A data generation method and system based on a building information model

Abstract

This invention discloses a data generation method and system based on Building Information Modeling (BIM), belonging to the field of BIM processing technology. The method includes: constructing a BIM topology graph by abstracting and representing the topological architecture of the BIM using graph theory; selecting a standard BIM node set from the topology graph and constructing a first BIM data set accordingly; optimizing the constructed BIM analysis neural network with at least two layers based on the first BIM data in the set, forming a deep BIM analysis neural network; extracting and loading the BIM data to be tested, and using the deep neural network to analyze and generate corresponding model-directed data, which serves as the representation data of the BIM data to be tested. This invention adapts the model to the feature dimensions and representation requirements of BIM data, improving the reliability and accuracy of BIM data representation generation.

Description

Technical Field

[0001] This invention relates to the field of building information processing technology, and more specifically to a data generation method and system based on building information modeling. Background Technology

[0002] Building Information Modeling (BIM) technology is a core data tool in engineering design, construction, and management. By integrating digitized and informational models of the entire building lifecycle, it enables the sharing and transmission of building information, providing a foundation for collaborative work among all project stakeholders. In BIM applications, the labeling data of the target building information model is crucial for describing its core characteristics and enabling information retrieval and management. However, current technologies often rely on manual, experience-based annotation for this labeling data, resulting in low reliability and insufficient accuracy.

[0003] Meanwhile, traditional BIM data generation methods do not fully consider the topological relationship characteristics of building information, and lack scientific quantitative indicators for the selection of building data nodes. This results in the exemplary data clusters used for neural network training containing a large amount of invalid data, affecting the optimization effect of the neural network. In addition, in the process of optimizing the building information analysis neural network, traditional methods only use a single feature mining optimization method, without combining the characteristics of BIM data that have both discrete features (such as building component types) and continuous features (such as component dimensions and parameters). The optimization dimension is single and the accuracy is insufficient, which further reduces the reliability of BIM data identification generation.

[0004] With the increasing complexity and intelligence of building engineering, higher demands are placed on the accuracy, topology fit, and intelligent analysis capabilities of BIM data generation. There is an urgent need for a BIM data generation method that integrates topology analysis, scientific node selection, and high-precision neural network optimization. This method can adapt the feature dimensions of building information data through formula transformation and parameter replacement, thereby addressing the aforementioned shortcomings of existing technologies. Summary of the Invention

[0005] In view of this, the present invention provides a data generation method and system based on Building Information Modeling (BIM). It constructs a BIM topology graph using graph theory, selects standard data nodes using the entropy-TOPSIS method, and simultaneously performs formula transformation and parameter replacement on a hybrid discrete-continuous action space model to adapt to the feature representation requirements of BIM data. This achieves multi-level, precise optimization of the BIM analysis neural network, solving the problems of low reliability, lack of topological feature integration, and insufficient neural network optimization accuracy in traditional BIM data representation generation, thereby improving the reliability and accuracy of BIM data representation generation. To achieve the above objectives, the present invention adopts the following technical solution: A data generation method based on Building Information Modeling (BIM) includes: The topological architecture of the building information model is abstracted and represented using graph theory, and a building information topology graph is constructed. A standard building information node set is obtained by filtering the building information topology map, and a first building information model data set is constructed based on the standard building information node set; Based on the first building information model data included in the first building information model dataset, the constructed building information analysis neural network is optimized at least two levels to form a corresponding deep building information analysis neural network. Extract the building information model data to be tested, load the building information model data to be tested, and use a deep building information analysis neural network to analyze and generate model pointing data corresponding to the building information model data to be tested. The model pointing data is the representation data of the building information model data to be tested.

[0006] Optionally, the filtering of the building information topology map to obtain a standard building information node set includes: 2a) Construct a building information node screening system by selecting the degree centrality index, compactness centrality index, edge betweenness centrality index, and information betweenness centrality index of building information nodes; 2b) Construct a multi-attribute decision matrix for building information based on centrality indicators; 2c) Standardize and normalize the multi-attribute decision matrix to obtain a normalized matrix; 2d) Calculate the information entropy, difference coefficient, and index weight of building information screening indicators based on the normalized matrix; 2e) Construct a weighted normalized decision matrix for building information based on information entropy, difference coefficients, and index weights; calculate the positive ideal solution, negative ideal solution, and the distance between each node and the optimal and worst solutions; 2f) Calculate the relative proximity of each building information node based on the positive ideal solution, the negative ideal solution, and the distance between each node and the optimal solution and the worst solution. Sort the nodes in descending order of importance based on the relative proximity and select the top η×100% of nodes to construct a standard building information node set as the node selection ratio.

