Metal plate forming equipment processing parameter determination method based on graph confrontation driven self-learning framework

Abstract

The invention discloses a method for determining processing parameters of sheet metal forming equipment based on a graph adversarial driven self-learning framework. The method comprises the following steps: establishing a database fusing adaptive evolution and key factor perception; establishing a graph adversarial driven self-learning framework theoretical model; dynamic parameter optimization of adversarial training and enhanced exploration; fixed machining parameter high-precision prediction based on a collaborative framework; and wide-area parameter prediction driven by self-adaptive regulation and control and an optimization strategy. Through adaptive evolution, adversarial learning, enhanced exploration and collaborative optimization, high-precision and high-adaptability processing parameter prediction and optimization are realized.

Description

Technical Field

[0001] The present invention relates to a method for determining machining parameters of a metal sheet forming machine tool, and more particularly to a method for determining machining parameters of a metal sheet forming device based on a graph adversarial driven self-learning framework. Background Art

[0002] In the field of metal sheet forming, the determination of machining parameters directly affects the forming quality, machining efficiency and equipment life. Traditional methods for determining machining parameters mainly rely on empirical formulas, experimental trial and error or statistical analysis based on limited data. In the face of complex non-linear machining processes, multi-source parameter interactions and dynamically changing thermal fields, these methods often suffer from problems such as insufficient accuracy, poor adaptability and low computational efficiency. In view of the above problems, the present invention proposes a method for determining machining parameters of a metal sheet forming machine tool based on a graph adversarial driven self-learning framework. This method innovatively combines graph neural networks (GNNs) with adversarial training to construct a self-learning collaborative optimization framework, providing an efficient and accurate parameter determination method for metal sheet forming, which has become an urgent problem to be solved. Summary of the Invention

[0003] Object of the Invention: The object of the present invention is to provide a method for determining machining parameters of a metal sheet forming device based on a graph adversarial driven self-learning framework.

[0004] Technical Solution: The method for determining machining parameters of a metal sheet forming device based on a graph adversarial driven self-learning framework according to the present invention includes the following steps:

[0005] (1) Establishment of a database integrating adaptive evolution and key factor perception;

[0006] (2) Establishment of a theoretical model of a graph adversarial driven self-learning framework;

[0007] (3) Dynamic parameter optimization of adversarial training and reinforcement exploration;

[0008] (4) High-precision prediction of fixed machining parameters based on a collaborative framework;

[0009] (5) Wide-area parameter prediction driven by adaptive regulation and optimization strategy.

[0010] Further, the step (1) includes:

[0011] Model the non-linear relationship between input features such as temperature, stress and displacement and machining quality. The Kriging model uses the Gaussian process regression method and can predict machining quality under different cutting process conditions by backpropagation and adaptively adjusting the parameters of the model.

[0012] Further, in the step (1), the Kriging model is approximately represented as the sum of an arbitrary bulk function and a polynomial:

[0013]

[0014] where \(g(x)\) is the unknown Kriging model; \(x = [x_1, x_2,\cdots, x_{ n \) is the design variable vector; \(f(x)=[f_1(x), f_2(x),\cdots, f_{ p}(x)]\) is the known function of \(x\), providing the global regression model; \(\beta = [\beta_1, \beta_2,\cdots, \beta_{ p \) is the undetermined coefficient of the regression function; \(z(x)\) represents a random process that follows a normal distribution with a mean of zero and a standard deviation of Its covariance matrix can be expressed as:

[0015] \(\text{cov}[z(x_{ (i)}), z(x_{ (j)})]=\sigma_{ 2}R[R(x_{ (i)}), R(x_{ (j)})]\) \(i, j = 1, 2,\cdots, m\)

[0016] where \(R\) is the correlation matrix; \(R(x_{ (i) , x_{ (j)})\) is the correlation function between any two sample points; \(M\) is the number of sample points, and \(R(x_{ (i) , x_{ (j)})\) can be selected in various forms. In this study, the Gaussian correlation function is selected, and the expression is:

[0017]

[0018] where \(\theta_{ k}(k = 1, 2,\cdots, m)\) are the unknown correlation parameters, \(y_{ pred}\) is the predicted value of the processing quality, \(f(x)\) is the processing quality predicted by the Kriging model, and \(\varepsilon\) is the error term;

[0019] Calculate the main effects, interaction effects, and high-order effects of each input feature, and evaluate their effects on the processing quality. Obtain the sensitivity index of each input parameter and determine the most important influencing factors.

