Artificial intelligence-based building design automation system

Abstract

The application relates to the technical field of architectural design and discloses an automatic architectural design system based on artificial intelligence, a collection module is used for collecting multi-source data of a target architectural project, and the multi-source data is subjected to standardized preprocessing; an extraction module is used for inputting the preprocessed multi-source data into a multi-modal feature enhancement network based on a Transformer attention mechanism, extracting and fusing target features of the multi-source data, and outputting a feature vector; a generation module is used for inputting the feature vector into a StyleGAN2-RL fusion generation model and outputting an initial architectural design scheme; an adjustment module is used for inputting the initial architectural design scheme into a dynamic constraint optimizer based on fuzzy logic and outputting an optimized design scheme meeting constraint requirements; and an output module is used for converting the optimized design scheme into a BIM model containing a professional interface, verifying the BIM model through BIM collision detection, and outputting a final architectural design scheme; the application improves the efficiency of architectural design automation.

Description

Technical Field

[0001] This invention relates to the field of architectural design technology, and more specifically to an automated architectural design system based on artificial intelligence. Background Technology

[0002] The architectural design industry is gradually transforming towards automation and intelligence. Although various automated architectural design systems have been initially applied, they are still insufficient to meet the design needs of complex architectural projects. They only perform format cleaning on basic data such as building codes and topography, without effectively extracting key features from multi-source data. This results in poor adaptability of design schemes to site conditions. Traditional generative adversarial networks or rule-driven models are often used, focusing only on spatial aesthetics or single functions, and failing to take into account multiple objectives such as structural safety, energy consumption optimization, and evacuation efficiency. Design constraints are processed with fixed weights, and the priority of constraints cannot be dynamically adjusted according to the project type. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing an automated building design system based on artificial intelligence.

[0004] This invention provides an artificial intelligence-based automated building design system, the system comprising:

[0005] The acquisition module is used to collect multi-source data of the target building project and perform standardized preprocessing on the multi-source data, wherein the multi-source data includes terrain LiDAR data, climate data, building design code data and user requirement text data.

[0006] The extraction module is used to input preprocessed multi-source data into a multimodal feature enhancement network based on the Transformer attention mechanism, extract and fuse target features from the multi-source data, and output feature vectors.

[0007] The generation module is used to input the feature vector into the StyleGAN2-RL fusion generation model. The generation model generates architectural design geometric parameters through the improved StyleGAN2 and optimizes the generation strategy with multiple objective reward items such as structural stress compliance rate, energy consumption index and space utilization rate through the RL module based on PPO algorithm, and outputs the initial architectural design scheme.

[0008] The adjustment module is used to input the initial architectural design scheme into a dynamic constraint optimizer based on fuzzy logic. The dynamic constraint optimizer dynamically adjusts the weights of the design constraints according to the project type, performs constraint adaptation optimization on the initial architectural design scheme, and outputs an optimized design scheme that meets the constraint requirements.

[0009] The output module is used to convert the optimized design scheme into a BIM model containing professional interfaces, and after verification by BIM collision detection, output the final architectural design scheme.

[0010] Optionally, in a first implementation of the present invention, the acquisition module includes:

[0011] The classification preprocessing submodule is used to classify and preprocess the collected multi-source data, convert the preprocessed terrain LiDAR data into raster data format, and normalize the data values ​​to the [0,1] interval according to the terrain slope and elevation range.

[0012] The first conversion submodule is used to convert the preprocessed climate data into energy efficiency impact correlation data and to map temperature and wind speed parameters according to climate zoning standards.

[0013] The second conversion submodule is used to convert structured building design code data into coded data, with each code requiring a unique identification code.

[0014] The third transformation submodule is used to convert the user's required text segmentation results into fixed-dimensional vectors through word embedding algorithms, and finally convert all types of data into tensors of the same dimension to complete the standardization preprocessing.

[0015] Optionally, in a second implementation of the present invention, the classification preprocessing submodule includes:

[0016] For terrain LiDAR data, a point cloud denoising algorithm is used to remove outliers caused by equipment errors and environmental interference. Then, a point cloud registration algorithm is used to unify the point cloud data collected at different times into the same coordinate system.

[0017] For climate data, data cleaning is used to remove missing and outlier values, and data interpolation algorithms are used to supplement a small number of missing time periods.

