A deep learning method for springback prediction of hyperbolic plates based on constitutive relations

Abstract

The present invention discloses a deep learning method for predicting the springback of a hyperbolic plate based on a constitutive relationship. The method comprises the following steps: collecting sample springback data of hyperbolic plates from plate specimens formed from various materials, aggregating the sample springback data from the various materials to form a sample dataset; constructing a deep convolutional neural network; adding a custom constitutive relationship layer to the deep convolutional neural network; training the deep convolutional neural network based on the training dataset to obtain a plate forming springback prediction model; inputting preprocessed springback prediction input samples into the plate forming springback prediction model, selecting stress-strain curves corresponding to the springback prediction input samples, adjusting the weights and bias parameters of the hyperbolic plate springback prediction model based on the stress-strain curves of the materials, and outputting predicted springback data. The present invention can effectively overcome the problem of existing algorithms being unable to consider material properties. By embedding a large number of material constitutive models, it can achieve springback prediction for plates made of various new materials and composite materials.

Description

Technical Field

[0001] The present invention belongs to the technical field of hyperbolic plate forming springback prediction, and specifically relates to a deep learning method for hyperbolic plate springback prediction based on constitutive relations. Background Art

[0002] With the advancement of technology, the requirements for sheet metal forming in various industrial fields are also increasing. At the same time, different structures have different requirements, which can be achieved through different materials. However, the material properties of many composite materials and other advanced materials currently used are relatively complex. Therefore, it is difficult to improve the accuracy of sheet metal forming when selecting materials to meet other requirements.

[0003] At present, the application of deep learning in the field of sheet metal forming for springback prediction has achieved good results. However, due to the complex constitutive relationships of many composite materials, when the three-dimensional forming feature information extracted from these different materials is used for springback prediction, the prediction accuracy is reduced, which seriously affects the forming accuracy of the sheet metal. Summary of the Invention

[0004] The purpose of the present invention is to provide a deep learning method for predicting the rebound of hyperbolic plates based on constitutive relations, which can effectively overcome the problem that existing algorithms cannot take material properties into account. By embedding a large number of material constitutive models, the rebound prediction of various new materials and composite materials can be realized.

[0005] To solve the above technical problems, the technical solution of the present invention is: a deep learning method for predicting the rebound of a hyperbolic plate based on constitutive relations, comprising the following steps:

[0006] S1. Conduct tensile tests on sheet metal specimens formed from various materials to obtain tensile force-displacement curves of the corresponding materials. Obtain stress-strain curves of the corresponding materials based on the tensile force-displacement curves. Collect springback values ​​of the sheet metal specimens formed from various materials during forming experiments and numerical simulations. Use the stress-strain curves and springback values ​​of the various materials as components of sample data. Aggregate the sample data from the various materials to form a sample dataset; the sample data include at least springback values, stress-strain curves, sheet metal thickness, and forming variation of the various materials. Preprocess the sample dataset to obtain a training dataset.

[0007] S2. Based on convolutional neural networks, a deep convolutional neural network is constructed for end-to-end springback prediction of sheet metal forming of multiple materials;

[0008] S3. Add a custom constitutive relationship layer to the deep convolutional neural network to map the constitutive relationship of the material. Train the deep convolutional neural network based on the material constitutive relationship according to the training data set to obtain a sheet metal forming springback prediction model.

[0009] S4. Input the preprocessed springback prediction input sample into the sheet metal forming springback prediction model, select the stress-strain curve corresponding to the springback prediction input sample, adjust the weight and bias parameters of the hyperbolic plate springback prediction model according to the stress-strain curve of the material, and output the predicted springback data.

[0010] The specific method of preprocessing in S1 is:

[0011] The hyperbolic plate plane is set as the xoy plane, the direction perpendicular to the plate surface is the z direction, and the xoy plane is discretized into an n×n grayscale image with a grayscale value range of [0, n-1]. The maximum value of the plate thickness of the corresponding material corresponds to n-1, and the minimum value corresponds to 0. The plate thickness of the corresponding material is scaled and the plate thickness data is written to the thickness file;

[0012] In RGB color mode, the maximum value of the forming variation and springback of the corresponding material corresponds to n-1, and the minimum value corresponds to 0. The forming variation of the corresponding material is scaled and the forming variation data and springback data of the corresponding material are written into the input file and output file respectively; wherein, the thickness file, input file and output file are all stored in the same preset format;

[0013] The thickness file is converted into a grayscale image, and the input file and output file are converted into RGB images as the input and output data sets; the input data set and the output data set are randomly divided into a training set, a validation set, and a test set according to a preset ratio, and used as the training data set; the input data includes the plate thickness and the forming change, and the output data includes the rebound amount.