[0020] Further, in the step (2), a self-learning framework based on graph adversarial learning is constructed, and a graph neural network is used to express the correlation between processing parameters. Each node represents a process parameter, and the weight of the edge represents their contribution to the final forming quality.

[0021] Further, it is characterized in that the process of constructing the self-learning framework based on graph adversarial learning includes:

[0022] Let the graph \(G=(V, E)\) composed of processing parameters be such that:

[0023] where \(V = \{v_1, v_2,\cdots, v_{n}\}\) is a set of \(N\) processing parameters, and each node \(v_i\) N represents a parameter. i

[0024] And \(E=\{e_{ij}\}\) is the connection relationship between parameters, where \(e_{ij}\) ij represents the influence degree between parameter \(v_i\) ij and \(v_j\). i j

[0025] At the \(t\)-th round of iteration, let the eigenvector of node \(v_i\) be represented as i \(\mathbf{h}_i^{(t)}\). Then the GNN updates the parameter relationship using the message passing mechanism: \(\mathbf{h}_i^{(t + 1)}=\sigma\left(\mathbf{W}_1\mathbf{h}_i^{(t)}+\mathbf{W}_2\sum_{j\in\mathcal{N}(i)}\alpha_{ij}\mathbf{h}_j^{(t)}\right)\)

[0026] where \(\mathbf{W}_1\), \(\mathbf{W}_2\) are trainable weight matrices, \(\sigma\) is a non - linear activation function, \(\mathcal{N}(i)\) is the neighbor set of node \(v_i\), and \(\alpha_{ij}\) i is the dynamic weight of the adjacency matrix, defined as: ij \(\alpha_{ij}=\frac{\exp\left(\mathbf{a}_{ij}\right)}{\sum_{k\in\mathcal{N}(i)}\exp\left(\mathbf{a}_{ik}\right)}\)

[0027]

[0028] A generative adversarial network (GAN) is used for data augmentation and parameter optimization. The generator \(G\) θ is responsible for generating processing parameter combinations that conform to physical laws, and the discriminator \(D\) φ is responsible for distinguishing between real processing data and the data generated by the generator.

[0029] The GAN adversarial training objective is:

[0030]

[0031] Through adversarial training, the learning ability of the model under scarce data and complex working conditions is enhanced.

[0032] Furthermore, step (3) includes:

[0033] Precise optimization is carried out using adversarial training and reinforcement exploration. The optimization objectives of the generator and discriminator are:

[0034]

[0035] At the same time, KL - divergence regularization is added to constrain the generator parameters to approximate the real parameter distribution:

[0036] ​​​

[0037] Final generator optimization objective:

[0038]

[0039] Among them, λ controls the weight of KL divergence regularization;

[0040] DQN uses Q-learning for parameter optimization:

[0041] Q(S t , A t ) = R t +γmax A′ Q(S t+1 , A′)

[0042] Among them: γ is the discount factor, and max A′ Q(S t+1 , A′) controls the importance of future rewards and selects the optimal future action;

[0043] DQN training loss function:

[0044]

[0045] Furthermore, the step (4) includes:

[0046] Establish a non-linear mapping using an adaptive database and a graph adversarial learning framework: F = g(σ y , E, v, d, T, H)

[0047] And through learning and optimization, minimize the prediction error: min θ E[(F 真实 -F 预测 ) 2

[0048] Through step (1), fuse the adaptive evolution and key factor perception database establishment to screen out the key factors affecting the forming force, and improve the data quality through evolutionary optimization.

