End-to-end intelligent building shear wall structure arrangement generation method and system

Abstract

The invention discloses an end-to-end intelligent building shear wall structure arrangement generation method. The method comprises the steps that a building plane arrangement sketch designed by an architect initial scheme is obtained through a sketch drawing tool; inputting the input building plane layout sketch into a pre-trained shear wall structure layout generation model to obtain a corresponding building shear wall structure layout feature vector; and inputting the arrangement feature vector of the building shear wall structure into a post-processing step based on computer vision so as to enable a generated structure wall body to be more smooth and continuous, and finally outputting a building shear wall structure arrangement diagram. The invention further comprises an end-to-end AI building shear wall intelligent structure design system. An end-to-end workflow which is easy to use is formed by the Autostruct developed by the method disclosed by the invention. In the system, an architect can quickly draw a building sketch and input the building sketch into a pre-training model, and the model generates reasonable structure scheme arrangement in an extremely short time, so that a timely structure arrangement reference is provided for the architect, and the design experience of a structure engineer is transmitted to the architect to become a structure design brain of the architect.

Description

Technical Field

[0001] The present invention belongs to the field of image generation, and in particular relates to an end-to-end intelligent building shear wall structure layout generation method and system. Background Art

[0002] The architectural design field is constantly striving to achieve intelligent design to reduce tedious, repetitive manual work. A well-designed building requires a close integration of architectural and structural design, requiring close collaboration between architects and structural engineers. In the architectural design process, the architect first designs a building layout (i.e., a design consisting of only basic wall layouts, without load-bearing structures). Then, the structural engineer designs the load-bearing structure and returns the results, allowing the architect to modify any inconsistencies in the original layout. This process is often repeated many times before a final layout is finalized. A significant gap exists between architects and structural engineers during this process. Architects lack experience in structural design and layout, and often fail to accurately assess whether a design meets structural requirements (e.g., designing a window where a load-bearing wall should be). Conversely, structural engineers are unable to provide architects with constant structural support, nor can they directly convey their abstract structural experience. Consequently, this process requires architects to repeatedly submit their complete design to the structural engineer, requiring constant communication and revisions. This results in a significant waste of time and energy, and architects are reluctant to try new and innovative layouts to avoid the hassle of back-and-forth communication. Therefore, this inhibits the innovation of architectural design. With the development of artificial intelligence, the use of deep learning to realize the automated design of building structures can bring new ideas to solve this problem.

[0003] According to McKinsey research, the construction industry generates approximately $265 billion in profits annually. Of this, total design costs account for approximately 2.5%, and preliminary design accounts for 5%. Statistics show that engineers typically spend over 3.5 hours designing a single structural drawing, while AI-assisted design can generate one in less than 30 seconds. AI-assisted design is expected to improve efficiency by more than 10 times, reducing time consumption by approximately one-third, and generating potential profits of $110 million (= $265 billion x 2.5% x 5% x 33.3%). Furthermore, AI-assisted design allows architects to instantly receive a building structural layout simply by inputting a sketch, providing a reference and reducing the time-consuming back-and-forth with structural engineers. This significantly improves efficiency and reduces communication costs, allowing architects to free up time to develop more innovative and livable building designs, resulting in superior architectural designs. With the advancement of urbanization, the demand for high-rise residential buildings continues to increase. Shear wall structures are one of the most common structural types for mid- and high-rise buildings, but their design requires engineers with extensive structural design experience. Therefore, using AI for automatic design of building shear wall structures will enable the acquisition of shear wall structure layout with a low threshold, significantly reduce design and communication costs, and improve design efficiency. It is one of the most worthy scenarios for applying structural automation design. Summary of the Invention

[0004] The present invention aims to overcome the above-mentioned shortcomings of the prior art and provides an end-to-end intelligent building shear wall structure layout generation method and system.