[0018] For building design code data, a text structuring algorithm is used to convert unstructured code clauses into recognizable item-based data;

[0019] For user-required text data, a text segmentation algorithm is used to break down long texts into semantic units and remove meaningless stop words.

[0020] Optionally, in a third implementation of the present invention, the extraction module includes:

[0021] The initialization submodule is used to initialize the multimodal feature enhancement network based on the Transformer attention mechanism and configure the network's hierarchical structure.

[0022] The allocation submodule is used to input standardized multi-source data into the network input layer, and the input layer allocates the data to each modality processing branch;

[0023] The computation submodule is used to perform attention calculations on single-modal data by each self-attention head after entering the Transformer encoder layer, by calculating the attention score of each element in the data with other elements;

[0024] The fusion submodule is used to obtain a single-modal feature matrix by concatenating and linearly transforming the calculation results of each attention head. The feature matrices of each modality are input into the feature fusion layer, and the residual connection algorithm is used to fuse the feature matrices of different modalities at the element level.

[0025] The processing submodule is used to convert the fused feature matrix into a feature vector through linear transformation and activation function processing of the fusion layer, thereby completing the extraction and fusion of target features from multi-source data and outputting the feature vector.

[0026] Optionally, in the fourth implementation of the present invention, the multimodal feature enhancement network is provided with 4 Transformer encoder layers, and each encoder layer is configured with 8 self-attention heads.

[0027] Optionally, in a fifth implementation of the present invention, the generation module includes:

[0028] The convolutional submodule is used to input feature vectors into the StyleGAN2-RL fusion generation model. The first convolutional layer of the StyleGAN2 generator converts the feature vectors into high-dimensional feature maps. Subsequent convolutional layers gradually reduce the dimensionality and refine the features. Each convolutional layer generates architectural geometric parameter features based on the feature information output by the previous layer. The last convolutional layer outputs a complete set of architectural design geometric parameters.

[0029] The receiving submodule is used by the RL module to receive the set of architectural design geometric parameters output by the StyleGAN2 generator and calculate the multi-objective reward item corresponding to each parameter scheme.

[0030] The comparison submodule is used to sum the multi-objective reward items according to preset weights to obtain the reward value of each scheme. Based on the PPO algorithm, it compares the reward value of the current scheme with the reward value of the historical scheme, calculates the advantage function of the policy network, and updates the generation policy of the StyleGAN2 generator.

[0031] The iteration submodule is used to filter the top 10 parameter schemes with the highest reward values ​​during the iteration process, convert them into initial architectural design schemes, and output them.

[0032] Optionally, in a sixth implementation of the present invention, the adjustment module includes:

[0033] The input submodule is used to input the initial architectural design scheme into the dynamic constraint optimizer, which then calls the data reading tool to extract the constraint parameters from the scheme.

[0034] The submodule is used to compare the extracted constraint parameters with the preset constraint compliance threshold, calculate the satisfaction of each constraint parameter, record the type and difference of the non-compliant constraint parameters, and generate a constraint satisfaction assessment report.

[0035] The execution submodule is used by the dynamic constraint optimizer to retrieve the initial constraint weights from the fuzzy logic rule base according to the project type, combine them with the constraint satisfaction evaluation report, and perform weight adjustment through fuzzy inference. After the adjustment is completed, a new constraint weight configuration table is generated.

[0036] The sending submodule is used by the dynamic constraint optimizer to send the new constraint weight configuration table to the generation module, which then drives the generation module to regenerate the architectural design geometric parameters based on the adjusted weights.

[0037] The output submodule is used to input the newly generated architectural design geometric parameters back into the dynamic constraint optimizer to evaluate the constraint satisfaction until all constraint parameters meet the threshold, and output the optimized design scheme that meets the constraint requirements.

[0038] Optionally, in a seventh implementation of the present invention, a method for implementing an AI-based automated building design system includes the following steps:

[0039] Collect multi-source data of the target building project and perform standardized preprocessing on the multi-source data, wherein the multi-source data includes terrain LiDAR data, climate data, building design code data and user requirement text data;

[0040] The preprocessed multi-source data is input into a multimodal feature enhancement network based on the Transformer attention mechanism to extract and fuse the target features of the multi-source data and output a feature vector.

