Building curved surface curtain wall component deformation prediction method and system based on GAN

Abstract

The invention belongs to the field of structure deformation prediction, and particularly discloses a GAN-based building curved surface curtain wall component deformation prediction method and system, and the method comprises the steps: building a data carrying out the sample collection of deformation related data after the processing of a building curved surface curtain wall component, and forming the database, the collection samples comprising an actual measurement sample and a simulation sample; building a sample generation model: generating new deformed samples after processing of various components based on GAN, and incorporating the generated samples into a database; building a deformation prediction model: building the deformation prediction model after processing of the component based on the CGAN, wherein samples comprise collection samples and generation samples; and deformation prediction: predicting the deformation condition of the processed component by using the model, and judging whether the predicted deformation condition meets the acceptance condition or not. According to the method, by generating the sample enhancement data, the prediction efficiency of deformation of the curtain wall component after processing is effectively improved, so that the rework cost caused by improper deformation control during bending processing of the curtain wall component is reduced.

Description

Technical Field

[0001] The present invention belongs to the field of structural deformation prediction, and more specifically, relates to a method and system for predicting deformation of architectural curved curtain wall components based on GAN. Background Art

[0002] As people's demands for building quality and functionality increase, building forms are becoming more diverse and nonlinear. The curtain wall system is a concrete manifestation of the building's appearance, including panels, keels, decorative lines and other components. In order to meet the nonlinear appearance requirements of different building forms, curtain wall components need to be bent by rolling, stretching, stamping and other forms. However, since curtain wall components often have high slenderness and large size, they are prone to rebound deformation and other defects during the bending process, which in turn affects the quality of the building's actual appearance and causes waste of resources due to rework. Therefore, it is particularly important to predict and control the deformation of curtain wall components after processing.

[0003] At present, the bending forming process of various complex components belongs to complex nonlinear deformation, which can be predicted by artificial intelligence algorithms. Scholars in the fields of shipbuilding and automobile manufacturing mostly use algorithms such as back propagation and support vector machines to predict the rebound deformation of simple components such as single curved circular tubes and stamping plates, or determine the deformation of components through finite element simulation. However, due to the large size of curtain wall components, complex and changeable shapes, high slenderness ratio, twisting, deadweight, material properties and other factors will affect the deformation of components after processing. Moreover, unlike the mass production of shipbuilding and automobile manufacturing, there may be thousands of single components with different shapes and cross-sections in a curved curtain wall project. The existing deformation prediction methods for mass production and single scenarios cannot meet the deformation prediction needs of complex and changeable building curtain wall components. At the same time, the existing finite element simulation methods require a lot of resources for modeling, which cannot adapt to the large number and multiple types of building curtain wall components. Therefore, in the field of curtain wall component processing quality control, it is very necessary to provide a new deformation prediction method and system. Summary of the invention

[0004] In view of the above defects or improvement needs of the prior art, the present invention provides a method and system for predicting the deformation of architectural curved curtain wall components based on GAN, which aims to achieve accurate prediction of the deformation of architectural curved curtain wall components after bending.

[0005] To achieve the above object, according to one aspect of the present invention, a method for predicting deformation of architectural curved curtain wall components based on GAN is proposed, comprising the following steps:

[0006] Model training phase:

[0007] Based on the initial database, the GAN is trained, and the number of samples in the database is expanded through the trained GAN; the initial database includes several samples, each of which includes the material parameters M, process parameters A, geometric parameters Ge, bending parameters B and deformation parameters De of the building curved curtain wall components;

[0008] Based on the expanded database, CGAN is trained, including:

[0009] CGAN includes a generator G2 and a discriminator D2. The generator G2 outputs a generated sample G2 (C, Z2) according to a condition C and a noise vector Z2 that obeys a Gaussian distribution. The condition C includes a material parameter M, a process parameter A, a geometric parameter Ge, and a bending parameter B. The discriminator D2 scores the authenticity of the real samples (C, De) and the generated samples G2 (C, Z2) in the database. During training, the generator G2 aims to make the generated samples G2 (C, Z2) more similar to the real samples (C, De), and the discriminator D1 aims to better distinguish the real samples (C, De) from the generated samples G2 (C, Z2).

