Metal forming limit curve prediction method based on neural network

Abstract

The invention belongs to the technical field of simulation application, and particularly relates to a metal forming limit curve prediction method based on a neural network, and the method comprises the steps: obtaining a basic performance capable of representing single-pull performance and a forming limit curve of a material; parameterizing the basic performance and the forming limit curve to obtain a strain parameter and a parameterization curve; obtaining a neural network model, and training the neural network model according to the strain parameters and the parameterized curve; and obtaining input parameters of a to-be-predicted material, and predicting a forming limit curve according to the data parameters and based on the trained neural network model. The data used for training the neural network is wide in source and can be obtained through self experiments according to the mode introduced in the patent, and any existing experimental data can also be used. Predicted results can be used for engineering problems including punch forming process analysis, sheet metal failure prediction and the like, and efficiency and accuracy of related simulation of sheet metal forming are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of simulation applications, and particularly relates to a method for predicting the metal forming limit curve based on a neural network. Background Art

[0002] In current industrial production, stamping parts are widely used in various engineering fields, and the forming limit curve (FLC) of sheet metal is crucial for stamping analysis and part design. Constructing a complete FLC often requires a large number of experiments with specimens of different geometries, which is time-consuming, laborious, and costly. At the same time, it is easily affected by factors such as specimen conditions and measurement methods. In recent decades, various theoretical methods have been established to analyze and evaluate the limiting strain, but the theoretical calculation methods often only apply to certain special cases, have large application limitations, and have high requirements for the use of engineering and technical personnel, making them not easy to popularize.

[0003] Currently, the methods for obtaining the forming limit mainly include forming limit experiments and theoretical calculation predictions. For example, Standard Document 1 (ISO 12004-2) clearly elaborates on experimental methods including Marciniak and Nakajima tests, as well as the methods for extracting corresponding experimental data and drawing the forming limit curve. Another example is Patent Document 2 (CN202310190831), which proposes a method for obtaining the forming limit through simulation calculation using the VonMises yield criterion constitutive in the lsdyna software.

[0004] The standard method in Standard Document 1 can accurately obtain the forming limit curve, but the number of experiments is large, the requirements for experimental accuracy are high, and the calibration process is difficult, which is not friendly to general laboratories and staff. The finite element calculation method described in Patent Document 2 has obvious limitations. VonMises is an isotropic yield criterion and is difficult to adapt to rolled sheet metal or aluminum alloy with strong anisotropy. In addition, the gissmo failure criterion requires obtaining the failure parameters of the material in advance and also requires a considerable amount of preliminary experimental data support. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for predicting the metal forming limit curve based on a neural network, which can be more convenient and accurate than traditional complex forming limit experiments or theoretical calculation methods.

[0006] The technical solution adopted by the present invention is specifically as follows:

[0007] A method for predicting the metal forming limit curve based on a neural network, comprising:

[0008] Obtaining the basic properties characterizing uniaxial tensile performance and the forming limit curve of the material;

[0009] Parametrize the basic performance and the forming limit curve to obtain strain parameters and a parametric curve;

[0010] Obtain a neural network model, and train the neural network model according to the strain parameters and the parametric curve to obtain connection weight coefficients with the minimum error from the target samples;

[0011] Obtain the input parameters of the material to be predicted, and predict the forming limit curve based on the data parameters and the trained neural network model.

[0012] In a preferred solution, the step of obtaining the basic performance that can characterize the uniaxial tensile performance includes:

[0013] Conduct a uniaxial tensile experiment;

[0014] Measure the time history including the deformation of the extensometer and the cross-section load, and calculate the corresponding engineering stress-strain curve;

[0015] Use a non-contact strain measurement device for testing to obtain the surface strain and calculate the anisotropy coefficient;

[0016] The surface strain includes the engineering stress-strain curves of the material in three directions and the anisotropy coefficients of the material in three directions.