[0007] Optionally, a first building information model data set is constructed based on the standard building information node set, including: 2g) Extract multiple initial first building information model data, each of the initial first building information model data having initial representation data; 2h) Perform data aggregation operation on each initial first building information model data corresponding to the same initial representation data to form the building information model data aggregation result corresponding to each initial representation data; 2i) Based on the initial representation data, analyze the hierarchical sorting relationship information corresponding to each initial first building information model data; 2j) Mark the data aggregation results of the building information model to form an optimized sub-level. Perform data aggregation operation on each optimized sub-level corresponding to the same level of sorting relationship information to form the optimized level aggregation result corresponding to each level of sorting relationship information. 2k) Based on the aggregation results of each optimization level, the first building information model data set is formed.

[0008] Optionally, the first building information model data in the first building information model dataset has optimized representation data, and the building information analysis neural network to be optimized belongs to the constructed building information analysis neural network or candidate building information analysis neural network. Optimizing the feature mining sub-network included in the building information analysis neural network to be optimized includes: a) Using the optimized building information analysis neural network, perform building information model analysis on the loaded first building information model data to generate analysis representation data corresponding to the loaded first building information model data; b) Calculate the representation data difference between the optimization representation data and the analysis representation data, and optimize the subnetwork parameters of the feature mining subnetwork included in the building information analysis neural network to be optimized based on the representation data difference, thereby completing the optimization of the building information analysis neural network to be optimized.

[0009] Optionally, the step of optimizing the constructed building information analysis neural network at least two levels based on the first building information model data included in the first building information model dataset to form a corresponding deep building information analysis neural network includes: 3a) Extract the first initial building information model data corresponding to the first optimization level, and optimize the feature mining sub-network included in the constructed building information analysis neural network based on the first initial building information model data to form the corresponding candidate building information analysis neural network. 3b) Extract the second initial building information model data corresponding to the next optimization level, and optimize the feature mining sub-network included in the candidate building information analysis neural network based on the second initial building information model data to form the corresponding optimized building information analysis neural network; 3c) Load the used building information model data corresponding to the second initial building information model data into the feature mining subnetworks included in the candidate building information analysis neural network and the optimized building information analysis neural network, respectively, and output the corresponding candidate building information feature mining results and optimized building information feature mining results; 3d) Construct a hybrid discrete-continuous action space model for building information using a conditional variational autoencoder and an embedding table. Based on the candidate building information feature mining results and optimized building information feature mining results corresponding to the same first building information model data, optimize the data transformation output sub-network included in the optimized building information analysis neural network in combination with the hybrid discrete-continuous action space model for building information to form a corresponding optimized candidate building information analysis neural network. 3e) Reverse the execution of step 3b) until the preset K levels of optimization are completed. When the optimization ends, mark the candidate building information analysis neural network corresponding to the end of the optimization as a deep building information analysis neural network.

[0010] Optionally, the construction of a hybrid discrete-continuous action space model of building information using a conditional variational autoencoder and an embedding table as described in step 3d) includes: 3d1) Embedding tables using learnable building information; 3d2) Using the state and embedding table of the building information model as conditions, the building information conditional variational autoencoder is used to map the continuous action parameters corresponding to discrete actions to latent variables, thus constructing a building information latent representation space. 3d3) Decode the latent variables of building information; 3d4) The embedding table and building information conditional variational autoencoder network and its parameters are trained by minimizing the loss function.

[0011] Optionally, the building information hybrid discrete-continuous action space model further uses a sub-network after the decoder of the conditional variational autoencoder to predict the building information state residual, defining the actual state residual and the predicted state residual for the building information sample. Based on the actual state residuals and the predicted state residuals, the L2 norm squared prediction loss is used as a regularization term: The total training loss function for the building information hybrid action representation model is constructed based on the L2 norm squared prediction loss and the variational autoencoder network loss.

[0012] Optionally, the candidate building information feature mining results include: , For candidate network feature mining units The results of the excavation; The optimized building information feature mining results are as follows: , To optimize network feature mining units The results of the excavation.