[0041] The feature vector is input into the StyleGAN2-RL fusion generation model. The generation model generates architectural design geometric parameters through the improved StyleGAN2. The generation strategy is optimized by the RL module based on the PPO algorithm with multiple objective reward items such as structural stress compliance rate, energy consumption index and space utilization rate, and outputs the initial architectural design scheme.

[0042] The initial architectural design scheme is input into a dynamic constraint optimizer based on fuzzy logic. The dynamic constraint optimizer dynamically adjusts the weights of the design constraints according to the project type, performs constraint adaptation optimization on the initial architectural design scheme, and outputs an optimized design scheme that meets the constraint requirements.

[0043] The optimized design scheme is converted into a BIM model with professional interfaces. After verification through BIM collision detection, the final architectural design scheme is output.

[0044] Optionally, in the eighth implementation of the present invention, a method for implementing an AI-based automated building design system includes the following steps:

[0045] Initialize the multimodal feature enhancement network based on the Transformer attention mechanism and configure the network's hierarchical structure;

[0046] The standardized multi-source data is input into the network input layer, which then distributes the data to each modality processing branch.

[0047] After entering the Transformer encoder layer, each self-attention head performs attention calculation on the single modal data, and calculates the attention score of each element in the data with other elements.

[0048] The calculation results of each attention head are concatenated and linearly transformed to obtain a single-modality feature matrix. The feature matrices of each modality are input into the feature fusion layer, and the residual connection algorithm is used to fuse the feature matrices of different modalities at the element level.

[0049] Through linear transformation and activation function processing in the fusion layer, the fused feature matrix is ​​converted into a feature vector, thus completing the extraction and fusion of target features from multiple sources and outputting the feature vector.

[0050] Optionally, in a ninth implementation of the present invention, a method for implementing an AI-based automated building design system includes the following steps:

[0051] The initial architectural design scheme is input into the dynamic constraint optimizer, which then uses a data reading tool to extract the constraint parameters from the scheme.

[0052] The extracted constraint parameters are compared with the preset constraint compliance thresholds to calculate the satisfaction level of each constraint parameter. At the same time, the types of constraint parameters that do not meet the standards and the difference are recorded to form a constraint satisfaction assessment report.

[0053] The dynamic constraint optimizer retrieves the initial constraint weights from the fuzzy logic rule base based on the project type, combines them with the constraint satisfaction evaluation report, performs weight adjustment through fuzzy inference, and generates a new constraint weight configuration table after the adjustment is completed.

[0054] The dynamic constraint optimizer sends the new constraint weight configuration table to the generation module, which then drives the generation module to regenerate the architectural design geometric parameters based on the adjusted weights.

[0055] The newly generated architectural design geometric parameters are input again into the dynamic constraint optimizer to evaluate the constraint satisfaction until all constraint parameters meet the threshold, and then the optimized design scheme that meets the constraint requirements is output.

[0056] The technical solution provided by this invention involves collecting multi-source data of the target building project and performing standardized preprocessing on the multi-source data; inputting the preprocessed multi-source data into a multimodal feature enhancement network based on the Transformer attention mechanism to extract and fuse target features from the multi-source data, outputting feature vectors; inputting the feature vectors into a StyleGAN2-RL fusion generation model, which generates architectural design geometric parameters through an improved StyleGAN2, and optimizes the generation strategy with multiple objective reward items such as structural stress compliance rate, energy consumption index, and space utilization rate through an RL module based on the PPO algorithm, outputting an initial architectural design scheme; inputting the initial architectural design scheme into a dynamic constraint optimizer based on fuzzy logic, which dynamically adjusts the weights of design constraints according to the project type, performs constraint adaptation optimization on the initial architectural design scheme, and outputs an optimized design scheme that meets the constraint requirements; and converting the optimized design scheme into a BIM model containing professional interfaces. After BIM collision detection verification, the final architectural design scheme is output. This invention solves the problem of heterogeneous multi-source data formats and difficulty in collaboration, eliminating the need for manual data supplementation or correction, significantly reducing manual intervention in the data processing stage, accurately extracting multimodal key features, improving the adaptability of the scheme to site conditions, and solving the problem of multi-objective imbalance in traditional generated models. The generated initial scheme does not require separate structural verification and energy consumption calculation, while meeting the requirements of structural safety, energy saving, and space utilization, reducing professional rework after scheme generation. It dynamically adapts to the constraint requirements of different projects, ensuring that the optimized scheme prioritizes meeting the core constraints of the project, reducing the manual decision-making cost in the constraint adjustment stage, eliminating the need for manual parameter conversion or launching third-party software, realizing end-to-end automation from design scheme to implementable BIM model, and significantly shortening the design result implementation cycle. It comprehensively improves the accuracy, efficiency, and implementability of architectural design automation, providing intelligent technical support for the scheme design of various architectural projects. Attached Figure Description

