Curtain wall profile standard establishment method and system based on deep learning

Abstract

According to the curtain wall profile standard establishment method and system based on deep learning, the deep learning model is established, automatic and intelligent analysis can be conducted on massive curtain wall profile section graphs through the model, and the method and system can be used for establishing the curtain wall profile standard through analysis of CAD primitives of local features. And obtaining the concentration ratio of the size of each primitive, and automatically and preferably selecting the primitive with high size concentration ratio to form a standard form of local features. According to the method, automatic and intelligent analysis of a large number of graphs which cannot be manually completed can be completed, meanwhile, the method is objective and public, acknowledged standardization is easily formed, and the standardization degree of curtain wall profiles is conveniently improved.

Description

Technical Field

[0001] This invention relates to the field of building engineering technology, and in particular to a method and system for establishing standards for curtain wall profiles based on deep learning. Background Technology

[0002] With the development of technology, building curtain walls have been widely used in various buildings. However, the standardization of curtain wall profiles is very low. Currently, in most curtain wall projects, the vast majority of profiles require new molds, resulting in a waste of manpower and resources. To improve the standardization of curtain wall profiles, curtain wall manufacturers have done a lot of standardization work. Based on their own technological development history and the personalized needs of architectural designs, each manufacturer has formed its own curtain wall solution system. However, due to differences in the technological development history, technical concepts, technical personnel capabilities, and the personalized architectural needs faced by various manufacturers, curtain wall profile design schemes have become diverse and varied. Since the methods for generating curtain wall profile standards themselves vary greatly, after many years of development, the standardization of curtain wall profiles remains very low, and with the increasing use of curtain walls, the standardization of curtain wall profiles shows a downward trend.

[0003] From a curtain wall design perspective, the overall dimensions of curtain wall profiles depend on the mechanical stress requirements of the curtain wall. The width and height dimensions are determined through mechanical calculations and approximate modules. Meanwhile, the detailed structures of each curtain wall profile depend on its specific functional requirements. A cross-section of a curtain wall profile reveals details such as: grooves for rubber strips, bolt grooves, and assembly fitting grooves. Traditionally, curtain wall standardization involves establishing a series of specifications for the width and height dimensions of the profile cross-section according to a certain module. This ensures that the main dimensions conform to the module requirements, resulting in a high degree of standardization. However, there are no universally accepted standards for the detailed structures, leading to a wide variety of variations and extremely low standardization. This results in a very low overall standardization of curtain wall profiles. Summary of the Invention

[0004] To address the shortcomings of the aforementioned standardization schemes for curtain wall profiles, the present invention aims to provide a method and system for establishing curtain wall profile standards based on deep learning, thereby solving the above problems.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] This invention provides a method for establishing standards for curtain wall profiles based on deep learning, characterized by comprising:

[0007] Step S1: Collect curtain wall profile detail drawings from curtain wall design projects and establish a standard detail database;

[0008] Step S2: Preprocess the curtain wall profile detail drawings to obtain the preprocessed image;

[0009] Step S3: Construct the mapping relationship between the pixels of the preprocessed image and the primitives of the curtain wall profile node drawing;

[0010] Step S4: Label the local features of the preprocessed image as a dataset for model training;

[0011] Step S5: Construct a deep learning model. Use the dataset from step S4 to train the model. The input to the model is an image, and the output of the model is local features.

[0012] Step S6: Perform the same processing on the curtain wall profile node drawings of the new project as in step S2, and input the processed image into the trained deep learning model to identify local features;

[0013] Step S7: Based on the local features identified in step S6, perform feature analysis to extract the primitives corresponding to the local features and the primitive sizes of the same parts of the local features.

[0014] Step S8: Establish a list of element size data for the same parts with the same local features, and sort them according to size concentration.

[0015] Step S9: Select one or more local features with the highest size concentration as the standard feature of that local feature.

[0016] Furthermore, the deep learning-based method for establishing curtain wall profile standards provided by this invention may also have the following feature: the preprocessing in step S2 includes the following steps:

[0017] 1) Standardization process: Convert the curtain wall profile detail drawings into a unified image format and adjust the images to a uniform resolution and size;

[0018] 2) Image enhancement: Performing contrast adjustment, noise reduction, and edge enhancement on the image;

[0019] 3) Drawing selection: Select the area of ​​the curtain wall profile details and exclude irrelevant parts;

[0020] 4) Layer Filtering: Select the specific layer containing profile elements and filter out irrelevant layers;

[0021] 5) Profile selection: Select the specific range of profiles.