[0014] The structure of the deep convolutional neural network in S2 is as follows: convolution layer, activation layer, maximum pooling layer, and transposed convolution layer; among them, the convolution layer is used for linear product summation, the activation layer is a rectified linear activation function, the maximum pooling layer is used for nonlinear downsampling, and the transposed convolution layer is used for upsampling.

[0015] S3 is specifically as follows: material tags are added to the RGB images in the training data set. After the deep convolutional neural network completes the first pooling, a custom constitutive relationship layer is added. In the custom constitutive relationship layer, the stress-strain curve of the corresponding material is identified by the material tag and element-by-element multiplication is performed. The output is passed to the subsequent convolutional layer and pooling layer. After each round of training based on the training data set, the weight of the custom constitutive relationship layer is updated according to the training accuracy feedback. After the preset rounds of training, the optimal weight of the custom constitutive relationship layer and the bias parameters of the sheet metal forming springback prediction model are determined to obtain the sheet metal forming springback prediction model.

[0016] In S4, the preprocessed springback prediction input sample is input, and the corresponding stress-strain curve is selected according to its material. The weight and bias parameters of the sheet metal forming springback prediction model are adjusted according to the stress-strain curve of the material. The sheet metal forming springback prediction model outputs an RGB image with a pixel size of n×n to represent the springback amount of the springback prediction input sample. The RGB image is reverse mapped according to the preprocessing method, and the RGB value of the RGB image is converted into the actual springback value and stored in a preset format, and it is used as the predicted springback data.

[0017] n is 256.

[0018] The preset ratio of the training set, validation set, and test set is 8:1:1.

[0019] The deep convolutional neural network is set to 42 layers.

[0020] The thickness file, input file, output file and actual springback value are all stored in .xlsx file format.

[0021] A computer-readable storage medium is also provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of any of the above methods are implemented.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] This paper applies deep learning to the field of hyperbolic plate forming by embedding a deep convolutional neural network algorithm within material curves. This approach is suitable for predicting the processing of a wider range of composite and advanced material ship hull plates. Unlike existing deep learning methods, this method can simultaneously predict the springback of multiple materials, including more complex materials. Using a deep convolutional neural network algorithm, the springback prediction of the plate can take into account the influence of the material's constitutive relationship on forming and springback. By matching the plate's shape input features with the plate's material features, the shape features of different materials are element-wise multiplied with the corresponding material features, incorporating the material features into subsequent training. This method fully considers the influence of the complex constitutive relationships of composite and advanced materials on the processing and forming of ship hull plates. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Schematic diagram of a flow chart of an embodiment of the present invention;

[0025] Figure 2 Schematic diagram of the overall framework of a deep convolutional neural network in an embodiment of the present invention;

[0026] Figure 3 This is a schematic diagram of the constitutive relationship customization process in an embodiment of the present invention;

[0027] Figure 4 Schematic diagram of the workflow of the sheet metal forming springback prediction model in an embodiment of the present invention;

[0028] Figure 5 Schematic diagram of the shape data processing flow in an embodiment of the present invention;

[0029] Figure 6 Schematic diagram of the matching process between the springback prediction input sample and the corresponding stress-strain curve in an embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining 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 may be combined with each other as long as they do not conflict with each other.

[0031] See Figures 1-6 , the technical solution of the present invention is:

[0032] A deep learning method for predicting springback of a hyperbolic plate based on constitutive relations includes the following steps:

[0033] S1. Conducting tensile tests on sheet metal specimens formed from various materials to obtain force-displacement curves of the corresponding materials. Based on the force-displacement curves, obtaining stress-strain curves of the corresponding materials. Then, collecting springback values ​​of the sheet metal specimens formed from various materials during sheet metal forming tests and numerical simulations. The stress-strain curves and springback values ​​of the various materials are used as components of sample data. The sample data of the various materials are aggregated to form a sample dataset; the sample data include at least springback values, stress-strain curves, sheet metal thickness, and forming variation of the various material sheets. Preprocessing the sample dataset to obtain a training dataset.