[0057] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0058] Figure 1 A flowchart of an AI-based automated building design system provided in an embodiment of the present invention;

[0059] Figure 2 This is a schematic diagram of the extraction module provided in an embodiment of the present invention;

[0060] Figure 3 This is a schematic diagram of the generation module provided in an embodiment of the present invention. Detailed Implementation

[0061] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0062] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 A schematic diagram of the structure of an AI-based automated building design system provided in this embodiment of the invention. The system includes:

[0063] The data acquisition module is used to collect multi-source data of the target building project and perform standardized preprocessing on the multi-source data, which includes terrain LiDAR data, climate data, building design code data and user requirement text data.

[0064] The extraction module is used to input preprocessed multi-source data into a multimodal feature enhancement network based on the Transformer attention mechanism, extract and fuse target features from the multi-source data, and output feature vectors.

[0065] The generation module is used to input the feature vector into the StyleGAN2-RL fusion generation model. The generation model generates architectural design geometric parameters through the improved StyleGAN2. The generation strategy is optimized by the RL module based on the PPO algorithm with multiple objective reward items such as structural stress compliance rate, energy consumption index and space utilization rate, and outputs the initial architectural design scheme.

[0066] The adjustment module is used to input the initial architectural design scheme into the dynamic constraint optimizer based on fuzzy logic. The dynamic constraint optimizer dynamically adjusts the weight of the design constraints according to the project type, performs constraint adaptation optimization on the initial architectural design scheme, and outputs an optimized design scheme that meets the constraint requirements.

[0067] The output module is used to convert the optimized design scheme into a BIM model with professional interfaces. After verification by BIM clash detection, the final architectural design scheme is output.

[0068] In this embodiment, the data acquisition module includes:

[0069] The classification preprocessing submodule is used to classify and preprocess the collected multi-source data, convert the preprocessed terrain LiDAR data into raster data format, and normalize the data values ​​to the [0,1] interval according to the terrain slope and elevation range.

[0070] The first conversion submodule is used to convert the preprocessed climate data into energy efficiency impact correlation data and to map temperature and wind speed parameters according to climate zoning standards.

[0071] The second conversion submodule is used to convert structured building design code data into coded data, with each code requiring a unique identification code.

[0072] The third transformation submodule is used to convert the user's required text segmentation results into fixed-dimensional vectors through word embedding algorithms, and finally convert all types of data into tensors of the same dimension to complete the standardization preprocessing.

[0073] In this embodiment, the classification preprocessing submodule includes:

[0074] For terrain LiDAR data, a point cloud denoising algorithm is used to remove outliers caused by equipment errors and environmental interference. Then, a point cloud registration algorithm is used to unify the point cloud data collected at different times into the same coordinate system.

[0075] For climate data, data cleaning is used to remove missing and outlier values, and data interpolation algorithms are used to supplement a small number of missing time periods.

[0076] For building design code data, a text structuring algorithm is used to convert unstructured code clauses into recognizable item-based data;

[0077] For user-required text data, a text segmentation algorithm is used to break down long texts into semantic units and remove meaningless stop words.

[0078] In this embodiment, please refer to Figure 2 The extraction module includes:

[0079] The initialization submodule is used to initialize the multimodal feature enhancement network based on the Transformer attention mechanism and configure the network's hierarchical structure.

[0080] The allocation submodule is used to input standardized multi-source data into the network input layer, and the input layer allocates the data to each modality processing branch;

[0081] The computation submodule is used to perform attention calculation on the single modality data by each self-attention head after entering the Transformer encoder layer. It calculates the attention score between each element in the data and other elements. The multimodal feature enhancement network is set with 4 Transformer encoder layers, and each encoder layer is configured with 8 self-attention heads.

