Building feature extraction and mapping method based on BIM technology

Abstract

The invention discloses a building feature extraction and mapping method based on a BIM technology, and the method comprises the following steps: constructing a three-level building feature system comprising microscopic, mesoscopic and macroscopic features, and carrying out the association with a BIM model component through a unique ID; exporting a BIM model and extracting feature data to form an original sample set; constructing a heterogeneous sample library through data enhancement; performing three-direction scanning along the X axis, the Y axis and the Z axis to generate a series of profile maps, plane maps and three-dimensional point cloud models, and performing standardized preprocessing; inputting the BIM parameters and the multi-view features into a two-channel fusion network, and outputting a high-dimensional feature tensor through tensor fusion and decomposition; converting the feature tensor into a vector graph through a mapping module; and carrying out pre-training and fine tuning on the network by adopting a mixed loss function, and finally realizing lightweight deployment. According to the method, efficient and automatic generation from the BIM model to the standard engineering drawing is realized, and the precision and efficiency of building information extraction and mapping are improved.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of building information technology and artificial intelligence, and specifically discloses a method for building feature extraction and mapping based on BIM technology. Background Technology

[0002] Building Information Modeling (BIM) technology, as the core of digital transformation in the construction industry, has been widely applied throughout the entire lifecycle of building design, construction, and operation and maintenance. BIM models contain rich geometric and non-geometric attribute information, providing strong support for project visualization, coordination analysis, and data management.

[0003] However, despite the significant progress made in BIM technology, there are still technical bottlenecks and challenges in the automated and intelligent conversion from BIM models to industry-standard 2D engineering drawings. These challenges are manifested in the following aspects: Insufficient depth of information extraction hinders advanced applications: Traditional BIM data utilization is mostly concentrated on basic applications such as model browsing, clash detection, and quantity surveying. Existing methods typically only extract basic component attributes, lacking a systematic and structured extraction and expression of multi-scale features of building space. This "rich model, poor feature" situation severely restricts the development of advanced BIM-based applications such as intelligent review, automated drawing generation, performance analysis, and optimization.

[0004] The automated drawing generation capability is weak, heavily reliant on manual intervention: Currently, the process of generating construction drawings from BIM models still largely depends on manual adjustments and revisions by designers. The generated views often suffer from information redundancy, missing annotations, and non-standard layouts, failing to directly meet national drafting standards. This process is not only inefficient and prone to human error, but also creates an information gap between the BIM model and the final delivered drawings, violating the original intention of BIM technology to be versatile.

[0005] The challenges of fusing and understanding multi-source heterogeneous data: A BIM model is a complex information carrier integrating geometric, attribute, and relational data from multiple sources. Current technologies lack effective means to deeply fuse and correlate BIM parametric data with geometric data such as multi-views and 3D point clouds. Relying solely on geometric or attribute information makes it difficult to fully and accurately understand architectural intent and generate semantically correct drawings, thus compromising the accuracy and reliability of automated drawing generation.

[0006] The generalization and intelligence levels of the methods need improvement: Many studies have attempted to automate map generation through rule bases or template matching, but these methods lack flexibility and are difficult to adapt to different design codes, building types, and complex and ever-changing architectural scenarios. The lack of end-to-end feature learning and mapping mechanisms based on artificial intelligence, especially deep learning, results in insufficient intelligence and adaptability of existing methods.