[0013] Optionally, in step 3d), optimizing the data transformation output sub-network of the optimized building information analysis neural network by combining the building information hybrid discrete-continuous action space model includes: 3d5) The mining results of each optimization network feature mining unit corresponding to the used building information model data are loaded into the corresponding optimization network feature mining unit in the optimization building information analysis neural network, and the corresponding local optimization building information feature transformation results are output. 3d6) Mark the one-to-one corresponding feature mining unit and data conversion output unit as the first feature mining unit and the first data conversion output unit; 3d7) Based on the mining results of the candidate network feature mining units corresponding to the same first building information model data and the results of local optimization building information feature transformation, combined with the training results of the building information hybrid discrete-continuous action space model, the corresponding data learning cost index is analyzed. 3d8) Based on each of the data learning cost indicators, the parameters of the corresponding optimized network feature mining unit in the optimized building information analysis neural network are optimized to form the corresponding optimized candidate building information analysis neural network.

[0014] Optionally, a data generation system based on Building Information Modeling includes: a Building Information Topology Construction Module: used to abstractly represent the topological architecture of the Building Information Modeling through graph theory and construct a Building Information Topology Graph; Standard Node Set and Data Set Construction Module: Used to filter the building information topology map to obtain a standard building information node set, and construct a first building information model data set based on the standard building information node set; Neural network optimization module: used to optimize the constructed building information analysis neural network at least two levels based on the first building information model data included in the first building information model data set, so as to form a corresponding deep building information analysis neural network; The test model data extraction module is used to extract the building information model data to be tested. Model representation data generation module: used to load the building information model data to be tested, and use a deep building information analysis neural network to analyze and generate model pointing data corresponding to the building information model data to be tested, wherein the model pointing data is the representation data of the building information model data to be tested.

[0015] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a data generation method and system based on building information modeling, which has the following beneficial effects: This invention constructs a building information topology graph using graph theory, abstracting building information into nodes and associated edges. Combined with the parameter-replaced entropy-TOPSIS method, it quantifies and filters standard building information nodes from a complex network perspective, eliminating invalid data interference. This makes the construction of the first dataset more closely resemble the information association characteristics of building projects, providing high-quality training data for neural network optimization. Targeted transformations are made to core formulas such as the hybrid discrete-continuous action space model, the entropy-TOPSIS method, and neural network optimization, replacing general parameters with parameters specific to the building information domain. This allows the model and method to deeply adapt to the discrete and continuous characteristics of BIM data, enhancing the model's ability to represent building information. The building information analysis neural network is optimized at least two levels, with targeted optimization of the feature mining sub-network and the data transformation output sub-network. Combined with the hybrid discrete-continuous action space model of building information, refined optimization is achieved, solving the problems of single-dimensional optimization and insufficient precision in traditional optimization methods, significantly improving the neural network's ability to analyze BIM data. By using standard node selection, high-quality dataset construction, and high-precision neural network optimization, the traditional manual experience-based annotation method is replaced, making BIM data representation and generation more automated and intelligent. Attached Figure Description

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

[0017] Figure 1 This invention provides a schematic flowchart of a data generation method based on Building Information Modeling. Detailed Implementation

[0018] The technical solutions of 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.

[0019] This invention discloses a data generation method based on Building Information Modeling (BIM), such as... Figure 1 As shown, it includes: Step 1): Use graph theory to abstractly represent the topological architecture of the Building Information Model (BIM) and construct a BIM topology graph. Where V is the set of all data nodes in the Building Information Model, and E is the set of edges connecting the data nodes. This is the adjacency matrix of building information nodes, with matrix elements... , This indicates that there is a relationship between building information nodes m and n. =0 indicates no association; Step 2): Filter the building information topology map to obtain a standard building information node set, and construct the first building information model data set based on the standard building information node set. The first building information model data set Includes first building information model data corresponding to at least two optimization levels. , This represents the total number of optimization levels. Step 3): Based on the first building information model data set Includes first building information model data The building information analysis neural network will be built At least two levels of optimization are performed to form the corresponding deep building information analysis neural network. ; Step 4) Extract the building information model data to be tested. ; Step 5) Process the building information model data to be tested. Loading is performed to utilize deep building information analysis neural networks. Analyze and generate the building information model data to be tested. The corresponding model points to the data. The model points to data. The building information model data to be tested The data represents the data.

[0020] Furthermore, step 2) describes the selection of a standard building information node set based on complex network theory and the entropy-TOPSIS method, including: 2a) Select the degree centrality index for building information nodes Density centrality index Border betweenness centrality index and information betweenness centrality index Construct a building information node filtering system, in which: , , This represents the total number of nodes in the building information topology graph. Let i be the number of associations with node i; , The shortest topological distance between building information nodes i and j; 2b) Construct a multi-attribute decision matrix for building information: , , ; 2c) For multi-attribute decision matrices Standardization and normalization are performed to obtain the normalized matrix. ,in: Standardized formula: ; Normalization formula: ; 2d) Calculate the information entropy of the j-th building information screening index. Difference coefficient and indicator weights ,in: ; ; ; 2e) Construct a weighted normalized decision matrix for building information: ; ; Calculate the ideal solution Negative ideal solution And the distance between each node and the optimal and worst solutions. , ,in: ; ; ; ; 2f) Calculate the relative proximity of each building information node. ,according to Nodes are sorted in descending order of importance, and the top η×100% of nodes are selected to construct a standard building information node set. , Set the node filtering ratio.