[0022] Furthermore, the method for establishing curtain wall profile standards based on deep learning provided by this invention may also have the following feature: in step S4, LabelMe is used to manually annotate local features.

[0023] Furthermore, the method for establishing curtain wall profile standards based on deep learning provided by this invention may also have the following feature: in step S5, the dataset is divided into a 70% training set, a 15% validation set, and a 15% test set.

[0024] Furthermore, the method for establishing curtain wall profile standards based on deep learning provided by this invention may also have the following feature: in step S5, the deep learning model is an end convolutional neural network model VGG16-UNet, and the model training adopts a supervised learning method.

[0025] Furthermore, the deep learning-based method for establishing standards for curtain wall profiles provided by this invention may also have the following feature: the model training process includes:

[0026] 1) Data augmentation of samples through rotation, scaling, translation, and flipping;

[0027] 2) Use grid search or random search methods to optimize the learning rate, batch size, and optimizer type;

[0028] 3) Use a multi-task loss function to evaluate classification loss and regression loss;

[0029] 4) Use Dropout and L2 regularization to prevent model overfitting.

[0030] Furthermore, the method for establishing curtain wall profile standards based on deep learning provided by this invention may also have the following features: the local features output by the model in step S6 include: overall shape, groove shape, profile connection method, support structure and geometric details.

[0031] Furthermore, the deep learning-based method for establishing curtain wall profile standards provided by this invention may also have the following feature: wherein the feature analysis in step S7 includes the following steps:

[0032] 1) Feature encoding: Encoding the local features extracted by the model into a high-dimensional vector representation;

[0033] 2) Similarity calculation: The similarity of the encoded local features is calculated to evaluate the similarity between the matched local features and the standard local features;

[0034] 3) Feature matching: Outputs the recognition result of local features, which is a certain part of the image converted from the cross-section of the corresponding profile;

[0035] 4) Mapping primitives: By mapping the pixel position of the image to the primitive coordinates, the associated primitives of local features are identified and the primitive size is extracted.

[0036] This invention also provides a deep learning-based curtain wall profile standard establishment system for executing the aforementioned deep learning-based curtain wall profile standard establishment method. The system comprises: a standard node database module for storing and managing curtain wall profile node drawing file data; a drawing preprocessing module for preprocessing curtain wall profile node drawings, including a format processing unit, a resolution and size processing unit, an image enhancement unit, a drawing frame selection unit, a layer filtering unit, and a profile frame selection unit; a mapping relationship module storing the mapping relationship between image pixels and drawing primitives; and a local feature extraction module that calls the trained deep learning algorithm. The learning model identifies local features of an image; the feature comparison module is used to encode, calculate similarity, match, map to primitives, and extract primitive dimensions of the local features identified by the model, including feature encoding unit, similarity calculation unit, feature matching unit, primitive extraction unit, and primitive size extraction unit; the primitive size list generation module is used to extract the primitive dimensions corresponding to the local features based on the feature comparison module and establish a primitive data list for each part with the same local features; the standard feature recognition result module sorts the data in the primitive data list according to the size concentration and outputs standard features based on the sorting result.

[0037] Furthermore, the deep learning-based curtain wall profile standard establishment system provided by this invention may also have the following features: the drawing preprocessing module includes: a format processing unit, a resolution and size processing unit, an image enhancement unit, a drawing frame selection unit, a layer filtering unit, and a profile frame selection unit; the feature comparison module includes: a feature encoding unit, a similarity calculation unit, a feature matching unit, and a mapping primitive unit.

[0038] The function and effects of this invention:

[0039] The deep learning-based method for establishing standards for curtain wall profiles in this invention constructs a deep learning model. This model enables automated and intelligent analysis of massive amounts of curtain wall profile cross-sectional graphics. By analyzing the CAD primitives of local features, the method determines the concentration of dimensions for each primitive and automatically selects primitives with high dimensional concentration to form standard forms for local features. Because this method incorporates deep learning, it can automate and intelligently analyze a large number of graphics that are impossible to perform manually. Furthermore, the method is objective and publicly available, facilitating the formation of universally accepted standards and improving the standardization of curtain wall profiles. Attached Figure Description

[0040] Figure 1 These are schematic diagrams of curtain wall profile nodes. (a), (b), and (c) are cross-sectional views of different curtain wall profile nodes. The areas labeled 1-1 to 1-5 represent different local features.