[0005] The purpose of this invention is to design an end-to-end AI-powered intelligent structural design system for building shear walls: AutoStruct. This system leverages neural networks to efficiently learn a large amount of structural drawing information. In this system, architects can quickly sketch a building and input it into the model. The model then generates a structural layout in a fraction of the time, providing architects with timely structural layout references. This system transfers the design experience of structural engineers to architects, effectively becoming their structural design brain. This significantly improves efficiency, reduces communication costs, and helps architects save time and create better designs.

[0006] In a first aspect, an embodiment of the present invention provides an end-to-end method for generating a shear wall structure layout for an intelligent building, the method comprising:

[0007] S1: Obtaining a building layout sketch of the architect's preliminary design through a sketching tool;

[0008] S2: Inputting the input building plan layout sketch into the pre-trained shear wall structure layout generation model to obtain the corresponding building shear wall structure layout feature vector;

[0009] S3: Inputting the building shear wall structure layout feature vector into a post-processing step based on computer vision to make the generated structural wall more flat and continuous, and finally outputting the building shear wall structure layout diagram.

[0010] Furthermore, the sketch drawing tool operation steps described in step S1 include:

[0011] S11: Select one of two commonly used architectural drawing methods: coordinate drawing method and directional drawing method to draw;

[0012] S12: Select the corresponding line segment thickness based on the three provided line segment thicknesses representing different wall thicknesses; and select the corresponding line segment thickness based on the four provided colors representing shear walls, ordinary walls, doors, and windows.

[0013] S13: In the coordinate drawing method, a line segment is drawn by inputting the coordinates of the starting point; in the directional drawing method, a line segment of a specific length is drawn according to the mouse direction and the input length;

[0014] S14: Use other features provided by the tool to assist in drawing: real-time display of the mouse coordinates to determine the drawing coordinates; automatic absorption of the endpoint coordinates within 10px of the line segment endpoint for quick and accurate positioning; users can import pictures and adjust the transparency as backgrounds to facilitate copying and drawing; use the undo function to undo incorrect drawings;

[0015] S15: Complete the sketch drawing and output the building layout sketch.

[0016] Furthermore, the shear wall structure arrangement generation model of step S2 is trained through the following steps:

[0017] S21: Obtain a training sample set; wherein the training sample set includes sample pairs of shear wall building layout drawings and shear wall structure layout drawings.

[0018] S22: Inputting the sample pairs into a TransUnet-based generator to fully learn the global and local layout features of the drawing and generate a corresponding shear wall structure layout drawing;

[0019] S23: Input the corresponding shear wall structure layout diagram and the shear wall structure layout diagram in the training set into a wavelet transform-based discriminator to determine whether the input image is an image from the generator; the wavelet transform-based discriminator will focus on the high-frequency information of the wall in the image;

[0020] S24: Perform adversarial generative training based on the TransUnet generator and the wavelet transform-based discriminator, calculate the VGG loss and the LSGAN loss, and update the parameters of the generator through error backpropagation to optimize the model to generate a more realistic and reliable structural layout result.

[0021] In a second aspect, an embodiment of the present invention further provides an end-to-end AI building shear wall intelligent structural design system that implements the method of the present invention, including AutoStruct, and:

[0022] The TransUnet-based generator learns both global and local layout features of the drawings;

[0023] A wavelet-based discriminator learns and identifies high-frequency information features of walls in drawings, forming a Transformer-Wavelet architecture. This gives AutoStruct a significant advantage in learning the feature information of shear wall data.

[0024] A post-processing process based on computer vision is designed to address the problems of discontinuous and irregular distribution in the generated walls. By alternating the specially designed closing and clustering functions, defects of different sizes in the generated walls are repaired, thereby improving the continuity and flatness of the generated walls.