[0021] Furthermore, in step 2), the standard building information node set is used as a basis. Constructing the first building information model dataset ,include: 2g) Extract multiple initial first building information model data Each of the initial first building information model data With initial representation data ,and ; 2h) for the same Each corresponding Perform data aggregation operations to form each The corresponding building information model data aggregation results ; 2i) Based on Analyze each Corresponding hierarchical sorting information , There is a one-to-one correspondence between the optimization level k; 2j) for Labeling to form optimized sub-hierarchies For the same Each corresponding Perform data aggregation operations to form each Corresponding optimization hierarchical aggregation results ; 2k) based on each This forms the first building information model dataset. .

[0022] Furthermore, the first building information model dataset The first building information model data in the data has optimized representation data. The neural network for building information analysis needs to be optimized. The building information analysis neural network described above Or the candidate building information analysis neural network The neural network for building information analysis that needs optimization will be used. Includes feature mining subnetwork Optimizations include: a) Utilizing the aforementioned building information analysis neural network that needs optimization The first building information model data is loaded and subjected to building information model analysis to generate analysis representation data corresponding to the first building information model data. ; b) Calculate the data used for optimization. and analysis of data The difference between the data represents the degree of data variation. ,in accordance with The neural network for building information analysis that needs optimization Includes feature mining subnetwork Subnetwork parameters The optimized update formula is as follows: ; Where t is the number of iterations. The learning rate for the feature mining subnetwork, To calculate the gradient of the difference with respect to the parameters, the neural network for building information analysis needs to be optimized. Optimization.

[0023] Furthermore, in step 3), the data set based on the first building information model is... Includes first building information model data The building information analysis neural network will be built At least two levels of optimization are performed to form the corresponding deep building information analysis neural network. The steps include: 3a) Extract the first initial building information model data corresponding to the first optimization level. Based on the first initial building information model data The building information analysis neural network will be built Includes feature mining subnetwork Optimize the neural network to form the corresponding candidate building information analysis network. ; 3b) Extract the second initial building information model data corresponding to the next optimization level. (k≥2), based on the second initial building information model data The candidate building information analysis neural network Includes feature mining subnetwork Optimize to form a corresponding optimized building information analysis neural network. ; 3c) Transfer the second initial building information model data respectively Corresponding Building Information Modeling Data Loaded into the candidate building information analysis neural network and the optimized building information analysis neural network Includes feature mining subnetwork In the process, the corresponding candidate building information feature mining results are output. And optimize the results of building information feature mining ; 3d) Constructing a hybrid discrete-continuous action space model of building information using a conditional variational autoencoder and an embedding table. Based on the same first building information model data and Combined with the model The optimized building information analysis neural network Includes data conversion output sub-network Optimization is performed to form a corresponding optimized candidate building information analysis neural network. ; 3e) Reverse the execution of step 3b) until the preset K levels of optimization are completed. If the optimization ends, mark the candidate building information analysis neural network corresponding to the end of the optimization as a deep building information analysis neural network. .

[0024] Furthermore, as described in step 3d), a hybrid discrete-continuous action space model of building information is constructed using a conditional variational autoencoder and an embedding table. ,include: 3d1) Embedding tables with learnable building information Build The implicit space of discrete actions of building information, in which each row of the table is embedded. Represents discrete actions of building information The corresponding one-dimensional continuous vector has a total dimension of , For embedded table parameters; 3d2) in the state of Building Information Model and embedded tables As a condition, the discrete actions are processed using a building information conditional variational autoencoder. Corresponding continuous motion parameters Mapping to latent variables In this process, a potential representation space for building information is constructed, and the encoding process is as follows: ; in, To encode network parameters, For the encoding process; 3d3) Decode the latent variables of building information. The decoding process is as follows: ; ; in, To decode network parameters, For the decoding process, Continuous action parameters for predicted building information. For building information feature conversion network, Reconstructing the fully connected layer for building information; 3d4) Minimize the embedding table by minimizing the loss function The building information conditional variational autoencoder network and its parameters are trained, and the loss function is: ; ; in, It is a building information replay buffer. for Vigaussian prior, for divergence, Let be the mathematical expectation.