[0034] S2. Based on convolutional neural networks, a deep convolutional neural network is constructed for end-to-end springback prediction of sheet metal forming of multiple materials;

[0035] S3. Add a custom constitutive relationship layer to the deep convolutional neural network to map the constitutive relationship of the material. Train the deep convolutional neural network based on the material constitutive relationship according to the training data set to obtain a sheet metal forming springback prediction model.

[0036] S4. Input the preprocessed springback prediction input sample into the sheet metal forming springback prediction model, select the stress-strain curve corresponding to the springback prediction input sample, adjust the weight and bias parameters of the hyperbolic plate springback prediction model according to the stress-strain curve of the material, and output the predicted springback data.

[0037] The specific method of preprocessing in S1 is:

[0038] The hyperbolic plate plane is set as the xoy plane, and the direction perpendicular to the plate surface is the z direction. The xoy plane is discretized into a 256×256 grayscale image with a grayscale value range of [0,255]. The maximum value of the plate thickness of the corresponding material corresponds to 255, and the minimum value corresponds to 0. The plate thickness of the corresponding material is scaled and the plate thickness data is written to the thickness file;

[0039] In RGB color mode, the maximum value of the forming variation and springback of the corresponding material corresponds to 255, and the minimum value corresponds to 0. The forming variation of the corresponding material is scaled and the forming variation data and springback data of the corresponding material are written into the input file and output file respectively; wherein, the thickness file, input file and output file are all stored in the same preset format;

[0040] The thickness file is converted into a grayscale image, and the input file and output file are converted into RGB images as the input and output data sets; the input data set and the output data set are randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1, and used as the training data set; the input data includes the plate thickness and the forming change, and the output data includes the rebound amount.

[0041] The deep convolutional neural network structure in S2 includes convolutional layers, activation layers, max pooling layers, and transposed convolutional layers. The convolutional layers perform linear product summation, the activation layers use rectified linear activation functions, the max pooling layers perform nonlinear downsampling, and the transposed convolutional layers perform upsampling. The deep convolutional neural network has 42 layers.

[0042] S3 is specifically as follows: material tags are added to the RGB images in the training data set. After the deep convolutional neural network completes the first pooling, a custom constitutive relationship layer is added. In the custom constitutive relationship layer, the stress-strain curve of the corresponding material is identified by the material tag and element-by-element multiplication is performed. The output is passed to the subsequent convolutional layer and pooling layer. After each round of training based on the training data set, the weight of the custom constitutive relationship layer is updated according to the training accuracy feedback. After the preset rounds of training, the optimal weight of the custom constitutive relationship layer and the bias parameters of the sheet metal forming springback prediction model are determined to obtain the sheet metal forming springback prediction model.

[0043] In S4, the preprocessed springback prediction input sample is input, and the corresponding stress-strain curve is selected according to its material. The weight and bias parameters of the sheet metal forming springback prediction model are adjusted according to the stress-strain curve of the material. The weight and bias parameters of the springback prediction model change with the stress-strain curves of different materials. The sheet metal forming springback prediction model outputs an RGB image with a pixel size of 256×256 to represent the springback amount of the springback prediction input sample. The RGB image is reverse mapped according to the preprocessing method, and the RGB value of the RGB image is converted into the actual springback value and stored in the .xlsx file format, and it is used as the predicted springback data.

[0044] It will be easily understood by those skilled in the art that the above description is merely 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 scope of protection of the present invention.