[0007] Therefore, it is necessary to invent a method for extracting and mapping building features based on BIM technology to solve the above problems. Summary of the Invention

[0008] To overcome the aforementioned deficiencies in existing technologies, this invention provides a method for building feature extraction and mapping based on BIM technology. This method constructs a three-tiered building feature system encompassing micro, meso, and macro features, and associates these features with BIM model components using unique IDs. It then exports the BIM model and extracts feature data to form an initial sample set. Through data augmentation, a heterogeneous sample library is constructed. A three-dimensional scan along the XYZ axes generates a series of cross-sectional views, plan views, and 3D point cloud models, which are then preprocessed using standardization. BIM parameters and multi-view features are input into a dual-channel fusion network, which outputs high-dimensional feature tensors through tensor fusion and decomposition. The feature tensors are converted into vector graphics using a mapping module. Finally, a hybrid loss function is used to pre-train and fine-tune the network, ultimately achieving lightweight deployment and effectively solving the problems mentioned in the background technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for extracting and mapping building features based on BIM technology, specifically including the following steps: S1. Construct a three-level building feature system that includes micro-features, meso-features, and macro-features, and define a set of tags that are associated with BIM model components through unique IDs according to the end application scenario; S2. Export the BIM model file from the BIM platform, extract the feature data from the three-level feature system, and form the original sample set; S3. The original sample set is augmented through geometric transformation, attribute perturbation and scene combination, and the augmented sample data is stored in a distributed database to construct a heterogeneous sample library. S4. Based on the spatial coordinate system of the BIM model, set the scanning parameters, perform three-way scanning along the XYZ axes, generate a series of cross-sectional views, plan views and three-dimensional point cloud models, and perform standardized preprocessing on the generated views; S5. Input BIM parameter features and multi-view features into a dual-channel fusion network. Map the dual-channel output features into a high-dimensional tensor through a tensor fusion layer. Perform dimensional compression and feature association through tensor decomposition to output a high-dimensional feature tensor. S6. Convert the high-dimensional feature tensor into vector graphics data in SVG format through the mapping module; S7. Construct a training dataset and design a hybrid loss function. Use pre-training and fine-tuning strategies to train the dual-channel fusion network. Implement lightweight processing through model pruning and convert it to ONNX format for deployment.

[0010] Preferably, the micro-features include component node features and edge features, the meso-features include component attribute features and spatial relationship features, and the macro-features include overall structural system features and spatial layout features.

[0011] Preferably, the geometric transformation includes scaling, rotating, and mirroring the BIM model; the attribute perturbation includes randomly adjusting component parameters within the limits allowed by the specifications; and the scene combination includes combining standard floors or components of different buildings to generate new samples.

[0012] Preferably, the three-dimensional scanning is performed along the XYZ axes with a set step size to generate a series of cross-sectional views in the YZ plane, a series of cross-sectional views in the XZ plane, and a series of planar views in the XY plane, and simultaneously saves the three-dimensional point cloud model.

[0013] Preferably, the dual-channel converged network includes: BIM parameter feature channel, used to filter key attribute features through fully connected layers and attention mechanisms; Multi-view feature channels are used to extract spatially continuous features using 3D-CNN; The tensor fusion layer is used to map the dual-channel output features into high-dimensional tensors and perform dimensionality compression through tensor decomposition.

[0014] Preferably, the training of the dual-channel fusion network adopts a hybrid loss function: Ls=α×Lf+β×Lg+γ×Lr, where Lf is the feature extraction MSE loss, Lg is the graph generation adversarial loss, Lr is the L2 regularization loss, and α, β, and γ are weight coefficients.

[0015] Preferably, the pre-training strategy used in step S7 includes unsupervised pre-training on a large-scale building BIM dataset, and supervised learning using labeled samples in the fine-tuning stage.

