Building feature extraction and mapping method based on BIM technology

Abstract

The application discloses a kind of based on BIM technology's building feature extraction and mapping method, comprising the following steps: constructing three-level building feature system including microcosmic, mesoscopic and macroscopic features, and being associated with BIM model component by unique ID;BIM model is exported and feature data is extracted, and original sample set is formed;By data enhancement, construct heterogeneous sample library;Three-way scanning is carried out along XYZ axis, and series of profile, plan and three-dimensional point cloud model are generated, and standardized pretreatment is carried out;BIM parameters and multi-view features are input into dual-channel fusion network, and high-dimensional feature tensor is output after tensor fusion and decomposition;Feature tensor is converted into vector graphics by mapping module;Mixed loss function is used to pretrain and fine-tune the network, and finally realize lightweight deployment.The method realizes efficient and automated generation from BIM model to standard engineering drawings, and improves the precision and efficiency of building information extraction and mapping.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of building information technology and artificial intelligence, and specifically discloses a building feature extraction and mapping method based on BIM technology. BACKGROUND

[0002] As the core of the digital transformation of the construction industry, the building information modeling (BIM) technology has been widely used in the whole life cycle of building design, construction and operation. The BIM model contains rich geometric information and non-geometric attribute information, which provides strong support for visualization, coordination analysis and data management of projects.

[0003] However, although the BIM technology has made great progress, there are still technical bottlenecks and challenges in the field of automatic and intelligent conversion from BIM models to two-dimensional engineering drawings conforming to industry standards, which are specifically manifested in the following aspects:

[0004] Insufficient information extraction depth, difficult to support advanced applications: Traditional BIM data utilization is mainly focused on model browsing, collision checking, engineering quantity statistics and other basic applications. Existing methods can only extract the basic attributes of components, and lack systematic and structured extraction and expression of multi-scale features of building space. This "rich model, poor feature" situation seriously restricts the development of advanced applications such as intelligent review, automatic drawing, performance analysis and optimization based on BIM.

[0005] Weak automatic mapping capability, heavily dependent on manual intervention: At present, the process of generating construction drawings from BIM models still largely relies on manual adjustment and modification by designers. The generated views often have information redundancy, missing labels, and non-standard layout, which cannot directly meet the national drawing standards. This process not only is inefficient and prone to human error, but also causes information gaps between BIM models and final delivery drawings, which violates the original intention of BIM technology.

[0006] Difficult to fuse and understand multi-source heterogeneous data: BIM model itself is a complex information carrier integrating geometry, attributes, relationships and other multi-source data. Existing technologies lack effective means to deeply fuse and correlate BIM parameter data with multi-view and three-dimensional point cloud geometry data of the model. Simply relying on geometric information or attribute information is difficult to fully and accurately understand the building intent and generate semantically correct drawings, resulting in difficulty in ensuring the accuracy and reliability of automatic mapping.

[0007] Method generalization and intelligence level need to be improved: Many studies attempt to achieve automatic mapping by rule base or template matching, but these methods have poor flexibility and are difficult to adapt to different design specifications, building types and complex building scenes. The lack of end-to-end feature learning and mapping mechanism based on artificial intelligence, especially deep learning, makes the intelligence level and adaptive ability of existing methods insufficient.

[0008] Therefore, it is necessary to invent a building feature extraction and mapping method based on BIM technology to solve the above problems. SUMMARY

[0009] In order to overcome the defects of the prior art, the present application provides a building feature extraction and mapping method based on BIM technology, which comprises the following steps:

[0010] To achieve the above purpose, the present application provides the following technical scheme: a building feature extraction and mapping method based on BIM technology, specifically comprising the following steps:

[0011] S1, a three-level building feature system including micro features, meso features and macro features is constructed, and a label set associated with BIM model components through unique ID is defined according to end application scenarios;

[0012] 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;

[0013] S2, the BIM model file is exported from the BIM platform, and the feature data in the three-level building feature system is extracted to form an original sample set;

[0014] S3, the original sample set is data enhanced by geometric transformation, attribute disturbance and scene combination, and the enhanced sample data is stored in a distributed database to build a heterogeneous sample library;

[0015] S4, set the scanning parameters based on the spatial coordinate system of the BIM model, perform three-way scanning along the XYZ axis, generate a series of section views, plan views and three-dimensional point cloud models, and standardize the generated views for pretreatment;