[0049] Furthermore, in the step (4), the forming force F is an important parameter of the metal sheet forming machine tool. Use the dynamic parameter optimization of adversarial training and reinforcement exploration in step (3) to predict the forming force F,

[0050] GNN uses a message passing mechanism to update node information:

[0051]

[0052] Finally, obtain the global processing parameter feature representation: H = GNN(G, X),

[0053] ​Adopt residual learning to correct the prediction error of the forming force:

[0054] Train a lightweight network to predict the error:

[0055] Finally correct the prediction:

[0056] Final optimization goal:

[0057] Furthermore, the step (5) includes:

[0058] Through adaptive regulation, deep learning modeling and optimization strategies, achieve high-precision prediction of any processing parameters to adapt to different material and process conditions,

[0059] Set the processing parameter set: X = {x1, x2, K x N}

[0060] where, x i represents a certain processing parameter, such as laser power, cutting speed, auxiliary gas pressure, forming force, and the goal is to construct a non-linear mapping: u = f(X)

[0061] where, y is the processing quality index, and optimize the parameters:

[0062] min θ E[(y 目标 - y 预测 ) 2

[0063] Based on the processing parameter features extracted by GNN, use GAN for data augmentation, and construct a prediction model to achieve the prediction of any processing parameters.

[0064] Furthermore, the step (5) using GAN for data augmentation includes:

[0065] Through adversarial training, make the processing parameters generated by GAN closer to the real data distribution, and enhance the adaptability of the model in the wide-area parameter space.

[0066] Adopt a multi-layer perceptron to predict any processing parameter:

[0067] h1 = σ(W1H + b1)

[0068] h2 = σ(W2H + b2)

[0069] y 预测 = W3h3 + b3

[0070] Adopt the mean square error (MSE) for error optimization: ​

[0071] Global parameter search using Bayesian optimization:

[0072] The expected improvement EI is used to select the optimal parameters: EI(X) = (μ(X) - y * )Φ(Z) + σ(X)φ(Z).

[0073] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: The present invention combines a graph neural network (GNN) with adversarial training to construct a self-learning collaborative optimization framework, which not only provides an efficient and accurate method for determining parameters in metal sheet forming processing, but also provides a new technical path for the intelligent optimization of complex manufacturing processes; through adaptive evolution, adversarial learning, reinforcement exploration, and collaborative optimization, high-precision and high-adaptability processing parameter prediction and optimization are achieved. Brief description of the drawings

[0074] Figure 1 It is a flowchart of the present invention. Detailed implementation manners

[0075] The technical solution of the present invention will be further described below with reference to the drawings.

[0076] As Figure 1 shown, the present invention includes the following steps:

[0077] (1) Establishment of a database integrating adaptive evolution and key factor perception

[0078] During the processing of a metal sheet forming machine tool, the influence of different process parameters on the forming quality is highly non-linear. In order to effectively extract key influencing factors and establish accurate data support, this step adopts an adaptive evolution and key factor perception mechanism to construct an efficient database.

[0079] First, a non-linear relationship between input features such as temperature, stress, and displacement and the processing quality is modeled. The Kriging model uses the Gaussian process regression method and can predict the processing quality under different cutting process conditions by backpropagation and adaptively adjusting the parameters of the model.

[0080] The Kriging model can be approximately expressed as the sum of an arbitrary bulk function and a polynomial.

[0081]

[0082] Among them, g(x) is the unknown Kriging model; x = [x1, x2,..., x n is the design variable vector (n is the number of design variables); is a known function of x, providing a global regression model; are the undetermined coefficients of the regression function; z(x) represents a stochastic process that follows a normal distribution with a mean of zero and a standard deviation of Its covariance matrix can be expressed as:

[0083] cov[z(x (i) ),z(x (j) )]=σ 2 R[R(x (i) ),R(x (j) )]i,j=1,2,…,m

[0084] where R is the correlation matrix; R(x (i) ,x (j) ) is the correlation function between any two sample points; M is the number of sample points. R(x (i) ,x (j) ) can take various forms. In this study, the Gaussian correlation function is selected, and its expression is:

[0085]

[0086] where θ k (k=1,2,...m) are unknown correlation parameters.