[0025] Furthermore, the building information hybrid discrete-continuous action space model A subnetwork is further used after the decoder of the conditional variational autoencoder. To predict the residuals of building information state, for building information samples Its actual state residual and predicted state residuals Defined as: ; ; The L2 norm squared prediction loss is used as a regularization term: ; The total training loss function for the building information hybrid action representation model is: ; in, It is the weight of the building information regularization term. ∈ (0,1).

[0026] Furthermore, let the deep building information analysis neural network be the building information analysis neural network under test, wherein the building information analysis neural network under test... The building information analysis neural network described above The candidate building information analysis neural network The optimized building information analysis neural network Or the deep building information analysis neural network , The neural network for analyzing building information under test Includes feature mining subnetwork and data conversion output subnetwork ; The feature mining subnetwork include Feature mining unit This is used to perform feature mining operations on the loaded first building information model data, and output the first building information feature mining results. Including each feature mining unit Results of mining local first building information features ,Right now ; The data conversion output subnetwork include Data conversion output unit ,and and One-to-one correspondence, used to convert the first building information feature mining results Perform data transformation and output operations to output the corresponding second building information feature mining results. ,in for right The conversion result.

[0027] Furthermore, the candidate building information feature mining results include: ; For candidate network feature mining units The mining results; the optimized building information feature mining results , To optimize network feature mining units The results of the excavation; Furthermore, in step 3d), the hybrid discrete-continuous action space model combining the building information is described. The optimized building information analysis neural network Includes data conversion output sub-network Optimizations include: 3d5) The used building information model data are respectively Each corresponding Loaded into the optimized building information analysis neural network The corresponding The transformation output corresponds to the transformation result of each locally optimized building information feature. ; 3d6) will correspond one-to-one and Marked as the first feature mining unit and the first data conversion output unit ; 3d7) Based on the same first building information model data and Combining Building Information Hybrid Discrete-Continuous Action Space Model The training results were analyzed to determine the corresponding data learning cost metrics. , , The cost of the fusion coefficient; 3d8) Based on each of the data learning cost metrics The optimized building information analysis neural network will be used. The corresponding parameters The formula has been optimized and updated as follows: ; in, The learning rate of the sub-network is transformed to form the corresponding optimized candidate building information analysis neural network. .

[0028] In a specific implementation, a data generation method based on Building Information Modeling (BIM) includes the following specific steps: (1) Building Information Topology Construction The topological architecture of Building Information Modeling (BIM) is abstracted and represented using graph theory, and a BIM topology graph is constructed. ,in It is the collection of all data nodes in the building information model (such as building components, equipment, and process data nodes). It is a set of edges that connect data nodes (such as spatial connections, logical connections, and process connections). This is the adjacency matrix of building information nodes, with matrix elements... , This indicates that there is a relationship between building information nodes m and n. =0 indicates no association. This step abstracts the Building Information Model into a topological structure, realizing a quantitative topological representation of building information, laying the foundation for subsequent standard node selection.

[0029] (2) Construction of standard node set and data set

[0030] A standard set of building information nodes was selected based on complex network theory and the entropy-TOPSIS method. Based on the standard node set Constructing the first building information model dataset The first building information model data set Includes first building information model data corresponding to at least two optimization levels. , This represents the total number of optimization levels.

[0031] The selection criteria are designed based on the characteristics of building information nodes, including degree centrality. Density centrality Border Centrality and information betweenness centrality And define the formulas for each indicator to adapt to the characteristics of building information topology; The core formula of the entropy-TOPSIS method is modified by replacing general parameters with those specific to the building information field, thereby enabling the scientific quantitative selection of building information nodes. The first dataset, constructed based on a standard node set, eliminates invalid data interference, providing high-quality training data for subsequent neural network optimization.

[0032] (3) Multi-level optimization of building information analysis neural network

[0033] Based on the first building information model data set Includes first building information model data The building information analysis neural network will be built At least two levels of optimization are performed to form the corresponding deep building information analysis neural network. This step is the core of the invention, and specifically includes: 3a) Extract the first initial building information model data corresponding to the first optimization level. ,in accordance with Neural Networks for Building Information Analysis Feature mining subnetwork Optimize the network to form a candidate building information analysis neural network. ; 3b) Extract the second initial building information model data corresponding to the next optimization level. (k≥2), based on right of Further optimization resulted in an optimized building information analysis neural network. ; 3c) Used Building Information Modeling (BIM) data Loaded to and of In the process, the feature mining results of candidate building information are output. And optimize the results of building information feature mining ; 3d) Construct a hybrid discrete-continuous action space model based on building information. The formulas of the general hybrid discrete-continuous action space model are transformed and replaced with parameters specific to the building information domain, combined with... and right Data conversion output subnetwork Optimization is performed to form an optimized candidate building information analysis neural network. ; 3e) Reverse the execution of step 3b) until the preset K levels of optimization are completed, and label the final candidate network as a deep building information analysis neural network. .