[0010] The generator G2 in the trained CGAN is used as the deformation prediction model;

[0011] Model application phase:

[0012] The material parameters M, process parameters A, geometric parameters Ge, and bending parameters B of the building curved curtain wall components to be predicted are input into the deformation prediction model to obtain the corresponding deformation parameters De, thereby realizing the deformation prediction of the building curved curtain wall components.

[0013] As a further preferred method, based on the initial database, the GAN is trained, and the number of samples in the database is expanded by the trained GAN, including:

[0014] GAN includes a generator G1 and a discriminator D1. The generator G1 outputs a generated sample G1(Z1) based on a noise vector Z1 that follows a Gaussian distribution. The discriminator D1 scores the authenticity of the real sample X in the initial database and the generated sample G1(Z1). During training, the generator G1 aims to make the generated sample G1(Z1) more similar to the real sample X, and the discriminator D1 aims to better distinguish the real sample X from the generated sample G1(Z1).

[0015] The trained generator G1 is used as the sample generation model, and multiple samples are generated by the sample generation model and added to the initial database, thereby expanding the database.

[0016] As a further preferred method, when training GAN:

[0017] Loss function of discriminator D1 as follows:

[0018]

[0019] The loss function of generator G1 as follows:

[0020]

[0021] Wherein, the superscript i represents the corresponding data of the i-th sample in the training set, and m represents the total number of training set samples randomly selected after the initial database sample mixture generator generates data samples; D1(X (i) )、D1(G1(Z1 (i) )) represent the scores of the discriminator D1 on the input X and G1(Z1) respectively.

[0022] As a further preferred embodiment, when the generator G1 and the discriminator D1 satisfy the expected formula, the training is terminated to obtain a trained GAN; the expected formula is as follows:

[0023]

[0024] in, It means that the goal of the generator G1 is to minimize the above formula, and the goal of the discriminator D1 is to maximize the above formula; It means that for the real data X, the discriminator D1 gives the probability that it is the real data, It means that for the false data G1(Z) generated by the generator G1, the discriminator D1 gives the probability that it is false data.

[0025] As a further preferred method, when training CGAN:

[0026] Loss function of discriminator D2 as follows:

[0027]

[0028] Among them, the superscript i represents the corresponding data of the i-th sample in the training set, and n represents the total number of training set samples randomly selected from the expanded database; Represents the same as C (i) Any De that does not match (i) ; D2(C (i) ,De (i) )、D2(C (i) ,G2(C (i) ,Z2 (i) ))、 They represent the discriminator D2 for input (C, De), G2 (C, Z2), Rating of θ d represents the discriminator D2 parameter, η represents the learning rate, Represents the d Perform gradient update;

[0029] The loss function of generator G2 as follows:

[0030]

[0031] Among them, θ g represents the generator G2 parameter, η represents the learning rate, Represents the g Perform gradient update.

[0032] As a further preferred embodiment, when the scores given by the discriminator D2 to the real samples (C, De) and the generated samples G2 (C, Z2) in the database are consistent, the training is terminated to obtain a trained CGAN.

[0033] As a further preference, the material parameter M refers to the elastic modulus and yield strength of the component material; the process parameter A refers to the tensile bending forming force of the component; the geometric parameter Ge refers to the cross-sectional size, cavity area ratio and material thickness of the component; the bending parameter B refers to the bending angle, bending radius, arc length and chord length parameters of the bending line before the component is deformed; the deformation parameter De refers to the absolute value of the chord length difference of the bending line before and after the component is deformed.

[0034] As a further preferred method, the curved connection line is determined by: determining all endpoints of the component, and for any two endpoints, making a shortest distance connection line for the two endpoints on the component surface; and selecting a curved connection line from all the connection lines as the curved connection line.