[0025] Sketching tools are designed in HTML5. Architects can quickly draw architectural sketches on the web and input them into the model, which then generates the structural layout generated by AutoStruct. AutoStruct thus forms a complete end-to-end workflow: architects simply draw sketches through sketching and input them into the system, which will generate the shear wall structure layout through a pre-trained model. The results will be automatically input into the post-processing process for post-processing optimization, and finally the shear wall structure layout will be output to the architect.

[0026] The main innovative features and characteristics of the technical solution adopted by the present invention include:

[0027] (1) Efficient Transformer-Wavelet architecture design: Using a TransUnet-based generator and matching it with a wavelet-based discriminator. This architecture takes into account the extraction of global features and local details, while significantly improving the model's ability to learn features of high-frequency key information such as walls.

[0028] (2) End-to-end Usage Process: This paper creates a simple sketching tool that is customizable to architects and also designs a targeted post-processing process. Combined with the Transformer-Wavelet architecture, the system forms a complete end-to-end usage process. Architects can quickly and easily create structural layouts using AutoStruct, bridging the gap in expertise.

[0029] (3) Excellent performance: AutoStruct significantly outperforms similar open-source shear wall automatic design models in four different shear wall structure design datasets, with higher generated layout quality and faster training convergence speed.

[0030] The present invention has developed an AI building shear wall structure intelligent design system: AutoStruct. The core of the system is an efficient Transformer-Wavelet architecture, which includes a generator based on TransUnet and a discriminator based on wavelet transform. This architecture takes into account the extraction of global features and local details of architectural design drawings, while improving the learning ability of high-frequency information features such as walls. Then, in view of the problems such as discontinuous and irregular distribution of generated walls, the present invention has designed a set of post-processing processes based on computer vision, which can repair defects of different sizes of walls and thus improve the continuity and flatness of generated walls. In addition, the system of the present invention also includes a sketching tool that conforms to the usage habits of architects. This tool enables architects to quickly draw architectural sketches on the web and input them into the model to generate structural layouts, realizing easy-to-use and end-to-end automated design. In summary, the AutoStruct developed by the present invention forms an easy-to-use end-to-end workflow. In this system, architects can quickly draw architectural sketches and input them into the model. The model will generate a reasonable structural layout in a very short time, providing architects with timely structural layout references, and transferring the design experience of structural engineers to architects, becoming the architect's structural design brain.

[0031] The advantages of this invention include an end-to-end, easy-to-use workflow that lowers the barrier to entry for architects. The efficient Transformer-Wavelet architecture extracts both global features and local details from design drawings, generating shear wall layouts close to those designed by engineers. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The overall structure and usage process of the present invention;

[0033] Figure 2 This is a diagram of the generator architecture of the present invention;

[0034] Figure 3This is a diagram of the wavelet discriminator architecture of the present invention;

[0035] Figure 4 This is a flowchart of the computer vision post-processing steps of the present invention;

[0036] Figure 5 Schematic diagram of the operating interface of the sketch drawing tool of the present invention. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0038] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0039] Example 1

[0040] Reference Figure 1 This embodiment provides an end-to-end intelligent building shear wall structure layout generation method, the generation method comprising:

[0041] S1: Obtaining a building layout sketch of the architect's preliminary design through a sketching tool;

[0042] S2: Inputting the input building plan layout sketch into the pre-trained shear wall structure layout generation model to obtain the corresponding building shear wall structure layout feature vector;

[0043] S3: Inputting the building shear wall structure layout feature vector into a post-processing step based on computer vision to make the generated structural wall more flat and continuous, and finally outputting the building shear wall structure layout diagram.

[0044] Furthermore, the sketch drawing tool operation steps described in step S1 include:

[0045] S11: Select one of two commonly used architectural drawing methods: coordinate drawing method and directional drawing method to draw;

[0046] S12: Select the corresponding line segment thickness based on the three provided line segment thicknesses representing different wall thicknesses; and select the corresponding line segment thickness based on the four provided colors representing shear walls, ordinary walls, doors, and windows.

