Stiffened steel plate shear wall design method based on image-text feature fusion of generative adversarial network
The proposed method addresses the challenge of integrating image and text data for structural design of reinforced steel plate shear walls using a generative adversarial network that integrates image and text features, resulting in structural solutions that meet design requirements and match classical topology optimization in performance.
Patent Information
- Application Number
- CN202411091737.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-08-09
AI Technical Summary
Existing generative adversarial networks struggle to simultaneously learn image and text data for structural design of reinforced steel plate shear walls, making it difficult to meet both structural image features and text requirements.
A design method for reinforced steel plate shear walls using a generative adversarial network that integrates image-text features, involving pre-processing of S.Mises and ESEDEN diagrams, classification, and a BraceGAN-TXT algorithm for intelligent generation and evaluation, followed by detailed design of stiffening ribs.
The method effectively generates structural solutions that meet design requirements, with design results comparable to classical topology optimization in plane design indicators and overall structural mechanical properties.
Smart Images

Figure CN119089770A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of civil engineering and computer deep learning application technology, and in particular to a design method for reinforced steel plate shear walls based on a generative adversarial network and a fusion image-text feature. Background Art
[0002] Although computers can learn the abstract semantic information contained in images, they cannot learn the abstract semantic information contained in the text of structural design conditions. Existing studies have shown that in addition to the input of building images, it is often necessary to input the text of seismic design conditions to guide the structural design in order to meet the requirements of seismic design. At present, it is difficult for existing generative adversarial networks to learn image data and text data at the same time. Therefore, it is necessary to conduct more in-depth research on the design method of reinforced steel plate shear walls to achieve the structural design goal of simultaneously meeting the structural image features and text requirements. Summary of the invention
[0003] Purpose of the invention: The technical problem to be solved by the present invention is to overcome the shortcomings of the existing design method of stiffened steel plate shear wall based on generative adversarial network and propose a design method of stiffened steel plate shear wall based on image-text features fusion based on generative adversarial network. This method can generate a stiffening rib arrangement diagram that meets the target design requirements by inputting steel plate shear wall training images and design texts. Compared with the existing design method of stiffened steel plate shear wall based on generative adversarial network, this method shows more outstanding design generation capabilities, and can effectively generate structural solutions that meet design requirements according to text instructions. Its design results are comparable to the classical topology optimization design method in terms of plane design indicators and overall structural mechanical properties.
[0004] Technical solution, in order to solve the above-mentioned technical problems, the present invention proposes a design method for reinforced steel plate shear wall based on image-text feature fusion of generative adversarial network, the method comprising the following steps:
[0005] Step S1, taking the S.Mises diagram and ESEDEN diagram of the steel plate shear wall to be designed, and determining its classification group;
[0006] Step S2, inputting the S.Mises diagram and ESEDEN diagram of the steel plate shear wall to be designed into the pre-trained BraceGAN-TXT to obtain the BraceGAN-TXT intelligently generated design diagram;
[0007] Step S3, based on the intelligently generated design drawing, perform in-depth design to obtain the final stiffening rib design.
[0008] Furthermore, in step S1, the production process of the S.Mises graph and the ESEDEN graph is as follows:
[0009] (2.1) According to the steel plate shear wall structure design drawings, the corresponding steel plate shear wall model is established in the finite element analysis software ABAQUS;
[0010] (2.2) Create field output in the analysis step module, with the scope being the middle steel plate of the steel plate shear wall, and the output variables being set to MISES, Mises equivalent stress and ELEDEN, all energy density components;
[0011] (2.3) Create and submit the Job file, run it to completion and open the Job.odb file, draw a cloud map on the undeformed map in the visualization module, and obtain the S.Mises map and ESEDEN map of the middle steel plate.
[0012] Furthermore, the modeling method in step (2.1) is:
[0013] (3.1) In terms of finite element modeling unit selection, the steel plate embedded in the middle of the steel plate shear wall belongs to the plate-shell structure, so the linear reduced thin shell unit is selected;
[0014] (3.2) The initial defect distribution adopts the deformation distribution pattern of the first-order elastic mode of the component, and the amplitude of the defect is determined based on actual measurements.