[0035] As a further preference, the samples in the initial database are obtained through bending tests and / or simulation using Abaqus finite element analysis software.

[0036] According to another aspect of the present invention, a GAN-based building curved curtain wall component deformation prediction system is provided, comprising a processor, wherein the processor is used to execute the GAN-based building curved curtain wall component deformation prediction method.

[0037] In general, the above technical solution conceived by the present invention has the following technical advantages compared with the prior art:

[0038] The present invention uses GAN (generative adversarial network) to generate deformation samples of curved curtain wall components, and uses CGAN (conditional generative adversarial network) to build a deformation prediction model for curved curtain wall components after expanding the database. By generating sample enhancement data, the prediction efficiency and accuracy of the deformation of curtain wall components after processing can be effectively improved, so as to improve the one-time qualified quality of processing and reduce the rework cost caused by improper deformation control during the bending processing of curtain wall components. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A schematic diagram of the bending connection line and chord length of the aluminum profile according to an embodiment of the present invention;

[0040] Figure 2 A flowchart for building an initial database for an embodiment of the present invention;

[0041] Figure 3 Generate a model flow chart for the sample of the embodiment of the present invention;

[0042] Figure 4 This is a flow chart of a deformation prediction model according to an embodiment of the present invention;

[0043] Figure 5 This is a flow chart of deformation prediction of building curved curtain wall components according to an embodiment of the present invention;

[0044] Figure 6 This is a flow chart of deformation prediction of curved curtain wall components of buildings based on GAN in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0046] The embodiment of the present invention provides a method for predicting deformation of architectural curved curtain wall components based on GAN. Figure 6 As shown, the following steps are included:

[0047] S1. Establish an initial database: Samples of deformation-related data of building curved curtain wall components before and after processing are collected to form an initial database. Each sample includes material parameters M, process parameters A, geometric parameters Ge, bending parameters B and deformation parameters De of building curved curtain wall components. The samples include measured samples and / or simulated samples, wherein the simulated samples are the deformation data of the curved components after processing obtained by Abaqus finite element analysis software, and the measured samples are the deformation data after processing obtained by on-site measurement after on-site bending tests.

[0048] like Figure 2 As shown, the details are as follows:

[0049] S11. Obtain basic information of the component, including component material parameters M, process parameters A and geometric parameters Ge. The material parameters M refer to parameters related to material plastic deformation, such as elastic modulus and yield strength; the process parameters A refer to the bending forming force, i.e., pre-force, forming force, post-compensation force and other parameters related to component processing and forming; the geometric parameters Ge refer to parameters related to the component geometry, such as component cross-sectional dimensions, cavity area ratio, material thickness and so on.

[0050] S12. Determine the endpoint bending line. There are a number of endpoints on a component. Draw the shortest distance line between any two endpoints on the surface where the component is located. b is the number of straight lines connecting the endpoints of the component. Obtain the bending parameter B of each bending line, which refers to the bending angle, bending radius, arc length, chord length and other parameters related to the bending properties in the horizontal or vertical direction. Figure 1 The aluminum profile component shown in the figure has 14 endpoints. If you take any two endpoints and connect them by the shortest distance on the surface of the component, you can get Root bending line, where The number of straight lines connecting the endpoints of the component.

[0051] S13, simulation and test. Simulate or test the components to obtain the deformation after bending. The simulation is to establish an orthogonal experimental table after Latin hypercube sampling of common material parameters M, process parameters A, and geometric parameters Ge of curtain wall components, which is completed by Abaqus finite element analysis software, and the test is completed by on-site bending. The deformation parameter De is obtained according to the deformation of the component, and De refers to the absolute value of the difference in the chord length of the bending line before and after the deformation of the component.