[0047] S13: In the coordinate drawing method, a line segment is drawn by inputting the coordinates of the starting point; in the directional drawing method, a line segment of a specific length is drawn according to the mouse direction and the input length;

[0048] S14: Use other features provided by the tool to assist in drawing: real-time display of the mouse coordinates to determine the drawing coordinates; automatic absorption of the endpoint coordinates within 10px of the line segment endpoint for quick and accurate positioning; users can import pictures and adjust the transparency as backgrounds to facilitate copying and drawing; use the undo function to undo incorrect drawings;

[0049] S15: Complete the sketch drawing and output the building layout sketch.

[0050] Furthermore, the shear wall structure arrangement generation model of step S2 is trained through the following steps:

[0051] S21: Obtain a training sample set; wherein the training sample set includes sample pairs of shear wall building layout drawings and shear wall structure layout drawings.

[0052] S22: Inputting the sample pairs into a TransUnet-based generator to fully learn the global and local layout features of the drawing and generate a corresponding shear wall structure layout drawing;

[0053] S23: Input the corresponding shear wall structure layout diagram and the shear wall structure layout diagram in the training set into a wavelet transform-based discriminator to determine whether the input image is an image from the generator; the wavelet transform-based discriminator will focus on the high-frequency information of the wall in the image;

[0054] S24: Perform adversarial generative training based on the TransUnet generator and the wavelet transform-based discriminator, calculate the VGG loss and the LSGAN loss, and update the parameters of the generator through error backpropagation to optimize the model to generate a more realistic and reliable structural layout result.

[0055] Example 2

[0056] See also Figure 1 , Figure 1 The overall architecture and usage flow chart of an end-to-end AI building shear wall intelligent structure design system for implementing the method of Example 1. Figure 1 As shown in , the system provided by the embodiment of the present application includes:

[0057] S101: TransUnet-based generator. This paper designs a TransUnet-based generator that combines the advantages of Transformer's ability to learn global contextual relationships and U-net's ability to learn local detail features, ensuring that it can simultaneously learn both global and local layout features of the drawing.

[0058] Regarding step S101 above, the specific structure and implementation of the TransUnet-based generator are as follows. TransUnet combines ResNet-50 and ViT (Vision Transformer) in its encoder. ResNet-50 extracts features at three image scales, which are then skip-connected with the image information passed through the Transformer layer to produce the final result.

[0059] In the Transformer layer, an image is first converted into a 2D The size of each patch (window) is P×P. According to the size of the dataset image, 32×32 is the most appropriate. The number of patches is The spatial information of the patch is encoded by adding a position embedding to the patch:

[0060]

[0061] in is the patch embedding projection, E pos ∈R N×D Embed for position.

[0062] The Transformer encoder consists of L layers of Multi-Head Self-Attention (MSA) and Multi-Layer Perceptron (MLP) modules. Therefore, the output of the L-th layer can be expressed as follows:

[0063] z' l =MSA(LN(z l-1 ))+z l-1 (2)

[0064] z l =MSA(LN(z' l ))+z' l (3)

[0065] Where LN(·) represents the layer normalization operator, z l Represents the encoded image representation. The encoded feature representation Upsample to full resolution to predict dense output.

[0066] As a medical segmentation model, TransUnet cannot be used directly in the generative model. At the same time, its feature extraction effect is not ideal when facing shear wall data. By modifying the output layer, TransUnet was modified from a segmentation model to a generator model. In addition, the original three-scale splicing mechanism of TransUnet has the problem of unequal scale scaling, which leads to incomplete feature extraction. To this end, the present invention improves the architecture of the model. When the image is passed into the model, it will be upsampled at five sizes of 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the image size to ensure sufficient feature extraction. At different patch (window) sizes, the model will take into account different features. At large patch sizes, the model can learn the connection between the wall and surrounding features and learn a wider range of global features. At small patch sizes, especially the smallest patch size, the model can fully learn the layout features of the detailed parts of the wall, which helps the model generate a fine structural layout.