[0015] Furthermore, the specific method for determining the grouping category in step S1 is as follows:
[0016] The height and aspect ratio of the steel plate shear wall determine whether it belongs to Class N or Class W in the classification group. "N" and "M" are classifications of design text conditions. "N" represents steel plates with a height less than 2600mm and a aspect ratio less than 1.5 or a height greater than 2600mm and a aspect ratio less than 1.2; "W" represents steel plates with a height less than 2600mm and a aspect ratio greater than 1.5 or a height greater than 2600mm and a aspect ratio greater than 1.2.
[0017] Furthermore, the following processing is required before inputting the pre-trained BraceGAN-TXT in step S2:
[0018] (5.1) Use the Python Imaging Library (PIL) to convert the S.Mises and ESEDEN images into PNG format, and set the pixel size to 2048p × 1024p.
[0019] (5.2) Take the ESEDEN graph as train_A; take the S.Mises graph as train_H.
[0020] Furthermore, the network architecture of the BraceGAN-TXT algorithm in step S2 is as follows:
[0021] (6.1) Generator, which accepts image and text data as input, uses convolutional neural network to extract high-dimensional features from the encoded image, and uses FastText to extract high-dimensional feature representation of text information;
[0022] (6.2) fusing the extracted image feature data and text feature data in a residual network (ResNet) to form fused image-text feature data, and fusing the text feature data with the fused image-text feature data again to enhance the expression effect of the features;
[0023] (6.3) The target image is generated using a deconvolutional neural network. The discriminator combines the text feature data with the image and text data input into the generator to determine whether the generated image is real or fake.
[0024] Furthermore, the text encoding and feature extraction process in the BraceGAN-TXT algorithm in step S2 is as follows:
[0025] (7.1) Extract keywords from the train_A and train_H image file names;
[0026] (7.2) Based on the keywords, if the image name contains "N", a sentence list containing the description of "N" is generated; if the image name contains "W", a sentence list containing the description of "W" is generated; the sentence list is converted into word vectors using the embedding function of FastText, and the word vectors and description lengths are saved in a .pth file. The file name is associated with the corresponding image name, and the folder is named train_C;
[0027] (7.3) In the encoder, the average hidden state of the text sequence is calculated based on the text data and the length of the text sequence, and the latent variables are predicted using a neural network to obtain text feature data; the text feature data is normalized and the dimension of the text feature data is made to match the dimension of the image feature data; the image feature data and the text feature data are concatenated together and passed to the residual block network "Residual blocks" to obtain the final merged feature merge;
[0028] (7.4) Add the output features of the encoder and the merged features to obtain a temporary feature representation; concatenate the temporary features and the text feature data together to obtain a concatenated feature tensor; use the concatenated feature tensor as input and decode it with the decoder to obtain the final output result;
[0029] (7.5) Transform the text feature data in dimension so that the text feature data and the image and text data input into the generator can match each other in dimension; concatenate the image and text data input into the generator and the text feature data together to form a comprehensive input feature, and obtain the final output result through the discriminator model.
[0030] Furthermore, the evaluation method and index for the BraceGAN-TXT intelligent generation design in step S2 are:
[0031] (8.1) An evaluation method and index suitable for BraceGAN-TXT are proposed, namely Score IoU , and Score IoU Depend on WIoU, SIoU are composed of three indicators. As a reduction factor to control the number of generated entity units, it represents the control coefficient of the proportion of stiffening ribs; WIoU measures the weighted generation of the four elements in the image and evaluates the pixel-level accuracy of the generated image; SIoU is used to measure the consistency between the generated design and the target design in the design of the stiffened steel plate shear wall structure, Score IoU The evaluation index evaluates the design from two aspects: the number of stiffeners and the rationality of their layout. IoU Greater than 0.5 indicates good design performance;
[0032] (8.2) Use formula (1) to calculate Score IoU , calculated using formula (2)
[0033] Score IoU =η SRratio ×(η SIoU ×SIoU+η WIoU ×WIoU) (1)
[0034]
[0035] Among them, WIoU and SIoU are weighted intersection-over-union ratio and stiffener contour intersection-over-union ratio, respectively; η SIoU and η WIoU They represent the weight coefficients of SIoU and WIoU, both of which are 0.5; Indicates the total amount of stiffeners in the steel plate shear wall in the BraceGAN design drawing. represents the total amount of stiffeners in the steel plate shear wall in the target design;
[0036] (8.3) Calculate WIoU using formula (3);
[0037]
[0038] Among them, k represents the category, that is, category 0 is the background, category 1 is the entity unit, category 2 is the virtual unit, and category 3 is the black grid line; i represents the true value, j represents the predicted value; P ij It means that i is predicted as j, which is a false negative, that is, FN; P ji It means that j is predicted as i, which is a false positive, or FP; P ii It means that i is predicted as i, which is true, that is, TP; w i is the weight factor, i.e. w 0 =0,w 1 =0.5, w 2 =0.4, w 3 =0.1;
[0039] (8.4) Calculate SIoU using formula (4);
[0040]
[0041] In the formula, A inter represents the intersection area of the entity units in the BraceGAN design and the target design, A union represents the union area of the solid units in the BraceGAN design and the target design, A union =A target +A GAN -A inter , A target Represents the area of the solid unit in the target design diagram, A GAN Represents the area of the solid unit in the BraceGAN design diagram.