[0082] The fusion submodule is used to obtain a single-modal feature matrix by concatenating and linearly transforming the calculation results of each attention head. The feature matrices of each modality are input into the feature fusion layer, and the residual connection algorithm is used to fuse the feature matrices of different modalities at the element level.

[0083] The processing submodule is used to convert the fused feature matrix into a feature vector through linear transformation and activation function processing of the fusion layer, thereby completing the extraction and fusion of target features from multi-source data and outputting the feature vector.

[0084] In this embodiment, please refer to Figure 3 The generation module includes:

[0085] The convolutional submodule is used to input feature vectors into the StyleGAN2-RL fusion generation model. The first convolutional layer of the StyleGAN2 generator converts the feature vectors into high-dimensional feature maps. Subsequent convolutional layers gradually reduce the dimensionality and refine the features. Each convolutional layer generates architectural geometric parameter features based on the feature information output by the previous layer. The last convolutional layer outputs a complete set of architectural design geometric parameters.

[0086] The receiving submodule is used by the RL module to receive the set of architectural design geometric parameters output by the StyleGAN2 generator and calculate the multi-objective reward item corresponding to each parameter scheme.

[0087] The comparison submodule is used to sum the multi-objective reward items according to preset weights to obtain the reward value of each scheme. Based on the PPO algorithm, it compares the reward value of the current scheme with the reward value of the historical scheme, calculates the advantage function of the policy network, and updates the generation policy of the StyleGAN2 generator.

[0088] The iteration submodule is used to filter the top 10 parameter schemes with the highest reward values ​​during the iteration process, convert them into initial architectural design schemes, and output them.

[0089] In this embodiment, the adjustment module includes:

[0090] The input submodule is used to input the initial architectural design scheme into the dynamic constraint optimizer, which then calls the data reading tool to extract the constraint parameters from the scheme.

[0091] The submodule is used to compare the extracted constraint parameters with the preset constraint compliance threshold, calculate the satisfaction of each constraint parameter, record the type and difference of the non-compliant constraint parameters, and generate a constraint satisfaction assessment report.

[0092] The execution submodule is used by the dynamic constraint optimizer to retrieve the initial constraint weights from the fuzzy logic rule base according to the project type, combine them with the constraint satisfaction evaluation report, and perform weight adjustment through fuzzy inference. After the adjustment is completed, a new constraint weight configuration table is generated.

[0093] The sending submodule is used by the dynamic constraint optimizer to send the new constraint weight configuration table to the generation module, which then drives the generation module to regenerate the architectural design geometric parameters based on the adjusted weights.

[0094] The output submodule is used to input the newly generated architectural design geometric parameters back into the dynamic constraint optimizer to evaluate the constraint satisfaction until all constraint parameters meet the threshold, and output the optimized design scheme that meets the constraint requirements.

[0095] In this embodiment, the preset project type-constraint weight mapping rule library, such as plot ratio constraint weight = 0.4 and daylighting constraint weight = 0.2 when the project type is commercial building, and terrain adaptation constraint weight = 0.6 and plot ratio constraint weight = 0.2 when the project type is mountain residential, is imported into the dynamic constraint optimizer based on fuzzy logic. At the same time, constraint satisfaction evaluation indicators are set, such as plot ratio compliance threshold ≥ 3.0 and foundation depth adaptation threshold ≥ 1.5m. The fuzzy inference rules for initializing constraint weight adjustment are initialized. If the plot ratio does not meet the standard, the plot ratio constraint weight is increased.

[0096] The initial architectural design scheme output by the generation module is input into the dynamic constraint optimizer. The optimizer calls the data reading tool to extract key constraint parameters from the scheme, such as plot ratio, foundation depth, and daylighting duration. The extracted constraint parameters are compared with the preset constraint compliance thresholds, and the satisfaction of each constraint parameter is calculated. For example, when the plot ratio is 2.8, the satisfaction rate is 2.8 / 3.0≈93.3%. At the same time, the types of constraint parameters that do not meet the standards and the difference are recorded, and a constraint satisfaction assessment report is generated.

[0097] The optimizer retrieves initial constraint weights from the rule base based on the project type, and, combined with the constraint satisfaction assessment report, performs weight adjustments through fuzzy inference: If a constraint parameter fails to meet the standard, such as a commercial building's floor area ratio of 2.8 < 3.0, the fuzzy inference rule "Commercial building + floor area ratio fails to meet the standard → increase the floor area ratio constraint weight" is triggered, adjusting the floor area ratio constraint weight from the initial 0.4 to 0.5; if the terrain adaptation constraint meets the standard, such as a mountain residential building's foundation depth of 1.6m ≥ 1.5m, the terrain adaptation constraint weight remains unchanged at 0.6; after the adjustment is completed, a new constraint weight configuration table is generated.