[0016] The technical effects and advantages of this invention are as follows: 1. By constructing a three-level building feature system that includes micro, meso, and macro levels, and associating it with BIM components through unique IDs, we have achieved comprehensive and complete feature information capture from component details to overall layout. This overcomes the shortcomings of traditional methods, such as single feature dimensions and fragmented information, and provides complete, structured, and semantically rich underlying data support for subsequent in-depth analysis and automated mapping. 2. Data augmentation is performed using various methods such as geometric transformation, attribute perturbation, and scene combination, and a heterogeneous sample library is constructed, which greatly expands the quantity and diversity of the training dataset; effectively avoiding the overfitting problem caused by insufficient training data or single scene in deep learning models, making the trained model more adaptable and accurate to architectural projects of different styles and scales. 3. Innovatively adopting a dual-channel fusion network to simultaneously process BIM parametric features and multi-view geometric features, and using tensor fusion and decomposition technology for efficient feature association and compression; fully utilizing the advantages of BIM model information-geometry integration, combining abstract attribute parameters with intuitive spatial forms, so that the final extracted feature tensors can reflect the physical properties of components and express their complex spatial relationships, with feature expression capabilities far exceeding those of methods using a single data source. 4. Training is guided by a hybrid loss function, combined with a pre-training-fine-tuning strategy, and lightweighted through model pruning. Finally, the model is converted to ONNX format for deployment. The hybrid loss function ensures the optimal balance between feature extraction and image generation. Lightweighting and ONNX format enable the trained model to be deployed efficiently and cross-platform in real-world engineering environments, meeting the requirements of practical applications for computational efficiency and ease of deployment. Attached Figure Description

[0017] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the overall process flow of the present invention.

[0019] Figure 2 This is a structural diagram of the dual-channel fusion network of the present invention. Detailed Implementation

[0020] 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.

[0021] This invention provides, for example Figure 1 The method for extracting and mapping building features based on BIM technology, as shown, specifically includes the following steps: S1. Construct a three-level building feature system that includes micro-features, meso-features, and macro-features, and define a set of tags that are associated with BIM model components through unique IDs according to the end application scenario; It should be further explained that BIM is an abbreviation for Building Information Modeling.

[0022] S2. Export the BIM model file from the BIM platform, extract the feature data from the three-level feature system, and form the original sample set; S3. The original sample set is augmented through geometric transformation, attribute perturbation and scene combination, and the augmented sample data is stored in a distributed database to construct a heterogeneous sample library. S4. Based on the spatial coordinate system of the BIM model, set the scanning parameters, perform three-way scanning along the XYZ axes, generate a series of cross-sectional views, plan views and three-dimensional point cloud models, and perform standardized preprocessing on the generated views; S5. Input BIM parameter features and multi-view features into a dual-channel fusion network. Map the dual-channel output features into a high-dimensional tensor through a tensor fusion layer. Perform dimensional compression and feature association through tensor decomposition to output a high-dimensional feature tensor. S6. Convert the high-dimensional feature tensor into vector graphics data in SVG format through the mapping module; It should be further noted that SVG is short for Scalable Vector Graphics.

[0023] S7. Construct a training dataset and design a hybrid loss function. Use pre-training and fine-tuning strategies to train the dual-channel fusion network. Implement lightweight processing through model pruning and convert it to ONNX format for deployment.

[0024] It should be further noted that ONNX is short for Open Neural Network Exchange, which means open neural network exchange.

[0025] Furthermore, in the above technical solution, the micro-features mentioned in step S1 include component node features and edge features, the meso-features include component attribute features and spatial relationship features, and the macro-features include overall structural system features and spatial layout features; Furthermore, the construction of the three-level building feature system described in step S1 is based on the international IFC standard, the national "Building Information Modeling Classification and Coding Standard", and the industry-standard component classification rules to ensure the universality and standardization of the feature system; Furthermore, the specific implementation scheme of step S1 is as follows: S101. Constructing a three-tiered architectural feature system: Micro-features: These include the node features of components in the BIM model, typically including connection points, endpoints, and edge features. Specific implementations include: beam-column connection points and wall-floor intersection lines.

[0026] Mesoscopic features: These include the attribute characteristics of components, typically including materials, dimensions, type, and spatial relationship characteristics, such as relative position and adjacency. Specific implementations include: wall thickness, opening direction of doors and windows, and topological relationships between components.

[0027] Macro-level characteristics include the overall structural system characteristics, typically including structural type, load transfer path, and spatial layout characteristics, such as functional zoning and circulation organization. Specific implementations include: frame structure vs. shear wall structure, floor plan layout, and space utilization efficiency.