[0042] Furthermore, the specific implementation process of the optimization scheme deepening design in step S3 is as follows:
[0043] (9.1) Design the plan layout of stiffeners for steel plate shear walls based on BraceGAN-TXT: Obtain the S.Mises diagram and ESEDEN diagram of the steel plate filled in the steel plate shear wall required for the design, call the PIL library (Python Imaging Library) in Python to convert the resolution of the image to 2048*1024, and classify it according to the classification standard. If it belongs to category N, add the prefix "N_" to the image file name. If it belongs to category W, add the prefix "W_" to the image file name. Input the modified image into BraceGAN-TXT, and output the stiffener layout diagram.
[0044] (9.2) Detailed design based on stiffening rib arrangement diagram: The equivalent bending stiffness method is used to carry out detailed design of stiffening ribs. When the post-buckling strength design of the stiffened steel plate shear wall is considered and the stiffening ribs are steel strips, the stiffening rib width-to-thickness ratio complies with the provisions of formula (5), where λs is the stiffener width-to-thickness ratio, which is the ratio of the stiffener plate's overhanging width to its thickness. The equivalent bending stiffness method is used to adjust the size of the stiffener so that the section moment of inertia of the stiffener before and after the design remains consistent. The section moment of inertia I is calculated using formula (6): s ;
[0045] 6≤λ s ≤12 (5)
[0046]
[0047] In the formula, h s Indicates the overhang width of the stiffening rib plate, t s Indicates the thickness of the stiffener plate;
[0048] (9.3) Construct the corresponding analysis model of the stiffened steel plate shear wall: Use ABAQUS finite element analysis software to model the structure and conduct performance evaluation on the buckling mode, bearing capacity, stiffness, out-of-plane deformation and energy dissipation capacity of the model.
[0049] Beneficial effects: Compared with the prior art, the present invention adopts the above technical solution to have the following beneficial technical effects:
[0050] (1) The invention proposes a design method for reinforced steel plate shear walls based on a generative adversarial network that integrates image and text features. The method applies deep learning methods in artificial intelligence to structural design, designs text encoding and feature extraction processes, and constructs a network structure that can process high-dimensional abstract features of images and texts, thus solving the problem that existing generative adversarial networks are difficult to learn image data and text data at the same time.
[0051] (2) The present invention utilizes an improved and optimized generative adversarial network model, which can effectively generate a structural solution that meets the design requirements according to text instructions. Its design results are comparable to those of the classical topology optimization design method in terms of plane design indicators and overall structural mechanical properties, and have strong practical significance. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:
[0053] Figure 1 Schematic diagram of the framework of the BraceGAN-TXT method in the method of the present invention;
[0054] Figure 2 Schematic diagram of the BraceGAN-TXT network architecture in the method of the present invention;
[0055] Figure 3 The standard floor plan of the structure of the example used to verify the effectiveness of the method of the present invention;
[0056] Figure 4 The plan layout of the steel plate shear wall used as an example to verify the effectiveness of the method of the present invention;
[0057] Figure 5 The elevation diagram of the steel plate shear wall used as an example to verify the effectiveness of the method of the present invention;
[0058] Figure 6 A schematic diagram of the design results of the stiffeners generated by BraceGAN-TXT in order to verify the effectiveness of the method of the present invention;
[0059] Figure 7 A schematic diagram of the design results of the stiffening ribs in the topology optimization design of the example for verifying the effectiveness of the method of the present invention;
[0060] Figure 8 The present invention is a flow chart of the method. DETAILED DESCRIPTION
[0061] In order to make the technical solution, advantages and ultimate purpose of the present invention clearer, the present invention will be further explained and illustrated in conjunction with the accompanying drawings and specific implementation methods.
