step 3: Processing Results

[0144] Step 3.1: Sort and deduplicate the adjacencyList (call the method IdsSort_with_Deduplication).

[0145] Step 3.2: Write the results to a CSV file (call the WriteToCSV method).

[0146] Secondly, the extracted building component information is input into the encoder-decoder module to train and obtain a vocabulary. The final vocabulary size is 67 * 128.

[0147] Map the information of building components to words (tokens) in the vocabulary list to complete the learning of building component layout.

[0148] First, according to the depth of quantization The sequence is converted into a building description sequence, where each building component is described by three terms. In the building description sequence, different components are arranged in the order of column -> beam -> wall -> door -> window. For the same component, they are arranged in ascending order of priority ZYX based on a three-dimensional Cartesian coordinate system. SOS (start token) and EOS (end token) are added to the first and last positions of the sequence, respectively.

[0149] Secondly, the architectural description sequence is used as the training corpus for learning the layout of architectural components in the pre-trained model.

[0150] Automated BIM modeling based on pre-trained model-predicted building description sequences can be completed in just 10 seconds from prediction to BIM modeling.

[0151] First, the pre-trained model generates building sequences in two ways: 1) generating building description sequences with random probability by inputting a start sign (SOS); 2) encoding and quantizing incomplete (partial) building components, then concatenating them with the start sign in the above-mentioned order to obtain the generated building description sequence as output. Figure 5 As shown. The predicted building description sequence is arranged according to quantization depth. The summation is then input into the decoder to obtain the decoded true features.

[0152] Secondly, the actual features of the obtained building components are based on the BIM modeling script. The generated IFC file is obtained by generating the corresponding IFC statement description, thus completing the parametric modeling of BIM.

[0153] Taking a parametric BIM model as an example, the result is as follows: Figure 6 As shown.

[0154] Pseudocode for some BIM modeling (Python):

[0155] Step 1: Generate building components

[0156] Step 1.1: Create the building component dictionary: walls_dict, doors_dict, windows_dict, column_dict, beam_dict.

[0157] Step 1.2: Iterate through each element in trans_result (including five categories: columns, beams, walls, doors, and windows):

[0158] Step 1.2.1: Obtain the element category class_key.

[0159] Step 1.2.2: Create the corresponding building component entity based on class_key, set its geometric representation and position, and then add it to the corresponding dictionary.

[0160] Step 2: Create floors and assign heights

[0161] Step 2.1: Create a floor height list (floor_list), a floor entity dictionary (floor_dic), and a floor height information dictionary (floor_h_dic).

[0162] Step 2.2: Divide the floor areas according to the height information, create an entity for each floor and set the height information.

[0163] Step 3: Assign floors to buildings

[0164] Step 3.1: Declare building as a building entity.

[0165] Step 3.2: Add all floor entities to the building.

[0166] Step 4: Traverse the transformed results and assign entities to floors.

[0167] Step 4.1: Declare trans_result as a list of transformed results.

[0168] Step 4.2: Iterate through each element in trans_result:

[0169] Step 4.2.1: Generate the corresponding IFC entity according to the element category, and set the geometric representation and position.

[0170] Step 4.2.2: Assign entities to the corresponding floors based on their height information.

[0171] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0172] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0173] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0174] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0175] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A pre-trained artificial intelligence method for automatically generating BIM images, characterized in that: Specifically, it includes: Clean up the BIM model and extract relevant building component information, including the actual characteristics and spatial characteristics of the building components; Discretize the real features of the building components to obtain the discretized real features; Graph data is constructed based on the discretized real features and the spatial features of building components. The graph data is then input into a graph neural network, a residual quantization module, and a decoder that are connected in sequence. The output is a vocabulary of building component descriptions, as well as a trained graph neural network, residual quantization module, and decoder. Based on relevant building component information and a glossary of building component descriptions, construct a description sequence for the building components; A pre-trained model is trained using the description sequences of building components to obtain a well-trained pre-trained model. The system obtains information about the building components to be generated, and uses a trained graph neural network and residual quantization module to obtain a word sequence of the building components to be generated. The word sequence of the building components to be generated is then input into a trained pre-trained model, which outputs a description sequence of the building components to be generated. The description sequence of the building component to be generated is quantized and deep summed, and then input into the trained decoder to output the real features of the building component to be generated. Generate a BIM model based on the actual characteristics of the building components to be generated.

2. The pre-trained artificial intelligence method for automatically generating BIM according to claim 1, characterized in that: The actual characteristics of the building components include: the type, orientation, size, and location of columns, beams, walls, doors, and windows; the spatial characteristics of the building components include: the adjacent relationships of columns, beams, walls, doors, and windows.

3. The pre-trained artificial intelligence method for automatically generating BIM according to claim 1, characterized in that: The discretization operation specifically includes: using the minimum difference in the size of the building components as the resolution, adjusting the position of the building components to the positive range in the three-dimensional coordinate system, and constructing the discretized real features of the building components based on the resolution and the adjusted position.

4. The pre-trained artificial intelligence method for automatically generating BIM according to claim 1, characterized in that: The graph neural network includes three sequentially connected SAGEConv convolutional layers; the quantization depth in the residual quantization module is set to 3; the decoder includes five sequentially connected residual network layers.

5. The pre-trained artificial intelligence method for automatically generating BIM according to claim 1, characterized in that: The step of constructing a description sequence for building components based on relevant building component information and a glossary of building component descriptions specifically includes: Retrieve the corresponding terms from the glossary of building component descriptions based on relevant building component information; The words are arranged in the order of pillars, beams, walls, doors, and windows. Words of the same type are arranged in ascending order of their coordinate values ​​on the Z, Y, and X axes of a three-dimensional Cartesian coordinate system. Then, add the starting word and the ending word to the first and last positions of the sequence respectively to obtain the descriptive sequence of the building components.

6. The pre-trained artificial intelligence method for automatically generating BIM according to claim 1, characterized in that: The pre-trained model uses a Transformer-Only Decoder model with a 3-layers-4-heads attention mechanism.

7. The pre-trained artificial intelligence method for automatically generating BIM according to claim 1, characterized in that: The process of obtaining information about the building components to be generated, and obtaining the vocabulary sequence of the building components to be generated through a trained graph neural network and residual quantization module, specifically includes: The information of the building components to be generated is obtained by taking the real features of some building components, inputting the real features of some building components into the trained graph neural network and residual quantization module, and outputting the word sequence of the building components to be generated.

8. The pre-trained artificial intelligence method for automatically generating BIM according to claim 1, characterized in that: The process of generating a BIM model based on the actual characteristics of the building components to be generated specifically includes: Create IFC entities corresponding to columns, beams, walls, doors, and windows, and obtain the category, orientation, size, and location based on the actual characteristics of the building components to be generated. Generate the corresponding IFC statements for the IFC entities to define the geometric and spatial characteristics of the building components, and complete the BIM modeling. The Z-axis positions of columns and walls are divided and statistically analyzed to filter out intervals with more than a threshold number of data points. The number of floors and heights are then determined based on the number of intervals and the minimum value within each interval, thus obtaining information about each floor in the BIM model.

9. A computer-readable storage medium, characterized in that: It stores a computer program that, when executed by a processor, implements a pre-trained artificial intelligence method for automatically generating BIM as described in any one of claims 1 to 8.

10. A computer device, characterized in that: include: Memory, used to store instructions; A processor for executing the instructions, causing the computer device to perform operations of a pre-trained artificial intelligence method for automatically generating BIM as described in any one of claims 1 to 8.

Images