[0034] (4) Data extraction of the model to be tested

[0035] Building Information Model Data to be Generated , It can be any BIM data throughout the entire lifecycle of a building project, such as structural component data, mechanical and electrical equipment data, construction process data, etc.

[0036] (5) Model representation data generation

[0037] For the building information model data to be tested Loading is performed, and the optimized deep building information analysis neural network is used. Analyze and generate model-pointing data corresponding to the test data. The model points to data. As data of the building information model to be tested The data representing, It contains core feature information of the data to be tested, such as component type, material, size, performance parameters, etc.

[0038] In a specific implementation, a data generation system based on Building Information Modeling (BIM) enables automated and high-precision generation of BIM data representation through the collaborative work of its modules, including: Building Information Topology Construction Module: Its core function is to abstract and represent the topological architecture of the building information model using graph theory, and to construct the building information topology graph. This enables the topological and quantitative representation of building information, providing basic data for subsequent standard node selection.

[0039] Standard Node Set and Data Set Construction Module: Based on complex network theory and the entropy-TOPSIS method, this invention uses a modified formula and replaced parameters to select a standard building information node set. And construct the first building information model data set based on the standard node set. This provides high-quality training data for neural network optimization.

[0040] Neural Network Optimization Module: This is the core module of the system, containing a hybrid discrete-continuous action space model of building information after formula transformation and parameter replacement. Used based on the first building information model data set The building information analysis neural network was built At least two levels of optimization are performed, sequentially optimizing the feature mining subnetwork and the data transformation output subnetwork, ultimately forming a deep building information analysis neural network. .

[0041] The test model data extraction module is used to extract the test building information model data to be processed from the building information model database. It supports the extraction of BIM data throughout the entire lifecycle, including structure, electromechanical systems, construction, and operation and maintenance.

[0042] Model representation data generation module: used to load the building information model data to be tested. Utilizing an optimized deep building information analysis neural network The model that generates the test data points to the data. and will The data to be measured is stored in the BIM database to achieve automated generation and management of the data.

[0043] In a specific embodiment, the generation of BIM model data representation of a commercial complex structure is taken as an example: This embodiment takes the generation of structural BIM model data of a commercial complex in the core area of ​​a city as an example. The method of this invention is applied to generate the BIM model representation data of the reinforced concrete columns of the complex's core tube. The specific implementation steps are as follows: (1) Building Information Topology Construction The structural components of a commercial complex, such as columns, beams, slabs, walls, and foundations, are abstracted into building information nodes. There are a total of 286 nodes; the spatial connections and force transfer relationships between components are abstracted into associated edges. Construct an adjacency matrix Matrix elements The building information topology diagram is generated by assigning values ​​of 0 or 1 based on the component relationships. .

[0044] (2) Construction of standard node set and data set

[0045] 2a) Calculate the 286 structural nodes Construct a multi-attribute decision matrix for building information ; 2b) For Standardization and normalization are performed to obtain the normalized matrix. Calculate the information entropy of each indicator. Difference coefficient and indicator weights ,in, ; 2c) Construct a weighted normalized decision matrix Calculate the relative proximity of each node. Set the node filtering ratio η=0.7, and select the first 200 nodes to construct a standard building information node set. ; 2d) From 800 sets of initial first-structure BIM data were extracted, aggregated according to the initial representation data (columns, beams, slabs, walls), and divided into 3 optimization levels (K=3) to construct the first building information model data set. .

[0046] (3) Multi-level optimization of building information analysis neural network

[0047] 3a) Constructing a building information analysis neural network Including feature mining subnetwork (6 feature mining units) (m=1-6) and data conversion output subnetwork (6 data conversion output units) ,and (One-to-one correspondence); setting the learning rate of the feature mining sub-network. ,use optimization parameters A neural network for analyzing candidate building information is formed. ; 3b) Utilize right of Further optimization resulted in an optimized building information analysis neural network. ; 3c) will Corresponding used data Load to and of In the middle, output and ; 3d) Construct a hybrid discrete-continuous action space model based on building information. Set up embedded building information table ( Regularization term weights Cost fusion coefficient Data conversion subnetwork learning rate Training using the modified loss function , combined and Calculate the cost of data learning metrics ,optimization of Parameters, forming ; 3e) Utilization Repeat steps 3b)-3d) to complete the 3-level optimization, and label the final candidate network as a deep building information analysis neural network. .