[0016] S5, input the 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, and perform dimension compression and feature association through tensor decomposition to output a high-dimensional feature tensor;

[0017] The dual-channel fusion network comprises:

[0018] The BIM parameter feature channel is used to filter key attribute features through a fully connected layer and an attention mechanism;

[0019] The multi-view feature channel is used to extract spatial continuous features using a three-dimensional convolutional neural network;

[0020] The tensor fusion layer is used to map the dual-channel output features into a high-dimensional tensor and perform dimension compression through tensor decomposition;

[0021] The training of the dual-channel fusion network adopts a hybrid loss function: Ls=α×Lf+β×Lg+γ×Lr, wherein Lf is a feature extraction MSE loss, Lg is a graph generation adversarial loss, Lr is an L2 regularization loss, and α, β and γ are weight coefficients;

[0022] S6, convert the high-dimensional feature tensor into vector graphics data in a scalable vector graphics format through a graph mapping module;

[0023] S7, construct a training data set and design a hybrid loss function, train the dual-channel fusion network using a pre-training and fine-tuning strategy, realize lightweight processing through model pruning, and convert it into an open neural network exchange format for deployment;

[0024] The pre-training strategy used in the step S7 comprises unsupervised pre-training on a large-scale building BIM data set, and the fine-tuning stage uses labeled samples for supervised learning.

[0025] Preferably, the geometric transformation comprises scaling, rotating and mirroring the BIM model; the attribute disturbance comprises randomly adjusting the component parameters within the specification allowable range; and the scene combination comprises combining the standard layers or components of different buildings to generate new samples.

[0026] Preferably, the three-way scanning is scanning along the XYZ axis at a set step length to generate a series of YZ plane section views, a series of XZ plane section views and a series of XY plane plan views, and simultaneously save a three-dimensional point cloud model.

[0027] Technical effects and advantages of the present application:

[0028] 1. By constructing a three-level building feature system including micro, meso and macro, and associating it with BIM components through a unique ID, the feature information from component details to overall layout is captured comprehensively and without omission, overcoming the drawbacks of single feature dimension and information fragmentation in traditional methods, and providing complete, structured and semantically rich bottom layer data support for subsequent depth analysis and automatic mapping;

[0029] 2. Data augmentation is performed using geometric transformation, attribute perturbation and scene combination, and a heterogeneous sample library is constructed, greatly expanding the quantity and diversity of the training data set. The overfitting problem of deep learning models caused by insufficient training data or single scene is effectively avoided, and the trained model has stronger adaptability and accuracy for different styles and scales of building projects;

[0030] 3. A dual-channel fusion network is innovatively used to process BIM parameter features and multi-view geometric features simultaneously, and efficient feature association and compression are performed through tensor fusion and decomposition technology. The advantages of BIM model information-geometry integration are fully utilized to combine abstract attribute parameters and intuitive spatial form, so that the final extracted feature tensor can reflect both the physical properties of components and their complex spatial relationships, and the feature expression ability is far superior to that of single data source methods;

[0031] 4. A hybrid loss function is used to guide training, combined with a pre-training-fine-tuning strategy, and model pruning is used to achieve lightweight, and finally converted to an open neural network exchange format for deployment. The hybrid loss function ensures the optimal balance of feature extraction and graph generation, and the lightweight and open neural network exchange format enables the trained model to be efficiently and cross-platform deployed in actual engineering environment, meeting the requirements of calculation efficiency and deployment convenience in practical applications. BRIEF DESCRIPTION OF DRAWINGS

[0032] The invention is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the invention. For ordinary skilled in the art, other drawings can be obtained without creative labor based on the following drawings.

[0033] Figure 1 The overall step flowchart of the invention.

[0034] Figure 2 The structure diagram of the dual-channel fusion network of the invention. DETAILED DESCRIPTION

[0035] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0036] The present application provides a building feature extraction and mapping method based on BIM technology as shown in Figure 1 The building feature extraction and mapping method based on BIM technology specifically comprises the following steps:

[0037] S1, a three-level building feature system containing micro features, meso features and macro features is constructed, and a label set associated with a BIM model component through a unique ID is defined according to an end application scenario;

[0038] It should be further explained that BIM is the abbreviation of Building Information Modeling, which means building information model.