[0017] In a preferred solution, the engineering stress-strain curves of the material in three directions are:

[0018]

[0019] l0，l is the extensometer length before and after the experiment, F is the load of the testing machine, A0 is the original cross-sectional area of the sample, combined ε eng , σ eng the engineering stress-strain curve can be obtained;

[0020] The calculation method of the anisotropy coefficient of the material in three directions is as follows:

[0021]

[0022] where w 0, w is the width of the measurement point before and after the experiment, that is, along the vertical direction of the tensile direction, h 0, h is the length of the measurement point before and after the experiment, that is, along the tensile direction, and r is the anisotropy coefficient in this test direction.

[0023] In a preferred solution, the forming limit curve of the material is obtained by conducting a forming limit experiment, and different stress limit strain scatter points need to be obtained in the forming limit experiment.

[0024] In a preferred embodiment, the method for obtaining the limit strain scatter points under different stresses includes:

[0025] Let ε1 , ε2 be the maximum principal and secondary strains at the cracking point measured in the bulging experiment, and n be the number of all measurement points. Then, the forming limit test data can be recorded as:

[0026]

[0027] In a preferred embodiment, the steps of parameterizing the basic properties and the forming limit curve to obtain the strain parameters and the parameterized curve include:

[0028] Perform true stress-strain conversion and extrapolation fitting on the engineering stress-strain curve to obtain the corresponding parametric strain parameters, which are used as the sample inputs of the neural network;

[0029] Perform curve fitting on the limit strain scatter points obtained from the forming limit experiment to obtain a parameterized curve, which is used as the neural network target sample corresponding to the single tension data input.

[0030] In a preferred embodiment, the steps for obtaining the strain parameters are as follows:

[0031] The engineering stress-strain curve is first truncated at the peak position;

[0032] Convert it into a true stress-strain curve through the following formula;

[0033] ε true = ln(1 + ε eng );

[0034] σ true = σ eng *(1 + ε eng );

[0035] Perform fitting on the true stress-strain curve through the extrapolation formula. The fitting method can be, but is not limited to, swift, Ghosh;

[0036] σ eq = k1*(k2 + ε eq ) n + k3;

[0037] The fitting parameters of the above true stress-strain curve and the obtained anisotropy coefficient r value can form the sample input items of the neural network. Assuming the number of samples is m, the sample input matrix of the single tension information is:

[0038]

[0039] Denote x i = [k 1-i , k 2-i , k 3-i , n i , r 00-i , r 45-i , r 90-i ​ , i = 1,2，3，...，m Then the sample input matrix is simply denoted as:

[0040]

[0041] In a preferred embodiment, the neural network model is a three - layer fully - connected neural network. The first layer is the input layer, the middle layer is the hidden layer, and the last layer is the output layer.

[0042] In a preferred embodiment, the steps of training the neural network model according to the strain parameters and the parametric curve include:

[0043] Use the strain parameters as the sample input of the neural network, and the parametric curve as the neural network target sample corresponding to the strain parameters to train the neural network model;

[0044] The actual training method can be but is not limited to batch gradient descent method or stochastic gradient descent method.

[0045] In a preferred embodiment, the steps of obtaining the input parameters of the material to be predicted and predicting the forming limit curve based on the data parameters and the trained neural network model include:

[0046] Take the input parameters of the material to be predicted;

[0047] Substitute the input parameters into the trained neural network model to output the forming limit parameters of the material to be predicted:

[0048] Substitute the forming limit parameters into the forming curve equation to obtain the forming limit curve of the material to be predicted.