[0098] The optimizer sends the new constraint weight configuration table to the generation module, which then regenerates the building design geometric parameters based on the adjusted weights. It focuses on correcting non-compliant constraint parameters, such as increasing the number of building floors from 15 to 18 to address insufficient floor area ratio. The newly generated scheme is then input into the optimizer again to evaluate constraint satisfaction. This evaluation-adjustment-optimization process is repeated until all constraint parameters meet the compliance threshold, at which point an optimized design scheme that meets the constraint requirements is output.

[0099] In this embodiment, the architectural design geometric parameters such as wall position, floor slab elevation, beam and column dimensions, and pipeline routing in the optimized design scheme output by the adjustment module are parsed to extract the professional attributes corresponding to each parameter, such as the beam cross-section dimensions of the structural profession and the pipe diameter of the mechanical and electrical profession. The API interface of the BIM platform such as Revit is called to convert the parsed parameters into a parameter format that the BIM platform can recognize, and establish a mapping relationship between the parameters and BIM model elements, such as the wall position parameter corresponding to the wall family instance position in the BIM model.

[0100] Based on the mapping relationship between parameters and model elements, the corresponding professional family libraries, such as structural beam families and MEP (Mechanical, Electrical, and Plumbing) families, are called through the BIM platform API interface to automatically create model element instances in the BIM platform: In terms of structural disciplines, structural elements such as beams, columns, and foundations are generated based on beam and column dimensions and foundation type parameters; in terms of MEP disciplines, water supply, drainage, and HVAC pipeline routing elements are generated based on pipeline routing and pipe diameter parameters; at the same time, parameter interfaces for each discipline, such as interfaces for adjusting the cross-sectional dimensions of structural beams and interfaces for modifying the diameter of MEP pipelines, are reserved in the BIM model to complete the BIM model construction.

[0101] Activate the clash detection function in the BIM platform, set the clash detection scope to structural beams and columns and mechanical and electrical pipes, between mechanical and electrical pipes, and between structural components, etc., of all professional model elements, and configure the clash detection accuracy such as the allowable spatial error range of ≤5mm; run the clash detection algorithm to compare the spatial positions of all elements in the BIM model, determine whether there is spatial overlap or collision between different elements, record the clash information such as the type, position, and overlap size of the colliding elements, and generate a clash detection report.

[0102] Analyze the clash detection report. If the report shows a clash rate ≤ 0.5%, which meets the acceptance standard, export the BIM model to a common format such as IFC format. At the same time, convert the design information in the BIM model into PDF design drawings, including floor plans, sections, and detailed drawings. If the clash rate > 0.5%, feed the clash detection report back to the adjustment module. The adjustment module will then optimize the design scheme. After the scheme is optimized, the conversion-construction-detection process will be executed again until the clash rate meets the standard. Finally, the BIM model and PDF design drawings will be output as the final architectural design scheme.