[0039] S2, a BIM model file is exported from a BIM platform, feature data in the three-level building feature system is extracted, and an original sample set is formed;

[0040] S3, the original sample set is subjected to data enhancement through geometric transformation, attribute disturbance and scene combination, and the enhanced sample data is stored in a distributed database to construct a heterogeneous sample library;

[0041] S4, scanning parameters are set based on the spatial coordinate system of the BIM model, three-way scanning is performed along the XYZ axis, a series of profile maps, plane maps and three-dimensional point cloud models are generated, and the generated views are standardized and pretreated;

[0042] S5, BIM parameter features and multi-view features are input into a dual-channel fusion network, dual-channel output features are mapped into a high-dimensional tensor through a tensor fusion layer, and dimension compression and feature association are performed through tensor decomposition, and a high-dimensional feature tensor is output;

[0043] S6, the high-dimensional feature tensor is converted into vector graphic data in a scalable vector graphics format through a mapping module;

[0044] S7, a training data set is constructed and a hybrid loss function is designed, a dual-channel fusion network is trained by using a pre-training and fine-tuning strategy, light-weight processing is realized through model pruning, and an open neural network exchange format is converted for deployment.

[0045] Further, in the above technical solution, the micro features 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.

[0046] Further, the construction of the three-level building feature system in step S1 is based on the international IFC standard, the national "Building Information Model Classification and Coding Standard", and the industry common component classification rules, to ensure the universality and normativity of the feature system.

[0047] Further, the specific implementation of step S1 is as follows:

[0048] S101. Construct a three-level building feature system:

[0049] Micro features: including the node features of components in the BIM model, typically including connection points, end points, and edge features, and specific implementations include: beam-column connection points, wall-floor intersection lines.

[0050] Meso features: including the attribute features of components, typically including materials, sizes, types, and spatial relationship features such as relative positions and adjacent relationships. Specific implementations include: wall thickness, door and window opening direction, and topological relationship between components.

[0051] Macro features: including overall structural system features, typically including structural types, load transfer paths, and spatial layout features such as functional zoning and streamline organization, and specific implementations include: frame structure vs. shear wall structure, floor plan layout, and spatial use efficiency.

[0052] S102. Define a label set and associate it with BIM components:

[0053] According to the end application scenario, a set of label system is defined; each label corresponds to one or more components in the BIM model, and is associated through a globally unique ID.

[0054] Further, the end application scenarios include but are not limited to: construction drawing drawing, quantity calculation, collision detection, energy saving analysis, compliance review, facility management;

[0055] Further, the label set is stored in JSON or XML format, each label contains: label name, attribute key-value pair, application scenario identifier, and is associated through the globally unique ID of the BIM component, and is stored in a relational database.

[0056] Label content example:

[0057] Component type, such as structural column, water supply pipe;

[0058] Functional attribute, such as load-bearing, non-load-bearing;

[0059] Spatial attribution, such as bedroom, hallway;

[0060] Drawing specification tags, such as visible, hidden line;

[0061] Further, in the above technical solution, the specific implementation of step S2 is as follows:

[0062] Sample collection exports model files of the same type of building from BIM platforms such as Revit and ArchiCAD, extracts feature data defined in step S1, and forms an original sample set.

[0063] Further, 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 the component parameters within the specification allowed range; and the scene combination includes combining standard layers or components of different buildings to generate new samples.

[0064] Further, the specific implementation of sample enhancement is as follows:

[0065] Geometric transformation: scaling, rotating, and mirroring the BIM model to simulate different scale building structures;

[0066] Attribute perturbation: randomly adjusting component parameters within the specification allowed range, such as concrete strength grade ±1 grade, wall thickness ±50mm;

[0067] Scene combination: combining standard layers or components of different buildings to generate new samples to simulate complex building forms.

[0068] The sample library uses a distributed database for storage, and each sample contains model ID, BIM file path, feature data, label set, and scan data index.

[0069] Further, the typical setting of randomly adjusting component parameters within the specification allowed range in the attribute perturbation is that the concrete strength grade is floating within one grade based on the design value, and the wall thickness is allowed to deviate by ±10mm to ±50mm;

[0070] Further, in the above technical solution, the three-way scanning in step S4 is scanning along the XYZ axis with a set step size to generate YZ plane series of cross-sections, XZ plane series of cross-sections, and XY plane series of plan views, and to save a three-dimensional point cloud model synchronously.