[0052] S14, data preprocessing. Match and clean the data, match the material parameter M, process parameter A, geometric parameter Ge, bending parameter B, and deformation parameter De corresponding to each bending line, and remove redundant, repeated, inconsistent, missing and other invalid data.

[0053] S15, establishing an initial database. Giving corresponding labels to the data in the database for easy calling, forming an initial data sample, and establishing an initial database.

[0054] S2. Based on the initial database, GAN (generative adversarial network) is trained to obtain a sample generation model, and samples are generated through the sample generation model to expand the number of samples in the database.

[0055] like Figure 3 As shown, the details are as follows:

[0056] S21. Data normalization. Since the dimensions of each parameter vary greatly, all data need to be normalized for comparison and analysis, that is,

[0057]

[0058] where x max and x min represent the maximum and minimum values ​​of parameter x respectively, Represents the result after normalization operation.

[0059] S22. Divide the training set and the test set. Randomly select samples from the initial database in a ratio of 8:2 to generate the training set and the test set of the model.

[0060] S23. Create generator G1. Introduce a noise vector Z1 that obeys Gaussian distribution as the input of generator G1, and generator G1 generates a data sample G1(Z1). The sample should contain material parameters M, process parameters A, geometric parameters Ge, bending parameters B, deformation parameters De and their corresponding labels, and there should also be correlation between the data.

[0061] S24. Create discriminator D1. The real samples X and G1(Z1) obtained from the experiment and simulation are used as the input of D1. Discriminator D1 makes sample quality judgments respectively and outputs judgment results D1(X) and D1(G1(Z1)). Discriminator D1 is trained to make the real sample X get a high score and generate sample G1.

[0062] (Z1) gets a low score. The training formula is:

[0063]

[0064] S25, train the generator G1. The goal is to train the generator G1 with the generated sample G1(Z1) obtaining a high score in the discriminator D1. The training formula is:

[0065]

[0066] Wherein, the superscript i represents the corresponding data of the i-th sample in the training set, and m represents the total number of training set samples randomly selected after the initial database sample mixture generator generates data samples; D1(X (i) )、D1(G1(Z1 (i) )) represent the scores of the discriminator D1 on the input X and G1(Z1) respectively.

[0067] S26, loop S24 and S25, when the generator G1 and the discriminator D1 satisfy the following expected formula, the training is terminated and the trained generator G1 is used as the sample generation model.

[0068]

[0069] in, is the objective function, where the goal of the generator G1 is to minimize the entire formula, making the generated data as close to the real data as possible, and let D1 judge the generated data as real data. The goal of the discriminator D1 is to maximize the entire formula and improve the ability to distinguish between real and generated data. It means that for the real data X, the discriminator gives the probability that it is the real data, It means that for the fake data G1(Z) generated by the generator, the discriminator gives the probability that it is fake data.

[0070] S27. Use the test set to test the accuracy of the sample generation model. Compare the generated samples based on the test set with the test set, and calculate the values ​​of the evaluation indicators RMSE and MAE. After verification, the samples generated by the sample generation model can be included in the database and obtain corresponding labels, thereby expanding the samples in the database.

[0071] S3. Based on the expanded database, CGAN (conditional generative adversarial network) is trained to obtain a deformation prediction model.

[0072] like Figure 4 As shown, the details are as follows:

[0073] S31. Data normalization Since the dimensions of each parameter vary greatly, all data need to be normalized for comparison and analysis, that is,

[0074]

[0075] where x max and x min represent the maximum and minimum values ​​of parameter x respectively, Represents the result after normalization operation.

[0076] S32, dividing the training set and the test set. Randomly extract samples in the database at a ratio of 8:2 to generate the training set and the test set of the model.

[0077] S33, create generator G2. List material parameter M, process parameter A, geometric parameter Ge, and bending parameter B as condition C, introduce noise vector Z2 that obeys Gaussian distribution, use C and Z2 as input of generator G2 at the same time, and generate data G2 (C, Z2) under the constraint of condition C, which is the predicted deformation parameter.