[0049] The technical effects achieved by the present invention are as follows: The method for quickly predicting the forming limit based on a simple uniaxial tension experiment and a neural network abandons the traditional complex forming limit experiment or theoretical calculation method. The forming limit curve predicted by the fully - trained neural network has the advantages of high efficiency and high accuracy, and the data source for training the neural network is extensive. It can be obtained by self - experiment according to the method introduced in this patent, or any existing experimental data can be used. The predicted results can be used for engineering problems including stamping process analysis and sheet metal failure prediction, improving the efficiency and accuracy of sheet metal forming - related simulations. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is the system block diagram of the present invention;

[0051] Figure 2 is the flowchart of predicting the forming limit through a neural network in the embodiment of the present invention;

[0052] Figure 3 It is the configuration diagram of neural network model 1 in the embodiments of the present invention;

[0053] Figure 4 It is the process of training the neural network using the batch gradient descent method in the embodiments of the present invention;

[0054] Figure 5 It is the uniaxial tensile test diagram in the embodiments of the present invention;

[0055] Figure 6 It is the uniaxial tensile strain data diagram in the embodiments of the present invention;

[0056] Figure 7 It is the anisotropy coefficient diagram in the embodiments of the present invention;

[0057] Figure 8 It is the bulging test diagram in the embodiments of the present invention;

[0058] Figure 9 It is the scatter plot data diagram of the forming limit strain in the bulging test in the embodiments of the present invention;

[0059] Figure 10 It is the diagram of basic performance fitting and sample input data integration in the embodiments of the present invention;

[0060] Figure 11 It is the diagram of forming limit curve parameter fitting and sample target integration in the embodiments of the present invention;

[0061] Figure 12 It is the basic performance diagram obtained from the new version of the gold single tensile test in the embodiments of the present invention;

[0062] Figure 13 It is the hardening curve parameter fitting diagram in the embodiments of the present invention;

[0063] Figure 14 It is the comparison diagram of the predicted forming limit curve and the experimental forming limit curve of the new sheet metal in the embodiments of the present invention;

[0064] Figure 15 It is the configuration diagram of neural network model 2 in the embodiments of the present invention. Detailed implementation manners

[0065] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will make a detailed description of the specific implementation manners of the present invention in conjunction with the accompanying drawings of the specification.

[0066] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention. However, the present invention may be practiced in other ways than those specifically described herein, and those skilled in the art can make similar extensions without departing from the spirit of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0067] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in a preferred embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments.

[0068] Thirdly, the present invention is described in detail in conjunction with schematic diagrams. When describing the embodiments of the present invention in detail, for the sake of convenience of explanation, the cross-sectional views showing the device structure will be enlarged locally in a non-general proportion, and the schematic diagrams are only examples and should not limit the scope of protection of the present invention herein. In addition, in actual production, three-dimensional spatial dimensions including length, width, and depth should be included.

[0069] Embodiment 1

[0070] Please refer to the Figures 1 - 4 As shown, this is the first embodiment of the present invention. This embodiment provides a method for predicting the metal forming limit curve based on a neural network, including:

[0071] S1. Obtain the basic properties that can characterize the uniaxial tensile performance and the forming limit curve of the material;

[0072] S2. Perform parametric processing on the basic properties and the forming limit curve to obtain strain parameters and a parametric curve;

[0073] S3. Obtain a neural network model, and train the neural network model according to the strain parameters and the parametric curve to obtain the connection weight coefficients with the minimum error from the target samples;

[0074] S4. Obtain the input parameters of the material to be predicted, and predict the forming limit curve based on the data parameters and the trained neural network model.

[0075] In the present invention, first, a neural network capable of characterizing the uniaxial tensile performance parameters and the forming limit curve parameters is established. Secondly, a certain number of sheet metal materials are selected for uniaxial tensile experiments and forming limit experiments to obtain sample points. Finally, the neural network is trained to determine the neural network parameters for predicting the forming limit. The trained neural network only needs to input the experimental data of uniaxial tension to predict the forming limit. Based on the simple uniaxial tensile experiment and the method of quickly predicting the forming limit by the neural network, the traditional complex forming limit experiment or theoretical calculation method is abandoned. The predicted forming limit curve by the fully trained neural network has the advantages of high efficiency and high accuracy, and the data source for training the neural network is extensive, which can be obtained by self-experiment according to the method introduced in this patent, or any existing experimental data can also be used. The predicted results can be used for engineering problems including stamping forming process analysis and sheet metal failure prediction, etc., to improve the efficiency and accuracy of sheet metal forming related simulations.