[0071] Further, the specific implementation of the three-way scanning is as follows:

[0072] S401. Based on the spatial coordinate system of the BIM model, set the scanning parameters:

[0073] Scan step: set according to the construction accuracy requirements, the conventional construction scene is set to 100mm / step;

[0074] View range: automatically identify the model bounding box, determine the scan start and end coordinates;

[0075] Layer filtering: retain structural components, hide auxiliary lines, annotations and other non-feature elements.

[0076] S402. Multi-view generation:

[0077] Scan along the X-axis: generate a series of YZ plane section views to reflect the characteristics in the depth direction of the building;

[0078] Scan along the Y-axis: generate a series of XZ plane section views to reflect the characteristics in the width direction of the building;

[0079] Scan along the Z-axis: generate a series of XY plane views to reflect the layout of each floor;

[0080] Save the three-dimensional point cloud model synchronously, and retain the spatial coordinate information of the components.

[0081] S403. View preprocessing:

[0082] Size normalization: uniformly scaled to 512x512 pixels;

[0083] Feature enhancement: use Canny edge detection algorithm to highlight the component outline;

[0084] Data format conversion: convert the image to tensor form (HxWxC) as the input of the dual-channel fusion network;

[0085] Further, in the tensor form (HxWxC), H represents the height, i.e. the number of pixels in the vertical direction of the tensor, W represents the width, i.e. the number of pixels in the horizontal direction of the tensor, and C represents the number of channels, i.e. the information dimension contained by each pixel point.

[0086] Further, in the above technical solution, the dual-channel fusion network in step S5 comprises:

[0087] BIM parameter feature channel, for filtering key attribute features through a fully connected layer and an attention mechanism;

[0088] Multi-view feature channel, for extracting spatial continuous features using a three-dimensional convolutional neural network;

[0089] Tensor fusion layer, for mapping the output features of the dual-channel to a high-dimensional tensor, and performing dimension compression through tensor decomposition.

[0090] Further, the structure of the dual-channel fusion network is as shown in Figure 2 .

[0091] It needs to be further explained that the multi-view feature channel selects a series of profile and plan view processed by a three-dimensional convolutional neural network, because the three-dimensional convolutional neural network can effectively capture the spatial context information between continuous views and is suitable for extracting the spatial continuous features of building components. For three-dimensional point cloud data, PointNet++ is selected instead of traditional CNN, because PointNet++ can directly process unstructured point cloud data and has permutation invariance and rotation invariance, and is more suitable for extracting building component features with complex spatial relationships.

[0092] Further, the specific embodiments of step S5 are as follows:

[0093] The BIM parameter feature channel takes the BIM parameter features in the heterogeneous sample library as input, and performs preliminary feature extraction through a fully connected layer (FC); an attention mechanism is introduced, which can be Self-Attention, to filter key attributes and output a high-dimensional attribute feature vector.

[0094] The multi-view feature channel takes the images and point cloud data generated by three-way scanning as input, uses a three-dimensional convolutional neural network to extract spatial continuous features, and uses PointNet++ to extract features from the point cloud data, outputting a high-dimensional geometric feature tensor.

[0095] The tensor fusion layer takes the high-dimensional attribute feature vector and the high-dimensional geometric feature tensor as input, fuses the attribute feature vector and the geometric feature tensor, such as outer product, splicing or attention weighted fusion, maps them into a unified high-dimensional tensor, uses tensor decomposition technology such as CP decomposition for dimension compression and feature association, and outputs a compressed high-dimensional feature tensor that combines semantic and geometric information.

[0096] Further, in the above technical solution, the mapping module in step S6 learns the mapping relationship from the feature tensor to the graphic elements through a generative adversarial network, and outputs a scalable vector graphics format vector graphics that meets the "Unified Standard for Architectural Drawing".

[0097] Further, the mapping module uses a generative adversarial network as the core architecture, and the generator is used to receive the high-dimensional feature tensor, gradually upsample and decode it into graphic elements such as lines, labels, and filled areas, and the discriminator judges whether the generated graphics conforms to the distribution of real engineering drawings, and pushes the generator to output graphics that better meet the standards.

[0098] Further, in the above technical solution, the training of the dual-channel fusion network in step S7 uses a hybrid loss function: Ls=α×Lf+β×Lg+γ×Lr, where Lf is the feature extraction MSE loss, Lg is the graphic generative adversarial loss, Lr is the L2 regularization loss, and α, β, γ are weight coefficients.