[0078] S34, create discriminator D2. Take (C, De) in the database and data G2(C, Z2) generated by G2 as input of D2, and discriminator D2 makes data quality judgments respectively, and outputs judgment results D2(C, De) and D2(G2(C, Z2)). Train discriminator D2 to make (C, De) get high scores, generate data G2(C, Z2) get low scores, and data with mismatched C and De sampling also get low scores. The training formula is:

[0079]

[0080] Among them, the superscript i represents the corresponding data of the i-th sample in the training set, and n represents the total number of training set samples randomly selected from the expanded database; Represents the same as C (i) Any De that does not match (i) ; D2(C (i) ,De (i) )、D2(C (i) ,G2(C (i) ,Z2 (i) ))、 They represent the discriminator D2 for input (C, De), G2 (C, Z2), Rating of θ d represents the discriminator D2 parameter, η represents the learning rate, Represents the d Perform gradient update. S35, train generator G2. With the goal of generating data G2 (C, Z2) to obtain a high score in discriminator D2, train generator G2, and the training formula is:

[0081]

[0082] Among them, θ g represents the generator G2 parameter, η represents the learning rate, Represents the g Perform gradient update.

[0083] S36, loop S34 and S35 until the deformation parameters predicted by generator G2 are consistent with the score De of the deformation parameters in the database, then stop training, and the trained generator G2 can be regarded as a deformation prediction model with reliable accuracy.

[0084] S37, using the test set to test the accuracy of the deformation prediction model. Compare the deformation prediction based on the test set with the deformation of the test set, and calculate the values ​​of the evaluation indicators RMSE and MAE.

[0085] S4. Deformation prediction: Input new parameters of the component to be predicted and the allowable error value of the curtain wall component in the basic parameter input module. The deformation prediction module predicts the deformation of the component after processing. The comparison and analysis module determines whether the deformation of the component after processing is within the allowable error range.

[0086] like Figure 5 As shown, the details are as follows:

[0087] S41: Input new component material parameters M, process parameters A, geometric parameters Ge, bending parameters B, and curtain wall component error allowable values ​​through the basic parameter input module.

[0088] S42: The deformation of the curtain wall component after processing is predicted through the deformation prediction module to obtain the deformation parameter De and realize the deformation value prediction.

[0089] S43: Compare the predicted deformation value with the allowable error value of the curtain wall component through the comparison analysis module to evaluate whether the component size deformation is within the error range.

[0090] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A GAN-based method for predicting deformation of building curved curtain wall components, characterized in that: The steps include: Model training phase: Based on the initial database, the GAN is trained, and the number of samples in the database is expanded through the trained GAN; the initial database includes several samples, each of which includes the material parameters M, process parameters A, geometric parameters Ge, bending parameters B and deformation parameters De of the building curved curtain wall components; Based on the expanded database, CGAN is trained, including: CGAN includes a generator G2 and a discriminator D2. The generator G2 outputs a generated sample G2 (C, Z2) according to a condition C and a noise vector Z2 that obeys a Gaussian distribution. The condition C includes a material parameter M, a process parameter A, a geometric parameter Ge, and a bending parameter B. The discriminator D2 scores the authenticity of the real samples (C, De) and the generated samples G2 (C, Z2) in the database. During training, the generator G2 aims to make the generated samples G2 (C, Z2) more similar to the real samples (C, De), and the discriminator D1 aims to better distinguish the real samples (C, De) from the generated samples G2 (C, Z2). The generator G2 in the trained CGAN is used as the deformation prediction model; Model application phase: The material parameters M, process parameters A, geometric parameters Ge, and bending parameters B of the building curved curtain wall components to be predicted are input into the deformation prediction model to obtain the corresponding deformation parameters De, thereby realizing the deformation prediction of the building curved curtain wall components.