[0041] Figure 2This is a flowchart of a method for establishing curtain wall profile standards based on deep learning in an embodiment of the present invention.

[0042] Figure 3 This is a block diagram of a deep learning-based curtain wall profile standard establishment system in an embodiment of the present invention. Detailed Implementation

[0043] To make the technical means, creative features, objectives and effects of this invention easier to understand, the following embodiments, in conjunction with the accompanying drawings, will specifically illustrate the technical solution of this invention.

[0044] See Figure 2 This invention provides a method for establishing standards for curtain wall profiles based on deep learning, comprising the following steps:

[0045] Step S1: Drawing Collection and Database Construction

[0046] We collected a large number of curtain wall profile detail drawings from previous curtain wall design projects. The drawings are in CAD format.

[0047] The collected drawings are compared and analyzed to define the local features of the curtain wall profile nodes. A standard node database is established to ensure that the drawings in the database cover curtain wall nodes of different types and complexities, so as to enhance the generalization ability of the model.

[0048] Step S2: Drawing Preprocessing

[0049] The curtain wall profile detail drawings are preprocessed to obtain the preprocessed image. The specific steps are as follows:

[0050] 1) Standardization process: Convert the curtain wall profile detail drawings into a unified image format, such as PNG or JPEG, and adjust the images to a uniform resolution and size to ensure the consistency of data input.

[0051] 2) Image enhancement: Perform contrast adjustment, noise reduction, and edge enhancement on the image to improve the clarity and detail of the drawing and enhance the deep learning model's ability to recognize features.

[0052] 3) Drawing selection: Users manually select the area of ​​the curtain wall profile node, excluding irrelevant parts, to ensure that the model focuses on the target node area.

[0053] 4) Layer filtering: Users can manually select specific layers containing profile elements and filter out irrelevant layers to further reduce interference information and improve the accuracy of feature extraction.

[0054] 5) Profile selection: Users manually select the specific range of the profile to ensure that the model receives the most relevant profile feature information.

[0055] Step S3: Construct mapping relationship

[0056] Establish a mapping relationship between the pixels of the preprocessed image and the CAD elements of the curtain wall profile node drawings.

[0057] Step S4: Construct the dataset for model training

[0058] Users manually annotate local features on preprocessed images using LabelMe, which serves as the dataset for model training. The annotated images should contain over 3000 different types of curtain wall node drawing samples to ensure the model can learn rich feature information during training. Each sample contains multiple views and details; these local features cover the overall shape of different profiles, groove shapes, profile connection methods, support structures, and geometric details. The support structure refers to the support structures on the profiles and their connection methods; geometric details include complex geometric shapes such as curves, angles, and holes.

[0059] Step S5: Model Training

[0060] A deep learning model is constructed and trained using the dataset from step S4. The model input is an image, and the model output is local features. The model identifies local features including the overall shape of the profile, the groove shape, the profile connection method, the support structure, and geometric details.

[0061] The VGG16-UNet convolutional neural network model was adopted, and the model was trained using a supervised learning method.

[0062] The training dataset is divided into a 70% training set, a 15% validation set, and a 15% test set to ensure that the model performs stably on different datasets.

[0063] The model training process includes the following steps:

[0064] 1) Data augmentation: Data augmentation is performed on samples by rotating, scaling, translating, and flipping to generate diverse training samples.

[0065] 2) Hyperparameter tuning: Use grid search or random search methods to optimize the learning rate, batch size and optimizer type to improve model training performance.

[0066] 3) Loss function design: A multi-task loss function is adopted to comprehensively consider classification loss and regression loss, ensuring that the model is optimized in both feature classification and localization.

[0067] 4) Regularization: Use Dropout and L2 regularization to prevent model overfitting and improve generalization ability.

[0068] During model training, the model structure and hyperparameters are adjusted based on the evaluation results of the validation set, such as increasing or decreasing the number of network layers and adjusting the learning rate, to improve model performance. Early stopping is applied to prevent overfitting during training. Cross-validation is used to train the model multiple times and evaluate its stability, ensuring consistent performance across different datasets.