[0067] At each scale, the image will be extracted by the corresponding ResNet block, and then passed to the 12-layer Transformer layer, and the final output is the hidden features of the image. The sequence is reshaped into After the shape is formed, the tensor and the features processed by ResNet-50 at the previous 5 scales are skipped and spliced ​​to form a CNN and Transformer structure and a U-shaped structure. The feature aggregation of different resolution levels of H×W is achieved, and the context connection is established by taking advantage of the Transformer to learn global features. The structure diagram of the generator can be seen Figure 2 Through this improved TransUnet generator, the model can effectively extract global and local information features of shear wall data. Compared with traditional generators, the generator of the present invention can more comprehensively learn data features, taking into account both global and detailed features, and can generate structural layouts that are more similar to engineers.

[0068] S102, wavelet transform-based discriminator. The present invention matches the generator with a wavelet transform-based discriminator, which can more efficiently learn and identify high-frequency information features of wall-like structures in drawings.

[0069] Regarding step S102 above, the specific principles and implementation of the wavelet transform-based discriminator are as follows. In the shear wall structure plan drawing task, high-frequency information (such as the shear wall portion indicated in red on the drawing) is of greater concern, while low-frequency information (such as the blank areas on the drawing) is less important. However, previous discriminator models did not specifically distinguish between frequency information in images. Using wavelet transforms can provide frequency information to the discriminator, thereby better identifying the high-frequency content missing from the generated image.

[0070] The idea behind the wavelet transform is to first represent the original signal using a scaling function. As the scale increases, the representation becomes increasingly blurry and coarse. Therefore, a wavelet function is introduced to represent the difference between the scaling function representation and the original signal. This allows the scaling function (a low-pass filter) to describe the low-frequency components of the signal and extract the smoother parts. The wavelet function (a high-pass filter) then describes the high-frequency components of the signal and extracts the signal's details. Together, these two functions form the wavelet basis functions. This approach offers the advantage of localized analysis of the spatial frequency of an image and the ability to gradually subdivide the signal into multiple scales through specialized operations.

[0071]

[0072]

[0073] In the task here, Haar wavelet will be used as the basis function. The scaling function of Haar wavelet can be expressed by formulas (4) and (5). And the wavelet function of Haar wavelet can be expressed by formulas (6) and (7). The present invention will use wavelet transform to decompose the image into a series of channels, each channel representing the content of a different frequency range. First, a first-level wavelet decomposition is used to process the image using some low-pass and high-pass wavelet filters to decompose the image into four sub-bands: LL, LH, HL and HH. Among them, LL corresponds to low-frequency information, which is visually similar to a blurred version of the input image. The remaining sub-bands LH, HL and HH correspond to high-frequency content in the horizontal, vertical and diagonal directions respectively. Thus, the discriminator will not only consider the RGB space of the image, but also their wavelet decomposition.

[0074]

[0075] Discrete wavelet transform (DWT) extracts features by analyzing the local features of the signal at multiple scales. Discrete wavelet transform can be implemented by recursive decomposition, which is expressed by formula (8). Where x(t) is the given signal, c j，k is the coefficient of wavelet transform, ψ j，k(t) is the scale- and translation-transformed wavelet basis function, also known as the Haar wavelet. The inverse wavelet transform (IWT) is the inverse of the wavelet transform. In the discriminator, the image first undergoes a DWT. It is then processed alternately through the Wavelet block and the Convolution block to extract features at different scales.

[0076] The specific process is as follows: the image first enters the Wavelet block (wavelet block). In the Wavelet block, the wavelet coefficients are combined into a full image through the IWT (inverse wavelet transform). The image is then bilinearly downsampled, and the low-resolution image is decomposed back into wavelet coefficients through the discrete wavelet transform (DWT) and sent to the next block. At the same time, the results of each Wavelet block are processed by a ConvLayer (convolution layer) and a convolution block before being sent to the next Wavelet block for skip connection. During this process, the size of the image will continue to decrease, and the discriminator will extract features of the image at different scales through wavelets.