[0062] like Figure 8 As shown, the present invention proposes a design method for reinforced steel plate shear wall based on image-text feature fusion based on generative adversarial network, and the method comprises the following steps:
[0063] Step S1, taking the S.Mises diagram and ESEDEN diagram of the steel plate shear wall to be designed, and determining its classification group;
[0064] Step S2, inputting the S.Mises diagram and ESEDEN diagram of the steel plate shear wall to be designed into the pre-trained BraceGAN-TXT to obtain the BraceGAN-TXT intelligently generated design diagram;
[0065] Step S3, based on the intelligently generated design drawing, perform in-depth design to obtain the final stiffening rib design.
[0066] Furthermore, in step S1, the production process of the S.Mises graph and the ESEDEN graph is as follows:
[0067] (2.1) According to the steel plate shear wall structure design drawings, the corresponding steel plate shear wall model is established in the finite element analysis software ABAQUS;
[0068] (2.2) Create field output in the analysis step module, with the scope being the middle steel plate of the steel plate shear wall, and the output variables being set to MISES, Mises equivalent stress and ELEDEN, all energy density components;
[0069] (2.3) Create and submit the Job file, run it to completion and open the Job.odb file, draw a cloud map on the undeformed map in the visualization module, and obtain the S.Mises map and ESEDEN map of the middle steel plate.
[0070] Furthermore, the modeling method in step (2.1) is:
[0071] (3.1) In terms of finite element modeling unit selection, the steel plate embedded in the middle of the steel plate shear wall belongs to the plate-shell structure, so the linear reduced thin shell unit is selected;
[0072] (3.2) The initial defect distribution adopts the deformation distribution pattern of the first-order elastic mode of the component, and the amplitude of the defect is determined based on actual measurements.
[0073] Furthermore, the specific method for determining the grouping category in step S1 is as follows:
[0074] The height and aspect ratio of the steel plate shear wall determine whether it belongs to Class N or Class W in the classification group. "N" and "M" are classifications of design text conditions. "N" represents steel plates with a height less than 2600mm and a aspect ratio less than 1.5 or a height greater than 2600mm and a aspect ratio less than 1.2; "W" represents steel plates with a height less than 2600mm and a aspect ratio greater than 1.5 or a height greater than 2600mm and a aspect ratio greater than 1.2.
[0075] Furthermore, the following processing is required before inputting the pre-trained BraceGAN-TXT in step S2:
[0076] (5.1) Use the Python Imaging Library (PIL) to convert the S.Mises and ESEDEN images into PNG format, and set the pixel size to 2048p × 1024p.
[0077] (5.2) Take the ESEDEN graph as train_A; take the S.Mises graph as train_H.
[0078] Furthermore, the network architecture of the BraceGAN-TXT algorithm in step S2 is as follows:
[0079] (6.1) Generator, which accepts image and text data as input, uses convolutional neural network to extract high-dimensional features from the encoded image, and uses FastText to extract high-dimensional feature representation of text information;
[0080] (6.2) fusing the extracted image feature data and text feature data in a residual network (ResNet) to form fused image-text feature data, and fusing the text feature data with the fused image-text feature data again to enhance the expression effect of the features;
[0081] (6.3) The target image is generated using a deconvolutional neural network. The discriminator combines the text feature data with the image and text data input into the generator to determine whether the generated image is real or fake.