[0028] S102. Define a tag set and associate it with BIM components: Based on the end-user application scenario, a tagging system is defined; each tag corresponds to one or more components in the BIM model and is associated with a globally unique ID.

[0029] Furthermore, the terminal application scenarios include, but are not limited to: construction drawing output, engineering quantity calculation, collision detection, energy-saving analysis, compliance review, and facility management; Furthermore, the tag set is stored in JSON or XML format, and each tag includes: tag name, attribute key-value pair, and application scenario identifier. It is associated with the globally unique ID tag of the BIM component and stored in a relational database.

[0030] Tag content example: Component types, such as structural columns and water supply pipes; Functional attributes, such as load-bearing or non-load-bearing; Spatial classification, such as bedroom, hallway; Drafting standard labels, such as marking visible and hidden lines; Furthermore, in the above technical solution, the specific implementation of step S2 is as follows: The sample collection process involves exporting model files of similar buildings from BIM platforms such as Revit and ArchiCAD, extracting the feature data defined in step S1, and forming the original sample set.

[0031] Furthermore, in the above technical solution, the geometric transformation in step S3 includes scaling, rotating, and mirroring the BIM model; the attribute perturbation includes randomly adjusting component parameters within the allowable range of the specifications; and the scene combination includes combining standard floors or components of different buildings to generate new samples.

[0032] Furthermore, the specific implementation methods for sample enhancement are as follows: Geometric transformations: scaling, rotating, and mirroring the BIM model to simulate building structures at different scales; Attribute disturbance: Randomly adjust component parameters within the allowable range of the specification, such as concrete strength grade ±1 and wall thickness ±50mm; Scene combination: Combine standard floors and components of different buildings into new samples to simulate complex building forms.

[0033] The sample library uses a distributed database for storage, and each sample contains a model ID, BIM file path, feature data, label set, and scan data index.

[0034] Furthermore, the typical setting for randomly adjusting component parameters within the allowable range of the specification in the attribute disturbance is that the concrete strength grade fluctuates by no more than one grade based on the design value, and the allowable deviation of the wall thickness is ±10mm to ±50mm. Furthermore, in the above technical solution, the three-dimensional scanning in step S4 is a scanning along the XYZ axis with a set step size to generate a series of cross-sectional views of the YZ plane, a series of cross-sectional views of the XZ plane, and a series of planar views of the XY plane, and simultaneously saves the three-dimensional point cloud model.

[0035] Furthermore, the specific implementation of the three-dimensional scanning is as follows: S401. Based on the spatial coordinate system of the BIM model, set the scanning parameters: Scan step size: Set according to the building's precision requirements; for typical construction scenarios, set to 100mm / step. View range: Automatically identifies the model bounding box and determines the start and end coordinates of the scan; Layer filtering: Preserve structural components and hide non-feature elements such as guide lines and labels.

[0036] S402. Multi-view generation: Scan along the X-axis: Generate a series of YZ planar cross-sectional views, reflecting the building's depth characteristics; Scan along the Y-axis: Generate a series of XZ planar cross-sectional views, reflecting the characteristics of the building's width direction; Scan along the Z-axis: Generate a series of XY plan views, reflecting the floor plan layout of each floor; The 3D point cloud model is saved synchronously, preserving the spatial coordinate information of the components.

[0037] S403. View Preprocessing: Size normalization: uniformly scaled to 512×512 pixels; Feature enhancement: The Canny edge detection algorithm is used to highlight the component outline; Data format conversion: Convert the image into tensor form (H×W×C) to serve as input for the dual-channel fusion network; Furthermore, in the tensor form (H×W×C), H represents the height, which is the number of pixels in the vertical direction of the tensor; W represents the width, which is the number of pixels in the horizontal direction of the tensor; and C represents the number of channels, which represents the information dimension contained in each pixel.