[0076] Based on a further supplement to the above steps, wherein the step of obtaining the basic performance capable of characterizing the uniaxial tensile performance includes:

[0077] Conduct uniaxial tensile experiments. Generally, for sheet metal, it is recommended to use the ISO6892-1 test method for tensile experiments of metallic materials at room temperature for sample preparation and experiments. For sampling, three types of tensile specimens along the rolling direction (if any), and at 45° and 90° to the rolling direction need to be considered;

[0078] Measure the time history including the extensometer deformation and the cross-sectional load, and calculate the corresponding engineering stress-strain curve;

[0079] Use a non-contact strain measurement device (DIC) or other similar tools or equipment for testing to obtain the surface strain, and calculate the anisotropy coefficient (Landford value);

[0080] Wherein the surface strain includes the engineering stress-strain curves of the material in three directions and the anisotropy coefficients of the material in three directions;

[0081] Wherein, the engineering stress-strain curves of the material in three directions are:

[0082]

[0083] l0，l is the extensometer length before and after the experiment, F is the load of the testing machine, A0 is the original cross-sectional area of the specimen, combined ε eng , σ eng the engineering stress-strain curve can be obtained;

[0084] The anisotropy coefficient of the material in three directions is calculated as follows:

[0085]

[0086] where w 0, w is the width of the measurement point before and after the experiment, that is, along the direction perpendicular to the stretching h 0, h is the length of the measurement point before and after the experiment, that is, along the stretching direction, and r is the anisotropy coefficient in this test direction.

[0087] After the above steps, it is also necessary to carry out the forming limit experiment method and data extraction. The forming limit curve of the material is obtained by carrying out the forming limit experiment. Among them, the limit strain scatter points under different stresses need to be obtained in the forming limit experiment. Generally, for sheet metal, it is recommended to use the ISO 12004-2 sheet metal forming limit test method for sample preparation and experiment. The experiment method and the way to obtain the forming limit scatter points refer to the above specifications and will not be elaborated here. The ways to obtain the limit strain scatter points under different stresses include:

[0088] Let ε1 , ε2 be the major and minor principal strains at the cracking point measured in the bulging experiment, n be the number of all measurement points, then the forming limit test data can be recorded as:

[0089]

[0090] In a preferred embodiment, the steps of parameterizing the basic performance and the forming limit curve to obtain the strain parameter and the parameterized curve include:

[0091] Perform true stress-strain conversion and extrapolation fitting on the engineering stress-strain curve to obtain the corresponding parameter strain parameter as the sample input of the neural network;

[0092] Perform curve fitting on the limit strain scatter points obtained from the forming limit experiment to obtain a parameterized curve as the neural network target sample corresponding to the uniaxial tension data input.

[0093] Based on the above embodiment, the steps to obtain the strain parameter are as follows:

[0094] The engineering stress-strain curve is first truncated at the peak position;

[0095] Convert it into a true stress-strain curve through the following formula;

[0096] ε true = ln(1 + ε eng );

[0097] σ true = σ eng *(1 +ε eng ):

[0098] The true stress-strain curve is fitted by an extrapolation formula, and the fitting method can be, but is not limited to, Swift, Ghosh;

[0099] σ eq = k1 * (k2 + ε eq ) n + k3;

[0100] The anisotropy coefficient r values obtained by attaching the fitting parameters of the above true stress-strain curve can form the sample input items of the neural network. Assuming the number of samples is m, the sample input matrix for uniaxial tension information is:

[0101]

[0102] Denote x i = [k 1-i , k 2-i , k 3-i , n i , r 00-i , r 45-i , r 90-i ​ , i = 1,2,3，...，m , then the sample input matrix is simplified and denoted as:

[0103]

[0104] Furthermore, the past steps of the parametric curve are as follows:

[0105] The measured maximum principal and secondary strains, ε1 , ε2 are converted into the principal-secondary strain ratio α and the equivalent cracking strain ε m .