[0082] Furthermore, the text encoding and feature extraction process in the BraceGAN-TXT algorithm in step S2 is as follows:
[0083] (7.1) Extract keywords from the train_A and train_H image file names;
[0084] (7.2) Based on the keywords, if the image name contains "N", a sentence list containing the description of "N" is generated; if the image name contains "W", a sentence list containing the description of "W" is generated; the sentence list is converted into word vectors using the embedding function of FastText, and the word vectors and description lengths are saved in a .pth file. The file name is associated with the corresponding image name, and the folder is named train_C;
[0085] (7.3) In the encoder, the average hidden state of the text sequence is calculated based on the text data and the length of the text sequence, and the latent variables are predicted using a neural network to obtain text feature data; the text feature data is normalized and the dimension of the text feature data is made to match the dimension of the image feature data; the image feature data and the text feature data are concatenated together and passed to the residual block network "Residual blocks" to obtain the final merged feature merge;
[0086] (7.4) Add the output features of the encoder and the merged features to obtain a temporary feature representation; concatenate the temporary features and the text feature data together to obtain a concatenated feature tensor; use the concatenated feature tensor as input and decode it with the decoder to obtain the final output result;
[0087] (7.5) Transform the text feature data in dimension so that the text feature data and the image and text data input into the generator can match each other in dimension; concatenate the image and text data input into the generator and the text feature data together to form a comprehensive input feature, and obtain the final output result through the discriminator model.
[0088] Furthermore, the evaluation method and index for the BraceGAN-TXT intelligent generation design in step S2 are:
[0089] (8.1) An evaluation method and index suitable for BraceGAN-TXT are proposed, namely Score IoU , and Score IoU By η SRratio , WIoU, SIoU three indicators, η SRratio As a reduction factor to control the number of generated entity units, it represents the control coefficient of the proportion of stiffening ribs; WIoU measures the weighted generation of the four elements in the image and evaluates the pixel-level accuracy of the generated image; SIoU is used to measure the consistency between the generated design and the target design in the design of the stiffened steel plate shear wall structure, Score IoU The evaluation index evaluates the design from two aspects: the number of stiffeners and the rationality of their layout. IoU Greater than 0.5 indicates good design performance;
[0090] (8.2) Use formula (1) to calculate Score IoU , calculated using formula (2)
[0091] Score IoU =η SRratio ×(η SIoU ×SIoU+η WIoU ×WIoU) (1)
[0092]
[0093] Among them, WIoU and SIoU are weighted intersection-over-union ratio and stiffener contour intersection-over-union ratio, respectively; η SIoU and η WIoU They represent the weight coefficients of SIoU and WIoU, both of which are 0.5; Indicates the total amount of stiffeners in the steel plate shear wall in the BraceGAN design drawing. represents the total amount of stiffeners in the steel plate shear wall in the target design;
[0094] (8.3) Calculate WIoU using formula (3);
[0095]
[0096] Among them, k represents the category, that is, category 0 is the background, category 1 is the entity unit, category 2 is the virtual unit, and category 3 is the black grid line; i represents the true value, j represents the predicted value; P ij It means that i is predicted as j, which is a false negative, that is, FN; P ji It means that j is predicted as i, which is a false positive, or FP; P ii It means that i is predicted as i, which is true, that is, TP; wi is the weight factor, i.e. w 0 =0,w 1 =0.5, w 2 =0.4, w 3 =0.1;
[0097] (8.4) Calculate SIoU using formula (4);
[0098]
[0099] In the formula, A inter represents the intersection area of the entity units in the BraceGAN design and the target design, A union represents the union area of the solid units in the BraceGAN design and the target design, A union =A target +A GAN -A inter , A target Represents the area of the solid unit in the target design diagram, A GAN Represents the area of the solid unit in the BraceGAN design diagram.
[0100] Furthermore, the specific implementation process of the optimization scheme deepening design in step S3 is as follows:
[0101] (9.1) Design the plan layout of stiffeners for steel plate shear walls based on BraceGAN-TXT: Obtain the S.Mises diagram and ESEDEN diagram of the steel plate filled in the steel plate shear wall required for the design, call the PIL library (Python Imaging Library) in Python to convert the resolution of the image to 2048*1024, and classify it according to the classification standard. If it belongs to category N, add the prefix "N_" to the image file name. If it belongs to category W, add the prefix "W_" to the image file name. Input the modified image into BraceGAN-TXT, and output the stiffener layout diagram.