[0048] (4) Data extraction of the model to be tested

[0049] BIM model data of the core tube reinforced concrete columns were extracted from the BIM model database of the commercial complex structure as the data to be measured. This column is a core load-bearing component, and a core feature representation including material, size, strength, and floor height needs to be generated.

[0050] (5) Model representation data generation

[0051] Will Load to In the middle, the analysis and generation model points to the data. Core tube column - reinforced concrete - C60 - cross-sectional dimensions 1200×1200mm - story height 4.5m - seismic resistance level 1. This data is for testing purposes. The data represents the data.

[0052] The representation data generated in this embodiment has a 99.3% match with the actual design and construction features of the core column, which is much higher than the 90.2% of traditional manual annotation; the representation data generation time is only 0.8s, compared with 2.1s of manual annotation, which is 61.9% more efficient.

[0053] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0054] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A data generation method based on Building Information Modeling, characterized in that, include: The topological architecture of the building information model is abstracted and represented using graph theory, and a building information topology graph is constructed. A standard building information node set is obtained by filtering the building information topology map, and a first building information model data set is constructed based on the standard building information node set; Based on the first building information model data included in the first building information model dataset, the constructed building information analysis neural network is optimized at least two levels to form a corresponding deep building information analysis neural network. Extract the building information model data to be tested, load the building information model data to be tested, and use a deep building information analysis neural network to analyze and generate model pointing data corresponding to the building information model data to be tested. The model pointing data is the representation data of the building information model data to be tested.

2. The data generation method based on Building Information Modeling according to claim 1, characterized in that, The filtered building information topology map yields a standard set of building information nodes, including: 2a) Construct a building information node screening system by selecting the degree centrality index, compactness centrality index, edge betweenness centrality index, and information betweenness centrality index of building information nodes; 2b) Construct a multi-attribute decision matrix for building information based on centrality indicators; 2c) Standardize and normalize the multi-attribute decision matrix to obtain a normalized matrix; 2d) Calculate the information entropy, difference coefficient, and index weight of building information screening indicators based on the normalized matrix; 2e) Construct a weighted normalized decision matrix for building information based on information entropy, difference coefficients, and index weights; calculate the positive ideal solution, negative ideal solution, and the distance between each node and the optimal and worst solutions; 2f) Calculate the relative proximity of each building information node based on the positive ideal solution, the negative ideal solution, and the distance between each node and the optimal solution and the worst solution. Sort the nodes in descending order of importance based on the relative proximity and select the top η×100% of nodes to construct a standard building information node set as the node selection ratio.

3. The data generation method based on Building Information Modeling according to claim 2, characterized in that, The first building information model data set is constructed based on the standard building information node set, including: 2g) Extract multiple initial first building information model data, each of the initial first building information model data having initial representation data; 2h) Perform data aggregation operation on each initial first building information model data corresponding to the same initial representation data to form the building information model data aggregation result corresponding to each initial representation data; 2i) Based on the initial representation data, analyze the hierarchical sorting relationship information corresponding to each initial first building information model data; 2j) Mark the data aggregation results of the building information model to form an optimized sub-level. Perform data aggregation operation on each optimized sub-level corresponding to the same level of sorting relationship information to form the optimized level aggregation result corresponding to each level of sorting relationship information. 2k) Based on the aggregation results of each optimization level, the first building information model data set is formed.

4. The data generation method based on Building Information Modeling according to claim 1, characterized in that, The first building information model data in the first building information model dataset has data for optimization. The building information analysis neural network to be optimized belongs to the constructed building information analysis neural network or the candidate building information analysis neural network. The feature mining sub-network included in the building information analysis neural network to be optimized is optimized, including: a) Using the optimized building information analysis neural network, perform building information model analysis on the loaded first building information model data to generate analysis representation data corresponding to the loaded first building information model data; b) Calculate the representation data difference between the optimization representation data and the analysis representation data, and optimize the subnetwork parameters of the feature mining subnetwork included in the building information analysis neural network to be optimized based on the representation data difference, thereby completing the optimization of the building information analysis neural network to be optimized.