Claims

1. A deep learning method for predicting the rebound of a hyperbolic plate based on constitutive relations, characterized by: The following steps are involved: S1. Conduct tensile tests on sheet metal specimens formed from various materials to obtain tensile force-displacement curves of the corresponding materials. Determine stress-strain curves of the corresponding materials based on the tensile force-displacement curves. Collect springback values ​​of the sheet metal specimens formed from various materials during forming experiments and numerical simulations. Use the stress-strain curves and springback values ​​of the various materials as components of sample data. Aggregate the sample data for the various materials to form a sample dataset. The sample data at least includes the springback amount, stress-strain curve, plate thickness and forming variation of various material plates; the sample data set is preprocessed to obtain a training data set; S2. Based on convolutional neural networks, a deep convolutional neural network is constructed for end-to-end springback prediction of sheet metal forming of multiple materials; S3. Add a constitutive relationship custom layer to the deep convolutional neural network to map the constitutive relationship of the material, train the deep convolutional neural network based on the material constitutive relationship according to the training data set, and obtain a sheet metal forming springback prediction model; S3 specifically includes: adding material tags to the RGB images in the training data set, adding a constitutive relationship custom layer after the deep convolutional neural network completes the first pooling, identifying the stress-strain curve of the corresponding material through the material tag in the constitutive relationship custom layer, performing element-by-element multiplication, and passing the output to the subsequent convolutional layer and pooling layer, and after each round of training according to the training data set, updating the weight of the constitutive relationship custom layer according to the training accuracy feedback, and determining the optimal weight of the constitutive relationship custom layer and the bias parameters of the sheet metal forming springback prediction model after the preset rounds of training to obtain the sheet metal forming springback prediction model; S4. Input the preprocessed springback prediction input sample into the sheet metal forming springback prediction model, select the stress-strain curve corresponding to the springback prediction input sample, adjust the weight and bias parameters of the hyperbolic plate springback prediction model according to the stress-strain curve of the material, and output the predicted springback data.

2. A deep learning method for predicting springback of a hyperbolic plate based on constitutive relations according to claim 1, characterized in that: The specific method of preprocessing in S1 is: The hyperbolic plate plane is set as the xoy plane, the direction perpendicular to the plate surface is the z direction, and the xoy plane is discretized into an n×n grayscale image with a grayscale value range of [0, n-1]. The maximum value of the plate thickness of the corresponding material corresponds to n-1, and the minimum value corresponds to 0. The plate thickness of the corresponding material is scaled and the plate thickness data is written to the thickness file; In RGB color mode, the maximum value of the forming variation and springback of the corresponding material corresponds to n-1, and the minimum value corresponds to 0. The forming variation of the corresponding material is scaled and the forming variation data and springback data of the corresponding material are written into the input file and output file respectively; wherein, the thickness file, input file and output file are all stored in the same preset format; The thickness file is converted into a grayscale image, and the input file and output file are converted into RGB images as the input and output data sets; the input data set and the output data set are randomly divided into a training set, a validation set, and a test set according to a preset ratio, and used as the training data set; the input data includes the plate thickness and the forming change, and the output data includes the rebound amount.

3. A deep learning method for predicting springback of a hyperbolic plate based on constitutive relations according to claim 2, characterized in that: The structure of the deep convolutional neural network in S2 is as follows: convolution layer, activation layer, maximum pooling layer, and transposed convolution layer; among them, the convolution layer is used for linear product summation, the activation layer is a rectified linear activation function, the maximum pooling layer is used for nonlinear downsampling, and the transposed convolution layer is used for upsampling.

4. The deep learning method for predicting springback of a hyperbolic plate based on constitutive relations according to claim 1, characterized in that: In S4, the preprocessed springback prediction input sample is input, and the corresponding stress-strain curve is selected according to its material. The weight and bias parameters of the sheet metal forming springback prediction model are adjusted according to the stress-strain curve of the material. The sheet metal forming springback prediction model outputs an RGB image with a pixel size of n×n to represent the springback amount of the springback prediction input sample. The RGB image is reverse mapped according to the preprocessing method, and the RGB value of the RGB image is converted into the actual springback value and stored in a preset format, and it is used as the predicted springback data.

5. A deep learning method for predicting springback of a hyperbolic plate based on constitutive relations according to claim 4, characterized in that: n is 256.

6. A deep learning method for predicting springback of a hyperbolic plate based on constitutive relations according to claim 2, characterized in that: The preset ratio of the training set, validation set, and test set is 8:1:

1.

7. The deep learning method for predicting springback of a hyperbolic plate based on constitutive relations according to claim 3 is characterized in that: The deep convolutional neural network is set to 42 layers.

8. The deep learning method for predicting springback of a hyperbolic plate based on constitutive relations according to claim 4 is characterized in that: The thickness file, input file, output file and actual springback value are all stored in .xlsx file format.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images