[0077] Together, the generator and discriminator form a Transformer-Wavelet architecture. This architecture allows the model to fully learn the global and local layout characteristics of shear wall data, paying particular attention to high-frequency information such as the wall itself. The generator and discriminator are trained against each other, ultimately generating a structural layout that closely matches the engineer's design.

[0078] S103, the post-processing process based on computer vision can repair defects of different sizes in the generated wall by alternately executing specially designed closing and clustering functions, thereby improving the continuity and flatness of the generated wall.

[0079] The process described in S103 will first obtain a mask of the possible shear wall generation area (i.e., the layout area of ​​all walls) based on the input image (i.e., the building drawing that does not contain shear walls). Then the present invention sets a closing function and a clustering function respectively. Unlike the common graphics closing operation, the present invention uses positive and negative maximum pooling to simulate the expansion and erosion operations, so that the tensor can be operated directly on the GPU, which greatly speeds up the processing speed. The clustering function will identify the mask area where the shear wall is generated, and cluster the colors that meet the conditions within the threshold and the colors that are similar after expansion into the colors belonging to the shear wall (in the data set, the present invention uses red to represent the shear wall). These two functions will be applied at multiple scales to achieve the best effect. The generated image will be closed under 4 convolution kernels of different sizes to deal with defects of different sizes in the generated shear wall. After four steps of layer-by-layer processing, the processed shear wall layout result will be generated. See the flowchart for detailed processing. Figure 4 .

[0080] Because the entire process is executed on the GPU, it does not add much additional time. Through this post-processing step, some of the generated fragmented walls can be connected together, significantly improving the continuity and flatness of the structural layout, making the generated structural layout drawings more realistic in application scenarios.

[0081] S104, the sketch drawing tool is designed based on the architect's drawing habits using HTML5 language. The tool can be executed on the web to achieve end-to-end structure generation.

[0082] The sketch drawing tool described in S104 provides two drawing methods based on the architect's drawing habits: coordinate drawing method and directional drawing method, which are the two most common drawing methods in CAD drawing. The coordinate drawing method draws a line segment by entering the coordinates of the starting point. The directional drawing method draws a line segment of a specific length according to the mouse direction and the input length. The sketch drawing tool provides 3 thicknesses of line segments to represent the thickness of different walls. At the same time, 4 colors are provided, namely red, gray, green, and blue to represent shear walls, ordinary walls, doors and windows. The sketch drawing tool also provides some additional functions: the interface will display the coordinates of the mouse in real time to determine the drawing coordinates. The endpoint coordinates will be automatically adsorbed within 10px of the line segment endpoint to facilitate quick and accurate positioning. Users can import pictures and adjust the transparency as a background to facilitate copying and drawing. In addition, the drawing tool also provides the function of undoing drawing. The tool interface can be found in detail at Figure 5 .

[0083] This tool allows architects to quickly create sketches and input them into the model using the drawing methods they are most familiar with. The shear wall structure layout is then calculated using the model, significantly reducing the barrier to entry for architects using the AutoStruct system. This also allows AutoStruct to form a complete usage process, enabling end-to-end structural layout generation.

[0084] Its working principle is: Based on the TransUnet generator. This paper designs a TransUnet-based generator, which combines the advantages of Transformer's ability to learn global contextual relationships and U-net's ability to learn local detail features, ensuring that it can simultaneously learn the global and local layout features of the drawing.

[0085] Wavelet-based discriminator. This invention matches the generator with a wavelet-based discriminator, enabling more efficient learning and identification of high-frequency features of wall-like structures in drawings. The combined Transformer-Wavelet architecture gives AutoStruct a significant advantage in learning the features of shear wall data, significantly improving model performance.