[0038] Furthermore, in the above technical solution, the dual-channel fusion network in step S5 includes: BIM parameter feature channel, used to filter key attribute features through fully connected layers and attention mechanisms; Multi-view feature channels are used to extract spatially continuous features using 3D-CNN; It should be further explained that 3D-CNN is short for 3D Convolutional Neural Network, which means three-dimensional convolutional neural network.

[0039] The tensor fusion layer is used to map the dual-channel output features into high-dimensional tensors and perform dimensionality compression through tensor decomposition.

[0040] Furthermore, the structure of the dual-channel fusion network is as follows: Figure 2 As shown.

[0041] It should be further explained that 3D-CNN was chosen to process the series of cross-sectional and plan views in the multi-view feature channel because 3D-CNN can effectively capture the spatial context information between consecutive views, making it suitable for extracting the spatial continuity features of building components. For 3D point cloud data, PointNet++ was chosen instead of traditional CNN because PointNet++ can directly process unstructured point cloud data and has permutation invariance and rotation invariance, making it more suitable for extracting features of building components with complex spatial relationships.

[0042] Furthermore, the specific implementation plan for step S5 is as follows: The BIM parameter feature channel takes BIM parameter features from a heterogeneous sample library as input and performs preliminary feature extraction through a fully connected layer (FC). An attention mechanism, specifically Self-Attention, is introduced to filter key attributes and output a high-dimensional attribute feature vector.

[0043] The multi-view feature channel takes the image and point cloud data generated by the three-way scan as input, uses 3D-CNN to extract spatial continuous features, uses PointNet++ to extract features from the point cloud data, and outputs a high-dimensional geometric feature tensor.

[0044] The tensor fusion layer takes high-dimensional attribute feature vectors and high-dimensional geometric feature tensors as inputs, and fuses the attribute feature vectors and geometric feature tensors, such as outer product, concatenation or attention weighted fusion, to map them into a unified high-dimensional tensor. Tensor decomposition techniques, such as CP decomposition, are used to compress dimensions and associate features, and outputs a compressed high-dimensional feature tensor that has both semantic and geometric information. Furthermore, in the above technical solution, the mapping module in step S6 learns the mapping relationship between feature tensors and graphic elements through generative adversarial networks, and outputs vector graphics in SVG format that conform to the "Unified Standard for Building Drawing".

[0045] Furthermore, the mapping module uses a generative adversarial network as its core architecture. The generator receives high-dimensional feature tensors, upsamples them step by step, and decodes them into graphic elements, such as lines, labels, and filled areas. The discriminator judges whether the generated graphics conform to the distribution of real engineering drawings and pushes the generator to output graphics that are more in line with the standard. Furthermore, in the above technical solution, the training of the dual-channel fusion network in step S7 adopts a hybrid loss function: Ls=α×Lf+β×Lg+γ×Lr, where Lf is the feature extraction MSE loss, Lg is the graph generation adversarial loss, Lr is the L2 regularization loss, and α, β, and γ are weight coefficients.

[0046] It should be further explained that MSE is short for Mean Square Error.

[0047] Furthermore, the weight coefficients α, β, and γ in the hybrid loss function are tuned on the validation set using Bayesian optimization. Typical values ​​are: α ∈ [0.4, 0.6], β ∈ [0.3, 0.5], γ ∈ [0.1, 0.2], and satisfy α + β + γ = 1. This setting aims to balance feature reconstruction accuracy, image generation quality, and model generalization ability.

[0048] Furthermore, the pre-training strategy adopted in step S7 includes unsupervised pre-training on a large-scale building BIM dataset, and supervised learning using labeled samples in the fine-tuning stage.