[0069] Step S6: Perform local feature recognition using the trained model.

[0070] Perform the same processing as step S2 on the curtain wall profile detail drawings for the new project.

[0071] The processed image is input into a trained deep learning model, which automatically identifies local features including the overall shape of the profile, the groove shape, the profile connection method, the support structure, and geometric details. See also Figure 1 The figure shows the cross-sectional diagrams of the curtain wall profile nodes (a), (b), and (c), which identify local features. The local features are labeled 1-1 to 1-5.

[0072] Step S7: Feature Analysis

[0073] Based on the local features identified in step S6, feature analysis is performed to extract the corresponding CAD primitives and the dimensions of the CAD primitives at the same location as the local feature. The feature analysis specifically includes the following steps:

[0074] 1) Feature encoding: Encoding the local features extracted by the model into a high-dimensional vector representation;

[0075] 2) Similarity calculation: The similarity of the encoded local features is calculated to evaluate the similarity between the matched local features and the standard local features;

[0076] 3) Feature matching: Outputs the recognition result of local features, which is a certain part of the image converted from the cross-section of the corresponding profile;

[0077] 4) Mapping primitives: By mapping the pixel position of the image to the coordinates of the primitives, the associated CAD primitives of local features are identified, and the dimensions of the CAD primitives are extracted.

[0078] Step S8: Create a list of element dimensions

[0079] Establish a list of element size data for the same parts with the same local features, and sort them according to size concentration;

[0080] Step S9: Generate standard features

[0081] The local features with the highest dimensional concentration are selected as the standard features for that local feature. The same method is used to identify, extract, and sort each feature, generating the standard features for each key feature, thus forming a local key feature atlas for the curtain wall design.

[0082] See Figure 3 The present invention also provides a deep learning-based curtain wall profile standard establishment system for executing the above-mentioned deep learning-based curtain wall profile standard establishment method, which includes computer function modules such as a standard node database module, a drawing preprocessing module, a mapping relationship module, a local feature extraction module, a feature analysis module, a graphic element size list generation module, and a standard feature result output module.

[0083] The standard node database module is used to store and manage the drawing file data of curtain wall profile nodes.

[0084] The drawing preprocessing module is used to preprocess the data of curtain wall profile detail drawings. The drawing preprocessing module includes a format processing unit, a resolution and size processing unit, an image enhancement unit, a drawing frame selection unit, a layer filtering unit, and a profile frame selection unit.

[0085] The mapping module stores the mapping relationship between image pixels and drawing primitives.

[0086] The local feature extraction module calls the trained deep learning model to identify local features of the image.

[0087] The feature analysis module is used to analyze the local features identified by the model, identify the relevant primitives of the local features, and extract the primitive dimensions. The feature analysis module includes a feature encoding unit, a similarity calculation unit, a feature matching unit, a primitive extraction unit, and a primitive dimension extraction unit.

[0088] The primitive size list generation module is used to extract primitive sizes corresponding to local features based on the feature comparison module, establish a primitive data list for each part with the same local features, and sort the data in the primitive data list according to the size concentration.

[0089] The standard feature output module outputs the local features with the highest size concentration (the specific number is defined by the user) based on the sorting results.

[0090] The advantages of the deep learning-based method and system for establishing curtain wall profile standards in this invention are as follows: Users can obtain standard features of local characteristics of curtain wall profile nodes through this method and system, and can apply these standard features to design the cross-section of the curtain wall profile. When designing curtain wall profiles, the overall size is determined by the mechanical stress requirements of the curtain wall and similar modules. Key features are selected according to the functional requirements of the curtain wall profiles. By creating standard feature sets for local features such as sealing strip grooves, bolt grooves, and assembly mating grooves using the above method, the standardization level of the curtain wall profiles will be improved.