[0103] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based automated building design system, characterized in that, The system includes: The acquisition module is used to collect multi-source data of the target building project and perform standardized preprocessing on the multi-source data, wherein the multi-source data includes terrain LiDAR data, climate data, building design code data and user requirement text data. The extraction module is used to input preprocessed multi-source data into a multimodal feature enhancement network based on the Transformer attention mechanism, extract and fuse target features from the multi-source data, and output feature vectors. The generation module is used to input the feature vector into the StyleGAN2-RL fusion generation model. The generation model generates architectural design geometric parameters through the improved StyleGAN2 and optimizes the generation strategy with multiple objective reward items such as structural stress compliance rate, energy consumption index and space utilization rate through the RL module based on PPO algorithm, and outputs the initial architectural design scheme. The adjustment module is used to input the initial architectural design scheme into a dynamic constraint optimizer based on fuzzy logic. The dynamic constraint optimizer dynamically adjusts the weights of the design constraints according to the project type, performs constraint adaptation optimization on the initial architectural design scheme, and outputs an optimized design scheme that meets the constraint requirements. The output module is used to convert the optimized design scheme into a BIM model containing professional interfaces, and after verification by BIM collision detection, output the final architectural design scheme. The generation module includes: The convolutional submodule is used to input feature vectors into the StyleGAN2-RL fusion generation model. The first convolutional layer of the StyleGAN2 generator converts the feature vectors into high-dimensional feature maps. Subsequent convolutional layers gradually reduce the dimensionality and refine the features. Each convolutional layer generates architectural geometric parameter features based on the feature information output by the previous layer. The last convolutional layer outputs a complete set of architectural design geometric parameters. The receiving submodule is used by the RL module to receive the set of architectural design geometric parameters output by the StyleGAN2 generator and calculate the multi-objective reward item corresponding to each parameter scheme. The comparison submodule is used to sum the multi-objective reward items according to preset weights to obtain the reward value of each scheme. Based on the PPO algorithm, it compares the reward value of the current scheme with the reward value of the historical scheme, calculates the advantage function of the policy network, and updates the generation policy of the StyleGAN2 generator. The iteration submodule is used to filter the top 10 parameter schemes with the highest reward values ​​during the iteration process, convert them into initial architectural design schemes, and output them. The adjustment module includes: The input submodule is used to input the initial architectural design scheme into the dynamic constraint optimizer, which then calls the data reading tool to extract the constraint parameters from the scheme. The submodule is used to compare the extracted constraint parameters with the preset constraint compliance threshold, calculate the satisfaction of each constraint parameter, record the type and difference of the non-compliant constraint parameters, and generate a constraint satisfaction assessment report. The execution submodule is used by the dynamic constraint optimizer to retrieve the initial constraint weights from the fuzzy logic rule base according to the project type, combine them with the constraint satisfaction evaluation report, and perform weight adjustment through fuzzy inference. After the adjustment is completed, a new constraint weight configuration table is generated. The sending submodule is used by the dynamic constraint optimizer to send the new constraint weight configuration table to the generation module, which then drives the generation module to regenerate the architectural design geometric parameters based on the adjusted weights. The output submodule is used to input the newly generated architectural design geometric parameters back into the dynamic constraint optimizer to evaluate the constraint satisfaction until all constraint parameters meet the threshold, and output the optimized design scheme that meets the constraint requirements.

2. The AI-based automated building design system as described in claim 1, characterized in that, The acquisition module includes: The classification preprocessing submodule is used to classify and preprocess the collected multi-source data, convert the preprocessed terrain LiDAR data into raster data format, and normalize the data values ​​to the [0,1] interval according to the terrain slope and elevation range. The first conversion submodule is used to convert the preprocessed climate data into energy efficiency impact correlation data and to map temperature and wind speed parameters according to climate zoning standards. The second conversion submodule is used to convert structured building design code data into coded data, with each code requiring a unique identification code. The third transformation submodule is used to convert the user's required text segmentation results into fixed-dimensional vectors through word embedding algorithms, and finally convert all types of data into tensors of the same dimension to complete the standardization preprocessing.

3. The AI-based automated building design system as described in claim 2, characterized in that, The classification preprocessing submodule includes: For terrain LiDAR data, a point cloud denoising algorithm is used to remove outliers caused by equipment errors and environmental interference. Then, a point cloud registration algorithm is used to unify the point cloud data collected at different times into the same coordinate system. For climate data, data cleaning is used to remove missing and outlier values, and data interpolation algorithms are used to supplement a small number of missing time periods. For building design code data, a text structuring algorithm is used to convert unstructured code clauses into recognizable item-based data; For user-required text data, a text segmentation algorithm is used to break down long texts into semantic units and remove meaningless stop words.

4. The AI-based automated building design system as described in claim 1, characterized in that, The extraction module includes: The initialization submodule is used to initialize the multimodal feature enhancement network based on the Transformer attention mechanism and configure the network's hierarchical structure. The allocation submodule is used to input standardized multi-source data into the network input layer, and the input layer allocates the data to each modality processing branch; The computation submodule is used to perform attention calculations on single-modal data by each self-attention head after entering the Transformer encoder layer, by calculating the attention score of each element in the data with other elements; The fusion submodule is used to obtain a single-modal feature matrix by concatenating and linearly transforming the calculation results of each attention head. The feature matrices of each modality are input into the feature fusion layer, and the residual connection algorithm is used to fuse the feature matrices of different modalities at the element level. The processing submodule is used to convert the fused feature matrix into a feature vector through linear transformation and activation function processing of the fusion layer, thereby completing the extraction and fusion of target features from multi-source data and outputting the feature vector.