[0049] Furthermore, the specific implementation scheme of step S7 is as follows: S701. Use the heterogeneous sample library constructed after data augmentation in step S3 as the source of training data. Each sample contains: model ID, three-level feature data extracted from BIM file path, label set, and scan data index. S702. Uses a hybrid loss function: Ls=α×Lf+β×Lg+γ×Lr, to guide training; S703. Unsupervised pre-training is performed using a large-scale unlabeled BIM dataset to learn a general building feature representation. Supervised learning is performed using labeled samples. Fine-tuning is performed for specific mapping tasks. During fine-tuning, some low-level parameters are fixed and only the weights of high-level networks are updated to improve training efficiency and generalization ability. S704. Use model pruning techniques to remove connections / channels with small weights or low contributions from the network, while retaining the ability to extract key features and reducing the number of model parameters and computational cost. S705. Convert the trained dual-channel fusion network to ONNX format to ensure that the model can perform efficient inference on different platforms.

[0050] Furthermore, the large-scale building BIM dataset is constructed by collecting model files from publicly available BIM model libraries, such as BIMobject and National BIM Library. Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting and mapping building features based on BIM technology, characterized in that, Specifically, the following steps are included: S1. Construct a three-level building feature system that includes micro-features, meso-features, and macro-features, and define a set of tags that are associated with BIM model components through unique IDs according to the end application scenario; S2. Export the BIM model file from the BIM platform, extract the feature data from the three-level feature system, and form the original sample set; S3. The original sample set is augmented through geometric transformation, attribute perturbation and scene combination, and the augmented sample data is stored in a distributed database to construct a heterogeneous sample library. S4. Based on the spatial coordinate system of the BIM model, set the scanning parameters, perform three-way scanning along the XYZ axes, generate a series of cross-sectional views, plan views and three-dimensional point cloud models, and perform standardized preprocessing on the generated views; S5. Input BIM parameter features and multi-view features into a dual-channel fusion network. Map the dual-channel output features into a high-dimensional tensor through a tensor fusion layer. Perform dimensional compression and feature association through tensor decomposition to output a high-dimensional feature tensor. S6. Convert the high-dimensional feature tensor into vector graphics data in SVG format through the mapping module; S7. Construct a training dataset and design a hybrid loss function. Use pre-training and fine-tuning strategies to train the dual-channel fusion network. Implement lightweight processing through model pruning and convert it to ONNX format for deployment.

2. The method for extracting and mapping building features based on BIM technology as described in claim 1, characterized in that: The micro-features include component node features and edge features, the meso-features include component attribute features and spatial relationship features, and the macro-features include overall structural system features and spatial layout features.

3. The method for extracting and mapping building features based on BIM technology as described in claim 1, characterized in that: The geometric transformations include scaling, rotating, and mirroring the BIM model; the attribute perturbations include randomly adjusting component parameters within the limits allowed by the specifications; and the scene combination includes combining standard floors or components of different buildings to generate new samples.

4. The method for extracting and mapping building features based on BIM technology as described in claim 1, characterized in that: The three-dimensional scanning is performed along the XYZ axes with a set step size to generate a series of cross-sectional views in the YZ plane, a series of cross-sectional views in the XZ plane, and a series of planar views in the XY plane, and simultaneously saves the three-dimensional point cloud model.

5. The method for extracting and mapping building features based on BIM technology as described in claim 1, characterized in that: The dual-channel fusion network includes: BIM parameter feature channel, used to filter key attribute features through fully connected layers and attention mechanisms; Multi-view feature channels are used to extract spatially continuous features using 3D-CNN; The tensor fusion layer is used to map the dual-channel output features into high-dimensional tensors and perform dimensionality compression through tensor decomposition.

6. The method for extracting and mapping building features based on BIM technology as described in claim 1, characterized in that: The training of the dual-channel fusion network adopts a hybrid loss function: Ls=α×Lf+β×Lg+γ×Lr, where Lf is the feature extraction MSE loss, Lg is the graph generation adversarial loss, Lr is the L2 regularization loss, and α, β, and γ are weight coefficients.

7. The method for extracting and mapping building features based on BIM technology as described in claim 1, characterized in that: The pre-training strategy used in step S7 includes unsupervised pre-training on a large-scale building BIM dataset, and supervised learning using labeled samples during the fine-tuning stage.

Images