2. The method for predicting deformation of curved curtain wall components based on GAN according to claim 1, characterized in that: Based on the initial database, GAN is trained and the number of samples in the database is expanded through the trained GAN, including: GAN includes a generator G1 and a discriminator D1. The generator G1 outputs a generated sample G1(Z1) based on a noise vector Z1 that follows a Gaussian distribution. The discriminator D1 scores the authenticity of the real sample X in the initial database and the generated sample G1(Z1). During training, the generator G1 aims to make the generated sample G1(Z1) more similar to the real sample X, and the discriminator D1 aims to better distinguish the real sample X from the generated sample G1(Z1). The trained generator G1 is used as the sample generation model, and multiple samples are generated by the sample generation model and added to the initial database, thereby expanding the database.

3. The method for predicting deformation of curved curtain wall components based on GAN as claimed in claim 2, characterized in that: When training a GAN: Loss function of discriminator D1 as follows: The loss function of generator G1 as follows: Wherein, the superscript i represents the corresponding data of the i-th sample in the training set, and m represents the total number of training set samples randomly selected after the initial database sample mixture generator generates data samples; D1(X (i) )、D1(G1(Z1 (i) )) represent the scores of the discriminator D1 on the input X and G1(Z1) respectively.

4. The method for predicting deformation of curved curtain wall components based on GAN as claimed in claim 3, characterized in that: When the generator G1 and the discriminator D1 satisfy the expected formula, the training is terminated and the trained GAN is obtained; the expected formula is as follows: in, It means that the goal of the generator G1 is to minimize the above formula, and the goal of the discriminator D1 is to maximize the above formula; It means that for the real data X, the discriminator D1 gives the probability that it is the real data, It means that for the false data G1(Z) generated by the generator G1, the discriminator D1 gives the probability that it is false data.

5. The method for predicting deformation of curved curtain wall components based on GAN according to claim 1, characterized in that: When training CGAN: Loss function of discriminator D2 as follows: Among them, the superscript i represents the corresponding data of the i-th sample in the training set, and n represents the total number of training set samples randomly selected from the expanded database; Represents the same as C (i) Any De that does not match (i) ; D2(C (i) ,De (i) )、D2(C (i) ,G2(C (i) ,Z2 (i) ))、 They represent the discriminator D2 for input (C, De), G2 (C, Z2), Rating of θ d represents the discriminator D2 parameter, η represents the learning rate, Represents the d Perform gradient update; The loss function of generator G2 as follows: Among them, θ g represents the generator G2 parameter, η represents the learning rate, Represents the g Perform gradient update.

6. The method for predicting deformation of curved curtain wall components based on GAN according to claim 5, characterized in that: When the scores of the discriminator D2 for the real samples (C, De) in the database and the generated samples G2 (C, Z2) are consistent, the training is terminated and the trained CGAN is obtained.

7. The method for predicting deformation of curved curtain wall components based on GAN according to claim 1, characterized in that: Material parameter M refers to the elastic modulus and yield strength of the component material; process parameter A refers to the tensile bending force of the component; geometric parameter Ge refers to the cross-sectional size, cavity area ratio and material thickness of the component; The bending parameter B refers to the bending angle, bending radius, arc length and chord length parameters of the bending line before the component is deformed; The deformation parameter De refers to the absolute value of the difference in chord length of the bending line before and after the component is deformed.

8. The method for predicting deformation of curved curtain wall components based on GAN according to claim 7, characterized in that: The method for determining the curved connection line is: determine all the endpoints of the component, and for any two endpoints, make a shortest distance connection line for the two endpoints on the component surface; and select a curved connection line from all the connections as the curved connection line.

9. The method for predicting deformation of architectural curved curtain wall components based on GAN according to any one of claims 1 to 8, characterized in that: The samples in the initial database are obtained through bending tests and / or Abaqus finite element analysis software simulations.

10. A GAN-based building curved curtain wall component deformation prediction system, characterized in that: It includes a processor, which is used to execute the GAN-based building curved curtain wall component deformation prediction method as described in any one of claims 1 to 9.

Images