5. The data generation method based on Building Information Modeling according to claim 1, characterized in that, The step of optimizing the constructed building information analysis neural network at least two levels based on the first building information model data included in the first building information model dataset to form a corresponding deep building information analysis neural network includes: 3a) Extract the first initial building information model data corresponding to the first optimization level, and optimize the feature mining sub-network included in the constructed building information analysis neural network based on the first initial building information model data to form the corresponding candidate building information analysis neural network. 3b) Extract the second initial building information model data corresponding to the next optimization level, and optimize the feature mining sub-network included in the candidate building information analysis neural network based on the second initial building information model data to form the corresponding optimized building information analysis neural network; 3c) Load the used building information model data corresponding to the second initial building information model data into the feature mining subnetworks included in the candidate building information analysis neural network and the optimized building information analysis neural network, respectively, and output the corresponding candidate building information feature mining results and optimized building information feature mining results; 3d) Construct a hybrid discrete-continuous action space model for building information using a conditional variational autoencoder and an embedding table. Based on the candidate building information feature mining results and optimized building information feature mining results corresponding to the same first building information model data, optimize the data transformation output sub-network included in the optimized building information analysis neural network in combination with the hybrid discrete-continuous action space model for building information to form a corresponding optimized candidate building information analysis neural network. 3e) Reverse the execution of step 3b) until the preset K levels of optimization are completed. When the optimization ends, mark the candidate building information analysis neural network corresponding to the end of the optimization as a deep building information analysis neural network.

6. The data generation method based on Building Information Modeling according to claim 5, characterized in that, Step 3d) describes constructing a hybrid discrete-continuous action space model of building information using a conditional variational autoencoder and an embedding table, including: 3d1) Embedding tables using learnable building information; 3d2) Using the state and embedding table of the building information model as conditions, the building information conditional variational autoencoder is used to map the continuous action parameters corresponding to discrete actions to latent variables, thus constructing a building information latent representation space. 3d3) Decode the latent variables of building information; 3d4) The embedding table and building information conditional variational autoencoder network and its parameters are trained by minimizing the loss function.

7. A data generation method based on Building Information Modeling according to claim 6, characterized in that, The building information hybrid discrete-continuous action space model further uses a sub-network after the decoder of the conditional variational autoencoder to predict the building information state residuals. For the building information sample, the actual state residual and the predicted state residual are defined. Based on the actual state residuals and the predicted state residuals, the L2 norm squared prediction loss is used as a regularization term: The total training loss function for the building information hybrid action representation model is constructed based on the L2 norm squared prediction loss and the variational autoencoder network loss.

8. A data generation method based on Building Information Modeling according to claim 5, characterized in that, The results of feature mining of the candidate building information include: ， For candidate network feature mining units The results of the excavation; The optimized building information feature mining results are as follows: ， To optimize network feature mining units The results of the excavation.

9. A data generation method based on Building Information Modeling according to claim 5, characterized in that, Step 3d) involves optimizing the data transformation output sub-network of the optimized building information analysis neural network by combining the building information hybrid discrete-continuous action space model, including: 3d5) The mining results of each optimization network feature mining unit corresponding to the used building information model data are loaded into the corresponding optimization network feature mining unit in the optimization building information analysis neural network, and the corresponding local optimization building information feature transformation results are output. 3d6) Mark the one-to-one corresponding feature mining unit and data conversion output unit as the first feature mining unit and the first data conversion output unit; 3d7) Based on the mining results of the candidate network feature mining units corresponding to the same first building information model data and the results of local optimization building information feature transformation, combined with the training results of the building information hybrid discrete-continuous action space model, the corresponding data learning cost index is analyzed. 3d8) Based on each of the data learning cost indicators, the parameters of the corresponding optimized network feature mining unit in the optimized building information analysis neural network are optimized to form the corresponding optimized candidate building information analysis neural network.

10. A data generation system based on Building Information Modeling, characterized in that, include: Building Information Topology Construction Module: Used to abstractly represent the topological architecture of the building information model through graph theory and construct the building information topology graph; Standard Node Set and Data Set Construction Module: Used to filter the building information topology map to obtain a standard building information node set, and construct a first building information model data set based on the standard building information node set; Neural network optimization module: used to optimize the constructed building information analysis neural network at least two levels based on the first building information model data included in the first building information model data set, so as to form a corresponding deep building information analysis neural network; The test model data extraction module is used to extract the building information model data to be tested. Model representation data generation module: used to load the building information model data to be tested, and use a deep building information analysis neural network to analyze and generate model pointing data corresponding to the building information model data to be tested, wherein the model pointing data is the representation data of the building information model data to be tested.

Images