[0102] (9.2) Detailed design based on stiffening rib arrangement diagram: The equivalent bending stiffness method is used to carry out detailed design of stiffening ribs. When the post-buckling strength design of the stiffened steel plate shear wall is considered and the stiffening ribs are steel strips, the stiffening rib width-to-thickness ratio complies with the provisions of formula (5), where λ s is the stiffener width-to-thickness ratio, which is the ratio of the stiffener plate's overhanging width to its thickness. The equivalent bending stiffness method is used to adjust the size of the stiffener so that the section moment of inertia of the stiffener before and after the design remains consistent. The section moment of inertia I is calculated using formula (6): s ;
[0103] 6≤λ s ≤12 (5)
[0104]
[0105] In the formula, h s Indicates the overhang width of the stiffening rib plate, t s Indicates the thickness of the stiffener plate;
[0106] (9.3) Construct the corresponding analysis model of the stiffened steel plate shear wall: Use ABAQUS finite element analysis software to model the structure and conduct performance evaluation on the buckling mode, bearing capacity, stiffness, out-of-plane deformation and energy dissipation capacity of the model.
[0107] In order to verify the performance of the present invention, a calculation example of steel plate shear wall structure design in the national building standard design drawing collection was selected, and the present invention was applied to actual engineering to design stiffening ribs. Figure 3 This is the standard floor plan of the structure. Figure 4 This is the plan layout of the steel plate shear wall. Figure 5 This is the elevation drawing of the steel plate shear wall. Figure 6 Detailed design results of the stiffeners designed for BraceGAN-TXT generation, Figure 7 The results of the in-depth design of the stiffening ribs for topology optimization design are shown in Table 1. Table 1 is a quantitative analysis of the performance of the intelligent generation design of the example BraceGAN-TXT, and Table 2 is a comparison of the mechanical properties of the BraceGAN-TXT generation design and the topology optimization design. Combining Table 1 and Table 2, it can be seen that the difference in mechanical properties of the two is within 10%. Through the intelligent plane and overall structural design analysis of the design case of the stiffened steel plate shear wall, it is found that the algorithm can effectively generate a structural solution that meets the design requirements according to text instructions. The design results are comparable to the classical topology optimization design method in terms of plane design indicators and overall structural mechanical properties, which further verifies the effectiveness of the stiffened steel plate shear wall design method based on the generative adversarial network fusion image-text feature proposed in this invention.
[0108] Table 1 Quantitative analysis of the performance of BraceGAN-TXT intelligent generation design
[0109]
[0110] Table 2 Comparison of mechanical properties between BraceGAN-TXT generative design and topology optimization design
[0111]
[0112]
[0113] The specific embodiment described above is only a preferred solution of the present invention, but it is not intended to limit the present invention. Any replacement or equivalent transformation method that can be easily thought of by any technician familiar with the present technical field within the technical scope disclosed by the present invention should be included in the protection scope of the present invention.
Claims
1. A design method for reinforced steel plate shear wall based on image-text feature fusion based on generative adversarial network, characterized in that: The method comprises the following steps: Step S1, taking the S.Mises diagram and ESEDEN diagram of the steel plate shear wall to be designed, and determining its classification group; Step S2, inputting the S.Mises diagram and ESEDEN diagram of the steel plate shear wall to be designed into the pre-trained BraceGAN-TXT to obtain the BraceGAN-TXT intelligently generated design diagram; Step S3, based on the intelligently generated design drawing, perform in-depth design to obtain the final stiffening rib design.
2. According to claim 1, a design method for reinforced steel plate shear wall based on image-text feature fusion based on generative adversarial network is characterized in that: In step S1, the production process of S.Mises graph and ESEDEN graph is as follows: (2.1) According to the steel plate shear wall structure design drawings, the corresponding steel plate shear wall model is established in the finite element analysis software ABAQUS; (2.2) Create field output in the analysis step module, with the scope being the middle steel plate of the steel plate shear wall, and the output variables being set to MISES, Mises equivalent stress and ELEDEN, all energy density components; (2.3) Create and submit the Job file, run it to completion and open the Job.odb file, draw a cloud map on the undeformed map in the visualization module, and obtain the S.Mises map and ESEDEN map of the middle steel plate.