[0099] It needs to be further explained that MSE is the abbreviation of Mean Square Error, which means mean square error.

[0100] Further, the weight coefficients a, b, g in the hybrid loss function are optimized on the validation set by Bayesian optimization. The typical value range is: a e [0.4, 0.6], b e [0.3, 0.5], g e [0.1, 0.2], and a + b + g = 1. This setting aims to balance the feature reconstruction accuracy, graph generation quality and model generalization ability.

[0101] Further, the pre-training strategy used in step S7 includes unsupervised pre-training on a large-scale building BIM dataset, and the fine-tuning stage uses labeled samples for supervised learning.

[0102] Further, the specific implementation of step S7 is as follows:

[0103] S701. Use the heterogeneous sample library constructed after data augmentation in step S3 as the training data source, each sample contains: model ID, three-level feature data extracted from BIM file path, label set, and scanning data index;

[0104] S702. Use the hybrid loss function: Ls = a x Lf + b x Lg + g x Lr to guide the training;

[0105] S703. Use a large-scale unlabeled BIM dataset for unsupervised pre-training to learn general building feature representation, use labeled samples for supervised learning, fine-tune for specific mapping tasks, fix part of the bottom layer parameters during fine-tuning, only update the high layer network weights, to improve training efficiency and generalization ability;

[0106] S704. Use model pruning technology to prune connections / channels with small weights or low contribution in the network, retain key feature extraction capability, reduce model parameter quantity and calculation amount;

[0107] S705. Convert the trained dual-channel fusion network into an open neural network exchange format to ensure efficient inference of the model on different platforms.

[0108] Further, the large-scale building BIM dataset is constructed as follows: collect public BIM model libraries such as BIMobject, National BIM Library model files;

[0109] Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art will appreciate that the technical solutions described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalent features, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for building feature extraction and mapping based on BIM technology, characterized in that, Specifically comprising the following steps: S1, constructing a three-level building feature system comprising micro features, meso features and macro features, and defining a tag set associated with the BIM model components through a unique ID according to the end application scenario; 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; S2, exporting the BIM model file from the BIM platform, extracting feature data in the three-level building feature system, and forming an original sample set; S3, data augmentation of the original sample set through geometric transformation, attribute disturbance and scene combination, and storing the augmented sample data in a distributed database to construct a heterogeneous sample library; S4, setting scanning parameters based on the spatial coordinate system of the BIM model, performing three-way scanning along the XYZ axis, generating a series of profile maps, plane maps and three-dimensional point cloud models, and performing standardized preprocessing on the generated views; S5, inputting BIM parameter features and multi-view features into a dual-channel fusion network, mapping dual-channel output features into a high-dimensional tensor through a tensor fusion layer, and performing dimension compression and feature association through tensor decomposition to output a high-dimensional feature tensor; The dual-channel fusion network comprises: A BIM parameter feature channel for filtering key attribute features through a fully connected layer and an attention mechanism; A multi-view feature channel for extracting spatial continuous features using a three-dimensional convolutional neural network; A tensor fusion layer for mapping dual-channel output features into a high-dimensional tensor and performing dimension compression through tensor decomposition; The training of the dual-channel fusion network uses 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, γ are weight coefficients; S6, converting the high-dimensional feature tensor into scalable vector graphics data in scalable vector graphics format through a mapping module; S7, constructing a training data set and designing a hybrid loss function, training the dual-channel fusion network using a pre-training and fine-tuning strategy, implementing lightweight processing through model pruning, and converting it into an open neural network exchange format for deployment; The pre-training strategy used in step S7 includes unsupervised pre-training on a large-scale building BIM data set, and supervised learning using labeled samples in the fine-tuning stage.

2. The building feature extraction and mapping method based on BIM technology according to claim 1, characterized in that: The geometric transformation includes scaling, rotating and mirroring the BIM model; the attribute disturbance includes randomly adjusting the component parameters within the specified range; and the scene combination includes combining standard layers or components of different buildings to generate new samples.

3. The building feature extraction and mapping method based on BIM technology according to claim 1, characterized in that: The three-way scanning is scanning along the XYZ axis with a set step size to generate a series of YZ plane profile maps, XZ plane profile maps and XY plane profile maps, and to save a three-dimensional point cloud model simultaneously.

Images