[0106]

[0107] Then, the experimental principal and secondary strain scatter vectors in step (2) can be converted into the following principal-secondary strain ratio α and equivalent cracking strain ε vm vectors;

[0108]

[0109] Then, the above experimental result vectors can be fitted by a piecewise fitting equation to convert the scatter points into the parameters of a regular forming limit curve. Substituting the above principal-secondary strain ratio α and equivalent cracking strain ε vm vectors into the following forming curve equation, the corresponding curve parameters ε0, ε1, θ1 and θ2 can be obtained by least squares fitting.

[0110]

[0111] Similarly, assuming the number of samples is m, the sample target matrix of the uniaxial tension information is as follows:

[0112]

[0113] Denote y i = [ε 0-i , ε 1-i , θ 1-i , θ 2-i ​ , i = 1，2,3，...，m , then the sample input matrix is simply denoted as:

[0114]

[0115] It should be noted here that using different types of hardening parameter models to describe the uniaxial tension performance and using different types of curve parameter equations and forming limit curves are all within the scope of this patent.

[0116] It should be noted here that the hidden relationship between the corresponding strain parameters as the sample input of the neural network and the output of the forming curve parameters is denoted as:

[0117] [ε 0-i , ε 1-i , θ 1-i , θ 2-i = f([k 1-i , k 2-i , k 3-i , n i , r 00-i , r 45-i , r 90-i );

[0118] Or

[0119] y i = f(x i ) i = 1, 2, 3,..., m;

[0120] To find the relationship between the above input and output, a neural network model with regression function such as a multi-layer perceptron neural network (MLP), a convolutional neural network (CNN), a recurrent neural network (RNN), or other types can be established to optimize and equivalent the above parameter equations. (Neural networks with different structures, numbers of layers, numbers of nodes, or introducing convolutional and pooling layers, and using different types of activation functions are all within the scope of this patent).

[0121] It should be emphasized here that neural networks with different structures, numbers of layers, numbers of nodes, or introducing convolution and using different types of activation functions are all within the scope of this patent.

[0122] Based on the above neural network model, the neural network model also needs to be trained according to the strain parameters and the parametric curve. The specific steps include:

[0123] The loss function of the neural network can use the mean squared error (MSE), and the gradients of the neural network parameters are calculated using the Backpropagation method and the appropriate learning rate is selected to update the neural network , the actual training method can adopt but is not limited to batch gradient descent (BGD) or stochastic gradient descent (SGD) See Figure 2 .

[0124] It should be noted that for the above sample input data, in addition to the above methods, any existing data can also be used, and the larger the number of samples m, the better the effect of the trained neural network.

[0125] Finally, the process of obtaining the forming limit curve of the material to be predicted is described as follows, where the material to be predicted is a sheet metal for which the forming limit needs to be predicted.

[0126] The steps of obtaining the input parameters of the material to be predicted and predicting the forming limit curve based on the data parameters and the trained neural network model include:

[0127] Take the input parameters of the material to be predicted, x = [k1, k2, k3, n i , r 00 , r 45 , r 90 , and the acquisition method of the input parameters is the same as the previous steps S1 - S2, so no additional elaboration is made here;

[0128] Substitute the input parameters into the trained neural network model to output the forming limit parameters of the material to be predicted, y = [ε0, ε1, θ1, θ2]:

[0129] Substitute the above - mentioned forming limit parameters into the forming curve equation to obtain the forming limit curve of the material to be predicted.

[0130] To better illustrate the purpose and advantages of the present invention, the following further describes the present invention in combination with specific embodiments.