3. The design method of reinforced steel plate shear wall based on image-text feature fusion based on generative adversarial network according to claim 2 is characterized in that: The modeling method in step (2.1) is: (3.1) In terms of finite element modeling unit selection, the steel plate embedded in the middle of the steel plate shear wall belongs to the plate-shell structure, so the linear reduced thin shell unit is selected; (3.2) The initial defect distribution adopts the deformation distribution pattern of the first-order elastic mode of the component, and the amplitude of the defect is determined based on actual measurements.
4. The design method of reinforced steel plate shear wall based on image-text feature fusion based on generative adversarial network according to claim 1 is characterized in that: The specific method for determining the grouping category in step S1 is as follows: The height and aspect ratio of the steel plate shear wall determine whether it belongs to Class N or Class W in the classification group. "N" and "M" are classifications of design text conditions. "N" represents steel plates with a height less than 2600mm and an aspect ratio less than 1.5 or a height greater than 2600mm and an aspect ratio less than 1.2; "W" represents steel plates with a height less than 2600mm and an aspect ratio greater than 1.5 or a height greater than 2600mm and an aspect ratio greater than 1.
2.
5. The design method of reinforced steel plate shear wall based on image-text feature fusion based on generative adversarial network according to claim 1 is characterized in that: The following processing is required before inputting the pre-trained BraceGAN-TXT in step S2: (5.1) Use the Python Imaging Library (PIL) to convert the S.Mises and ESEDEN images into PNG format, and set the pixel size to 2048p × 1024p. (5.2) Take the ESEDEN graph as train_A; take the S.Mises graph as train_H.
6. The design method of reinforced steel plate shear wall based on image-text feature fusion based on generative adversarial network according to claim 1 is characterized in that: The network architecture of the BraceGAN-TXT algorithm in step S2 is as follows: (6.1) Generator, which accepts image and text data as input, uses convolutional neural network to extract high-dimensional features from the encoded image, and uses FastText to extract high-dimensional feature representation of text information; (6.2) fusing the extracted image feature data and text feature data in a residual network (ResNet) to form fused image-text feature data, and fusing the text feature data with the fused image-text feature data again to enhance the expression effect of the features; (6.3) The target image is generated using a deconvolutional neural network. The discriminator combines the text feature data with the image and text data input into the generator to determine whether the generated image is real or fake.
7. The design method of reinforced steel plate shear wall based on image-text feature fusion based on generative adversarial network according to claim 1 is characterized in that: The text encoding and feature extraction process in the BraceGAN-TXT algorithm in step S2 is as follows: (7.1) Extract keywords from the train_A and train_H image file names; (7.2) Based on the keywords, if the image name contains "N", a sentence list containing the description of "N" is generated; if the image name contains "W", a sentence list containing the description of "W" is generated; the sentence list is converted into word vectors using the embedding function of FastText, and the word vectors and description lengths are saved in a .pth file. The file name is associated with the corresponding image name, and the folder is named train_C; (7.3) In the encoder, the average hidden state of the text sequence is calculated based on the text data and the length of the text sequence, and the latent variables are predicted using a neural network to obtain text feature data; the text feature data is normalized and the dimension of the text feature data matches the dimension of the image feature data; the image feature data and the text feature data are concatenated together and passed to the residual block network "Residual blocks" to obtain the final merged feature merge; (7.4) Add the output features of the encoder and the merged features to obtain a temporary feature representation; concatenate the temporary features and the text feature data together to obtain a concatenated feature tensor; use the concatenated feature tensor as input and decode it with the decoder to obtain the final output result; (7.5) Transform the text feature data in dimension so that the text feature data and the image and text data input into the generator can match each other in dimension; concatenate the image and text data input into the generator and the text feature data together to form a comprehensive input feature, and obtain the final output result through the discriminator model.