5. The AI-based automated building design system as described in claim 4, characterized in that, The multimodal feature enhancement network consists of four Transformer encoder layers, each with eight self-attention heads.

6. A method for implementing the artificial intelligence-based automated building design system as described in claim 1, characterized in that, The method includes the following steps: Collect multi-source data of the target building project and perform standardized preprocessing on the multi-source data, wherein the multi-source data includes terrain LiDAR data, climate data, building design code data and user requirement text data; The preprocessed multi-source data is input into a multimodal feature enhancement network based on the Transformer attention mechanism to extract and fuse the target features of the multi-source data and output a feature vector. The feature vectors are input into the StyleGAN2-RL fusion generation model. This model generates architectural design geometric parameters using an improved StyleGAN2. An RL module based on the PPO algorithm optimizes the generation strategy with multiple objective rewards, including structural stress compliance rate, energy consumption index, and space utilization rate, outputting an initial architectural design scheme. The first convolutional layer of the StyleGAN2 generator converts the feature vectors into high-dimensional feature maps. Subsequent convolutional layers progressively reduce dimensionality and refine the features. Each convolutional layer is based on the features output from the previous layer. The system generates architectural geometric parameter features from information, and the last convolutional layer outputs a complete set of architectural design geometric parameters. The RL module receives the set of architectural design geometric parameters output by the StyleGAN2 generator, calculates the multi-objective reward item corresponding to each parameter scheme, sums the multi-objective reward items according to preset weights to obtain the reward value of each scheme, compares the reward value of the current scheme with the reward values ​​of historical schemes based on the PPO algorithm, calculates the advantage function of the policy network, and updates the generation strategy of the StyleGAN2 generator. The top 10 parameter schemes with the highest reward values ​​during the iteration process are selected, converted into initial architectural design schemes, and output. The initial architectural design scheme is input into a fuzzy logic-based dynamic constraint optimizer. The dynamic constraint optimizer dynamically adjusts the weights of design constraints according to the project type, performing constraint adaptation optimization on the initial architectural design scheme and outputting an optimized design scheme that meets the constraint requirements. The process involves: inputting the initial architectural design scheme into the dynamic constraint optimizer; the dynamic constraint optimizer using a data reading tool to extract constraint parameters from the scheme; comparing the extracted constraint parameters with preset constraint compliance thresholds, calculating the compliance degree of each constraint parameter, and recording the types and differences of non-compliant constraint parameters to form a constraint compliance assessment report; the dynamic constraint optimizer retrieving initial constraint weights from the fuzzy logic rule base according to the project type, combining them with the constraint compliance assessment report, and performing weight adjustments through fuzzy inference; after adjustment, generating a new constraint weight configuration table; the dynamic constraint optimizer sending the new constraint weight configuration table to the generation module, driving the generation module to regenerate the architectural design geometric parameters based on the adjusted weights; and inputting the newly generated architectural design geometric parameters back into the dynamic constraint optimizer to evaluate constraint compliance until all constraint parameters meet the compliance thresholds, outputting an optimized design scheme that meets the constraint requirements. The optimized design scheme is converted into a BIM model with professional interfaces. After verification through BIM collision detection, the final architectural design scheme is output.

7. A method for implementing the artificial intelligence-based automated building design system as described in claim 1, characterized in that, The method includes the following steps: Initialize the multimodal feature enhancement network based on the Transformer attention mechanism and configure the network's hierarchical structure; The standardized multi-source data is input into the network input layer, which then distributes the data to each modality processing branch. After entering the Transformer encoder layer, each self-attention head performs attention calculation on the single modal data, and calculates the attention score of each element in the data with other elements. The calculation results of each attention head are concatenated and linearly transformed to obtain a single-modality feature matrix. The feature matrices of each modality are input into the feature fusion layer, and the residual connection algorithm is used to fuse the feature matrices of different modalities at the element level. Through linear transformation and activation function processing in the fusion layer, the fused feature matrix is ​​converted into a feature vector, thus completing the extraction and fusion of target features from multiple sources and outputting the feature vector.

Images