[0131] a. Uniaxial tension experiment and raw data processing. For specific data, please refer to Figures 5 - 7 ;

[0132] b. Bulging experiment and forming limit strain scatter data. For specific data, please refer to Figures 8 - 9 ;

[0133] c. True stress - strain conversion, extrapolation fitting of the engineering stress - strain curve and integration of sample input data. Please refer to Figure 10 as shown;

[0134] d. Fitting of forming limit curve parameters and sample targets. Please refer to Figure 11 as shown;

[0135] e. Establishment and training of the neural network. Please refer to Figure 3 , and the following is one implementation of the establishment and training of the neural network;

[0136] Neural network model 1: MLP multi-layer perceptron neural network model;

[0137] Let this network be a three-layer fully connected neural network. The first layer is the input layer, the middle layer is the hidden layer, and the last layer is the output layer. The activation function of the middle layer adopts ReLU, and the loss function can use the mean square error. Let the number of nodes in the hidden layer be s, then the structure of this neural network is as follows Figure 3 shown;

[0138] Two-layer connection weight coefficients w ij-1 and w ij-2 can be expressed as

[0139]

[0140] and

[0141]

[0142] The following is another implementation of the establishment and training of the neural network;

[0143] Neural network model 2: 1D-CNN one-dimensional convolutional neural network;

[0144] Let this network be a five-layer convolutional neural network. The first layer is the input layer, the middle three layers are the hidden layers, and the last layer is the output layer. The middle three layers are the convolutional layer, the pooling layer, and the fully connected layer respectively. The convolutional layer uses a one-dimensional convolutional kernel for convolution; the pooling layer is average pooling; the activation function all adopts leaky_ReLU, and the loss function can use the mean square error. Then the structure of this neural network is as follows Figure 15 shown:

[0145] Then use the input in (c) and the target obtained in (d) to train the neural network, as Figure 4 shown.

[0146] f. Predict the forming limit curve based on the uniaxial data of the new sheet metal based on the fully trained neural network;

[0147] Obtain the uniaxial tensile test data of the new sheet metal, as Figure 12 shown;

[0148] Fit the hardening curve parameters, as Figure 13 shown;

[0149] Uniaxial performance parameters, neural network input:

[0150] x = [k1, k2, k3, n i , r 00 , r 45 , r 90= [1048.8, 0.0026, 0.0, 0.1228, 0.60, 0.73, 1.12];

[0151] Forming curve parameters, neural network model 1: Output result of the MLP multi-layer perceptron neural network model:

[0152] y = [ε0, ε1, θ1, θ2] = [0.269, 0.710, 1.041, 1.886]

[0153] Or neural network model 2: Output result of the 1D-CNN one-dimensional convolutional neural network:

[0154] y = [ε0, ε1, θ1, θ2] = [0.272, 0.690, 1.055, 2.002]

[0155] Calculate the forming limit curve according to the corresponding forming limit curve equation, and the result is consistent with the strain scatter points of the forming limit curve measured in the actual bulging experiment. The result is as Figure 14 shown.

[0156] In summary, the method of the present invention for quickly predicting the forming limit based on a simple uniaxial tension experiment and a neural network abandons the traditional complex forming limit experiment or theoretical calculation method. The forming limit curve predicted by the fully trained neural network has the advantages of high efficiency and high accuracy, and the data source for training the neural network is extensive, which can be obtained through self-experimentation according to the method described in this patent, or any existing experimental data can be used. The predicted results can be used for engineering problems including stamping process analysis and sheet metal failure prediction, improving the efficiency and accuracy of sheet metal forming-related simulations.

[0157] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention. The structures, devices, and operation methods not specifically described and explained in the present invention are implemented according to the conventional means in the art without special description and limitation.

Claims

1. A metal forming limit curve prediction method based on neural network, characterized in that: include: Obtain the basic properties that can characterize the single-pull performance and the forming limit curve of the material; Perform parameterization on basic properties and forming limit curves to obtain strain parameters and parameterized curves; Obtaining a neural network model, and training the neural network model according to strain parameters and parameterized curves; The input parameters of the material to be predicted are obtained, and the forming limit curve is predicted according to the data parameters and based on the trained neural network model.