8. The design method of reinforced steel plate shear wall based on image-text feature fusion based on generative adversarial network according to claim 1 is characterized in that: The evaluation method and index for the BraceGAN-TXT intelligent generation design in step S2 are: (8.1) An evaluation method and index suitable for BraceGAN-TXT are proposed, namely Score IoU , and Score IoU Depend on WIoU, SIoU are composed of three indicators. As a reduction factor to control the number of generated entity units, it represents the control coefficient of the proportion of stiffening ribs; WIoU measures the weighted generation of the four elements in the image and evaluates the pixel-level accuracy of the generated image; SIoU is used to measure the consistency between the generated design and the target design in the design of the stiffened steel plate shear wall structure, Score IoU The evaluation index evaluates the design from two aspects: the number of stiffeners and the rationality of their layout. IoU Greater than 0.5 indicates good design performance; (8.2) Use formula (1) to calculate Score IoU , calculated using formula (2) Score IoU =the SRratio ×(the SIoU ×SIoU+η WIoU ×WIoU) (1) Among them, WIoU and SIoU are weighted intersection-over-union ratio and stiffener contour intersection-over-union ratio, respectively; η SIoU and η WIoU They represent the weight coefficients of SIoU and WIoU, both of which are 0.5; Indicates the total amount of stiffeners in the steel plate shear wall in the BraceGAN design drawing. represents the total amount of stiffeners in the steel plate shear wall in the target design; (8.3) Calculate WIoU using formula (3); Among them, k represents the category, that is, category 0 is the background, category 1 is the entity unit, category 2 is the virtual unit, and category 3 is the black grid line; i represents the true value, j represents the predicted value; P ij It means that i is predicted as j, which is a false negative, that is, FN; P ji It means that j is predicted as i, which is a false positive, or FP; P ii It means that i is predicted as i, which is true, that is, TP; w i is the weight factor, i.e. w0 = 0, w1 = 0.5, w2 = 0.4, w3 = 0.1; (8.4) Calculate SIoU using formula (4); In the formula, A inter represents the intersection area of the entity units in the BraceGAN design and the target design, A union represents the union area of the solid units in the BraceGAN design and the target design, A union =A target +A GAN -A inter , A target Represents the area of the solid unit in the target design diagram, A GAN Represents the area of the solid unit in the BraceGAN design diagram.
9. The design method of reinforced steel plate shear wall based on image-text feature fusion based on generative adversarial network according to claim 1, characterized in that: The specific implementation process of the optimization scheme deepening design in step S3 is as follows: (9.1) Design the plan layout of stiffeners for steel plate shear walls based on BraceGAN-TXT: Obtain the S.Mises diagram and ESEDEN diagram of the steel plate filled in the steel plate shear wall required for the design, call the PIL library (Python Imaging Library) in Python to convert the resolution of the image to 2048*1024, and classify it according to the classification standard. If it belongs to category N, add the prefix "N_" to the image file name. If it belongs to category W, add the prefix "W_" to the image file name. Input the modified image into BraceGAN-TXT, and output the stiffener layout diagram. (9.2) Detailed design based on stiffening rib arrangement diagram: The equivalent bending stiffness method is used to carry out detailed design of stiffening ribs. When the post-buckling strength design of the stiffened steel plate shear wall is considered and the stiffening ribs are steel strips, the stiffening rib width-to-thickness ratio complies with the provisions of formula (5), where λ s is the stiffener width-to-thickness ratio, which is the ratio of the stiffener plate's overhanging width to its thickness. The equivalent bending stiffness method is used to adjust the size of the stiffener so that the section moment of inertia of the stiffener before and after the design remains consistent. The section moment of inertia I is calculated using formula (6): s ; 6≤λ s ≤12 (5) In the formula, h s Indicates the overhang width of the stiffening rib plate, t s Indicates the thickness of the stiffener plate; (9.3) Construct the corresponding analysis model of the stiffened steel plate shear wall: Use ABAQUS finite element analysis software to model the structure and conduct performance evaluation on the buckling mode, bearing capacity, stiffness, out-of-plane deformation and energy dissipation capacity of the model.
Citation Information
Patent Citations
Physical-data driven intelligent shear wall building structure design method and device
CN113779675A
Method for constructing personalized artificial intelligence assistant for shear wall layout design
CN116861540A
High-rise shear wall residential building information reconstruction method for structural seismic capacity evaluation
CN118171362A
Design method for topological optimization of stiffening steel plate shear wall based on generative adversarial network
CN118428203A
Physics-enhanced data-driven method and device for intelligent structural design of shear wall building
US20240211645A1