[0086] Computer vision-based post-processing process. To address the problems of discontinuous and irregular distribution in generated walls, this paper has designed a computer vision-based post-processing process. By alternating between specially designed closing and clustering functions, defects of varying sizes in the generated wall can be repaired, thereby improving the continuity and smoothness of the generated wall.

[0087] Sketching tool. The present invention uses HTML5 language and designs a sketching tool based on the architect's drawing habits. Architects can quickly draw architectural sketches on the web and input them into the model, and then obtain the structural layout generated by AutoStruct, which greatly reduces the threshold for using the model. Thus, AutoStruct forms a complete end-to-end workflow: architects draw simple sketches through sketching work and input them into the system, and the system will generate the shear wall structure layout through the pre-trained model. The results will be automatically input into the post-processing process for post-processing optimization, and finally the shear wall structure layout drawing will be output to the architect. The entire process will take no more than 2 minutes, and it is expected to increase efficiency by more than 60 times compared to manual communication.

[0088] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The scope of protection of the present invention should not be regarded as limited to the specific forms described in the embodiments. The scope of protection of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. An end-to-end intelligent building shear wall structure layout generation method, characterized in that: The steps include: S1: Obtaining a building layout sketch of the architect's preliminary design through a sketching tool; S2: Inputting the input building plan layout sketch into the pre-trained shear wall structure layout generation model to obtain the corresponding building shear wall structure layout feature vector; S3: Inputting the building shear wall structure layout feature vector into a post-processing step based on computer vision to make the generated structural wall more flat and continuous, and finally outputting the building shear wall structure layout diagram.

2. An end-to-end intelligent building shear wall structure layout generation method according to claim 1, characterized in that: The operation steps of the sketch drawing tool described in step S1 include: S11: Select one of two commonly used architectural drawing methods: coordinate drawing method and directional drawing method to draw; S12: Select the corresponding line segment thickness based on the three provided line segment thicknesses representing different wall thicknesses; and select the corresponding line segment thickness based on the four provided colors representing shear walls, ordinary walls, doors, and windows. S13: In the coordinate drawing method, a line segment is drawn by inputting the coordinates of the starting point; in the directional drawing method, a line segment of a specific length is drawn according to the mouse direction and the input length; S14: Use other features provided by the tool to assist in drawing: real-time display of the mouse coordinates to determine the drawing coordinates; automatic absorption of the endpoint coordinates within 10px of the line segment endpoint for quick and accurate positioning; users can import pictures and adjust the transparency as backgrounds to facilitate copying and drawing; use the undo function to undo incorrect drawings; S15: Complete the sketch drawing and output the building layout sketch.

3. An end-to-end intelligent building shear wall structure layout generation method according to claim 1, characterized in that: The training steps of the shear wall structure arrangement generation model described in step S2 include: S21: Obtain a training sample set; wherein the training sample set includes sample pairs of shear wall building layout drawings and shear wall structural layout drawings; S22: Inputting the sample pairs into a TransUnet-based generator to fully learn the global and local layout features of the drawing and generate a corresponding shear wall structure layout drawing; S23: Input the corresponding shear wall structure layout diagram and the shear wall structure layout diagram in the training set into a wavelet transform-based discriminator to determine whether the input image is an image from the generator; the wavelet transform-based discriminator will focus on the high-frequency information of the wall in the image; S24: Perform adversarial generative training based on the TransUnet generator and the wavelet transform-based discriminator, calculate the VGG loss and the LSGAN loss, and update the parameters of the generator through error backpropagation to optimize the model to generate a more realistic and reliable structural layout result.