2. The method for predicting metal forming limit curves based on neural networks according to claim 1, characterized in that: The step of obtaining basic performance that can characterize single pull performance includes: Carry out uniaxial tensile tests; Measure the time history including extensometer deformation and cross-sectional load, and calculate the corresponding engineering stress-strain curve; Use non-contact strain measurement equipment to test, obtain surface strain, and calculate anisotropy coefficient; The surface strain includes the engineering stress-strain curve of the material in three directions and the anisotropy coefficient of the material in three directions.

3. The method for predicting metal forming limit curves based on neural networks according to claim 2, characterized in that: The engineering stress-strain curves of the materials in the three directions are: I0，l is the extensometer length before and after the experiment, F is the test machine load, A0 is the original cross-sectional area of ​​the sample, combined e eng ,s eng Then the engineering stress-strain curve can be obtained; The anisotropy coefficient of the material in the three directions is calculated as follows: in, w 0, w is the width of the measuring point before and after the experiment, that is, along the vertical direction of stretching, h 0, h is the length of the measuring point before and after the experiment, that is, along the tensile direction, and r is the anisotropy coefficient in the test direction.

4. The method for predicting metal forming limit curves based on neural networks according to claim 1, characterized in that: The forming limit curve of the material is obtained by performing a forming limit experiment, wherein the limit strain scatter points under different stresses need to be obtained in the forming limit experiment.

5. The method for predicting metal forming limit curves based on neural network according to claim 4, characterized in that: The method of obtaining the limit strain scatter points under different stresses includes: set up ε1 , ε2 The maximum principal and secondary strains at the cracking point measured by the bulging experiment are: n is the number of all measuring points, the forming limit test data can be recorded as:

6. The method for predicting metal forming limit curves based on neural network according to claim 5, characterized in that: The step of performing parameterization processing on the basic performance and the forming limit curve to obtain the strain parameter and the parameterized curve comprises: After true stress-strain conversion and extrapolation fitting of the engineering stress-strain curve, the corresponding parameter strain parameters are obtained as sample input of the neural network; The parameterized curve is obtained by curve fitting the limit strain scatter points obtained from the forming limit experiment, which is used as the target sample of the neural network corresponding to the single-pull data input.

7. The method for predicting metal forming limit curves based on neural networks according to claim 1, characterized in that: The steps of obtaining the strain parameters are as follows: The engineering stress-strain curve is first truncated at the peak position; Convert to true stress-strain curve through the following formula; ε true ＝l(1+ ε eng )； σ true ＝ σ eng *(1+ ε eng )； The real stress-strain curve is fitted by extrapolation formula, and the fitting method can be used but not limited to Swift and Ghosh; s eq =k1*(k2+ε eq ) n +k3; The anisotropy coefficient r value obtained by adding the fitting parameters of the above true stress-strain curve can constitute the sample input item of the neural network. Assuming that the number of samples is m, the sample input matrix of the single-pull information is: remember x i =[k 1-i ,k 2-i ,k 3-i ,n i ,r 00-i ,r 56-i ,r 90-i ] , i=1,2,3,...,m , then the sample input matrix is ​​simplified as:

8. The method for predicting metal forming limit curves based on neural network according to claim 1, characterized in that: The neural network model is MLP multi-layer perceptron neural network model or 1D-CNN one-dimensional convolutional neural network .

9. The method for predicting metal forming limit curves based on neural network according to claim 1, characterized in that: The step of training the neural network model according to the strain parameters and the parameterized curve comprises: The strain parameters are used as sample inputs of the neural network, and the parameterized curves are used as target samples of the neural network corresponding to the strain parameters to train the neural network model. The actual training method may adopt but is not limited to batch gradient descent method or stochastic gradient descent method.

10. The method for predicting metal forming limit curves based on neural network according to claim 1, characterized in that: The step of obtaining input parameters of the material to be predicted and predicting the forming limit curve according to the data parameters and based on the trained neural network model includes: Get the input parameters of the material to be predicted; Substitute the input parameters into the trained neural network model and output the forming limit parameters of the material to be predicted: Substituting the forming limit parameters into the forming curve equation, the forming limit curve of the material to be predicted is obtained.

Images