[0091] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for establishing curtain wall profile standards based on deep learning, characterized in that: include: Step S1, collecting curtain wall profile node drawings in the curtain wall design project and establishing a standard node database; Step S2, preprocessing the curtain wall profile node drawing to obtain a preprocessed image; Step S3, constructing a mapping relationship between pixels of the preprocessed image and graphic elements of the curtain wall profile node drawing; Step S4, annotating local features of the preprocessed image as a data set for model training; Step S5, building a deep learning model, using the data set in step S4 to train the model, the input of the model is the image, and the output of the model is the local feature; Step S6, performing the same processing as step S2 on the curtain wall profile node drawings of the new project, and inputting the processed images into the trained deep learning model to identify local features; Step S7, based on the local features identified in step S6, feature analysis is performed to extract the graphic element corresponding to the local features and the size of the graphic element at the same part of the local features; Step S8, establishing a list of primitive size data of the same parts of the same local features, and sorting them according to the size concentration; Step S9, taking one or more local features with the highest size concentration as the standard features of the local features.

2. The method for establishing curtain wall profile standards based on deep learning as claimed in claim 1, characterized in that: in, The preprocessing of step S2 includes the following steps: 1) Standardization processing: Convert the curtain wall profile node drawings into a unified image format and adjust the images to a unified resolution and size; 2) Image enhancement: image contrast adjustment, denoising, and edge enhancement; 3) Drawing selection: select the area of ​​the curtain wall profile node and exclude irrelevant parts; 4) Layer filtering: select specific layers containing profile elements and filter irrelevant layers; 5) Profile selection: select the specific range of profiles.

3. The method for establishing curtain wall profile standards based on deep learning as claimed in claim 1, characterized in that: in, In step S4, LabelMe is used to manually annotate local features.

4. The method for establishing curtain wall profile standards based on deep learning according to claim 1, characterized in that: in, In step S5, the data set is divided into a 70% training set, a 15% validation set, and a 15% test set.

5. The method for establishing curtain wall profile standards based on deep learning according to claim 1, characterized in that: in, The deep learning model described in step S5 is an end-to-end convolutional neural network model VGG16-UNet, and the model training adopts a supervised learning method.

6. The method for establishing curtain wall profile standards based on deep learning as claimed in claim 5, Features: The model training process includes: 1) Enhance the data of samples by rotation, scaling, translation, and flipping; 2) Use grid search or random search methods to optimize learning rate, batch size and optimizer type; 3) Use multi-task loss function to evaluate classification loss and regression loss; 4) Use Dropout and L2 regularization methods to prevent model overfitting.

7. The method for establishing curtain wall profile standards based on deep learning as claimed in claim 1, characterized in that: in, The local features output by the model in step S6 include: overall shape, notch shape, profile connection method, support structure and geometric details.

8. The method for establishing curtain wall profile standards based on deep learning as claimed in claim 1, Features: The feature analysis in step S7 includes the following steps: 1) Feature encoding: Encode the local features extracted by the model into a high-dimensional vector representation; 2) Similarity calculation: perform similarity calculation on the encoded local features to evaluate the similarity between the matched local features and the standard local features; 3) Feature matching: output the recognition result of local features, which is a certain part of the image converted from the corresponding profile cross section; 4) Mapping primitives: Through the mapping relationship between the pixel position of the image and the primitive coordinates, the associated primitives of the local features are identified and the primitive sizes are extracted.

9. A curtain wall profile standard establishment system based on deep learning, used to execute the curtain wall profile standard establishment method based on deep learning as claimed in any one of claims 1 to 8, characterized in that: include: Standard node database module, used to store and manage curtain wall profile node drawing file data; Drawing preprocessing module, used for data preprocessing of curtain wall profile node drawings; A mapping relationship module stores the mapping relationship between image pixels and drawing primitives; The local feature extraction module calls the trained deep learning model to identify the local features of the image; A feature analysis module is used to analyze the local features identified by the model, identify the relevant primitives of the local features and extract the size of the primitives; A primitive size list generation module is used to extract primitive sizes corresponding to local features based on the feature comparison module, establish a primitive size data list of the same parts of the same local features, and sort the data in the primitive size data list according to the size concentration; The standard feature result output module is used to filter out standard features based on the sorting results.

10. The curtain wall profile standard establishment system based on deep learning as claimed in claim 9, Features: The drawing preprocessing module includes: a format processing unit, a resolution and size processing unit, an image enhancement unit, a drawing frame selection unit, a layer screening unit, and a profile frame selection unit; The feature comparison module includes: a feature encoding unit, a similarity calculation unit, a feature matching unit, and a mapping primitive unit.

Images