4. The end-to-end intelligent building shear wall structure layout generation method according to claim 3, characterized in that: The TransUnet-based generator described in step S22 specifically includes: S221: TransUnet, a model that combines the advantages of Transformer and U-net, is used as the basis of the generator; S222: TransUnet was modified from a segmentation model to a generator model by modifying the output layer. The original three-scale splicing mechanism of TransUnet was improved, which had the problem of unequal scale scaling. When the image is passed to the model, it will be upsampled at five sizes of 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the image size to ensure sufficient feature extraction. S223: At different window sizes, the model will take into account different features. At large window sizes, the model can learn the connection between the wall and surrounding features, learning a wider range of global features. At small window sizes, the model can fully learn the layout features of the wall's detailed parts, helping the model generate a detailed structural layout. At each scale, the image will be feature extracted by the corresponding ResNet block, and then passed to the 12-layer Transformer layer, and the final output is the hidden features of the image. S224: The image hidden features are finally converted into image tensor output through convolution.

5. The end-to-end intelligent building shear wall structure layout generation method according to claim 3, characterized in that: The wavelet transform-based discriminator described in step S33 specifically includes: S331: Using Haar wavelet as basis function; using wavelet transform to decompose the image into a series of channels, each channel represents the content of a different frequency range; using a first-level wavelet decomposition, using a number of low-pass and high-pass wavelet filters to process the image, the image is decomposed into four sub-bands: LL, LH, HL and HH; where LL corresponds to low-frequency information, which is visually similar to a blurred version of the input image; the remaining sub-bands LH, HL and HH correspond to high-frequency content in the horizontal, vertical and diagonal directions respectively; thus, the discriminator will not only consider the RGB space of the image, but also their wavelet decomposition; S332: The image first enters the wavelet block; in the Wavelet block, the wavelet coefficients are combined into a full image through the inverse wavelet transform IWT, and then the image is bilinearly downsampled, and the low-resolution image is decomposed back into wavelet coefficients through the discrete wavelet transform DWT, and sent to the next block; at the same time, the result of each Wavelet block will be processed by a convolution layer and pass through a convolution block, and then sent to the skip connection with the result extracted by the next Wavelet block; in this process, the size of the image will continue to decrease, and the discriminator will extract the features of the image at different scales through wavelets.

6. An end-to-end intelligent building shear wall structure layout generation method according to claim 1, characterized in that: The computer vision-based post-processing operation described in step S3 includes: S31: Based on the input building drawing that does not contain shear walls, a mask of the area where shear walls may be generated is obtained; and at the same time, a structural layout feature vector of the building shear walls is input; S32: Based on a closure function and a clustering function, respectively, positive and negative maximum pooling are used to simulate dilation and erosion operations. The clustering function will identify the masked area of ​​the generated shear wall and cluster the colors that meet the conditions within the threshold and the colors that are similar after dilation as the colors belonging to the shear wall. These two functions will be applied at multiple scales to achieve the best effect. The generated image will be closed under four convolution kernels of different sizes to deal with defects of different sizes in the generated shear wall. S33: After four steps of layer-by-layer processing, the processed shear wall structure layout drawing will be generated.

7. An end-to-end AI building shear wall intelligent structure design system that implements the end-to-end intelligent building shear wall structure layout generation method according to claim 1, characterized in that: Includes AutoStruct, and: The TransUnet-based generator learns both global and local layout features of the drawings; A wavelet-based discriminator learns and identifies high-frequency information features of walls in drawings, forming a Transformer-Wavelet architecture. This gives AutoStruct a significant advantage in learning the feature information of shear wall data. A post-processing process based on computer vision is designed to address the problems of discontinuous and irregular distribution in the generated walls. By alternating the specially designed closing and clustering functions, defects of different sizes in the generated walls are repaired, thereby improving the continuity and flatness of the generated walls. Sketching tools, designed in HTML5, allow architects to quickly sketch buildings on the web and input them into the model, which then generates the structural layout generated by AutoStruct. This creates a complete end-to-end workflow for AutoStruct: architects simply sketch and input the system, which then generates the shear wall structure layout using a pre-trained model. The results will be automatically input into the post-processing process for post-processing optimization, and finally the shear wall structure layout drawing will be output to the architect.

Images