[0087] where y pred is the predicted value of the processing quality, f(x) is the processing quality predicted by the Kriging model, and ε is the error term.

[0088] Calculate the main effects, interaction effects, and higher-order effects of each input feature, and evaluate their effects on the processing quality. Obtain the sensitivity index of each input parameter to determine the most important influencing factors.

[0089] (2) Establishment of the theoretical model of the graph adversarial-driven self-learning framework

[0090] To capture the complex multivariable correlation relationships in the metal sheet forming process, this step constructs a self-learning framework based on graph adversarial learning (Graph-GAN), using a graph neural network (GNN) to express the correlations between processing parameters. Each node represents a process parameter, and the weight of the edge represents their contribution to the final forming quality.

[0091] Let the graph G=(V, E) composed of processing parameters, where:

[0092] γ={v1,v2,...,v n} is a set of N processing parameters, and each node v i represents a parameter.

[0093] E={eij} is the connection relationship between parameters, e ij represents parameter v i and v j The degree of influence between them.

[0094] At the t-th round of iteration, the eigenvector representation of node v i is expressed as Then the GNN updates the parameter relationship using the message passing mechanism:

[0095] where W1 and W2 are trainable weight matrices, and σ is a non-linear activation function, is the neighbor set of node v i , and α ij is the dynamic weight of the adjacency matrix, defined as:

[0096]

[0097] To improve the applicability of the model, a generative adversarial network (GAN) is used for data augmentation and parameter optimization.

[0098] The generator G θ is responsible for generating machining parameter combinations that conform to physical laws.

[0099] The discriminator D φ is responsible for distinguishing between real machining data and data generated by the generator.

[0100] GAN adversarial training objective:

[0101]

[0102] Through adversarial training, the learning ability of the model under scarce data and complex working conditions is enhanced.

[0103] The reinforcement learning (RL) mechanism is adopted to continuously adjust the machining parameters through environmental feedback to make them tend to the optimal solution.

[0104] Policy Gradient (PG) optimization is adopted:

[0105]

[0106] where π θ (A t |S t ) is the output of the policy network.

[0107] The Deep Q-Network (DQN) is used for value function approximation:

[0108] Q(S t , A t ) = Rt +γmax A′ Q(S t+1 ,A′)

[0109] where γ is the discount factor.

[0110] Self-learning by combining GAN and RL:

[0111]

[0112] Combined with transfer learning, improve the adaptability of the model to different plates and working conditions and accelerate the convergence speed.

[0113] (3) Dynamic parameter optimization of adversarial training and reinforcement exploration

[0114] During the processing, the dynamic optimization of process parameters is crucial for the forming quality. This step uses adversarial training and reinforcement exploration for precise optimization. Optimization objectives of the generator and discriminator:

[0115]

[0116] At the same time, add KL divergence regularization to constrain the generator parameters to approach the true parameter distribution:

[0117]

[0118] Final optimization objective of the generator:

[0119]

[0120] where λ controls the weight of KL divergence regularization.

[0121] DQN uses Q-learning for parameter optimization:

[0122] Q(S t ,A t )=R t +γmax A′ Q(S t+1 ,A′)

[0123] where: γ is the discount factor, max A′ Q(S t+1 ,A′) controls the importance of future rewards. Select the optimal future action.

[0124] DQN training loss function:

[0125]

[0126] (4) High-precision prediction of fixed processing parameters based on a collaborative framework

[0127] The forming force F is an important parameter of a metal sheet forming machine tool, which is affected by multiple process factors, such as: material parameters (yield strength, elastic modulus), equipment parameters (slider speed, die clearance), environmental variables (temperature, humidity).

[0128] Establish a non-linear mapping using an adaptive database and a graph adversarial learning framework: F = g(σ y , E, v, d, T, H)

[0129] And through learning and optimization, minimize the prediction error: min θ E[(F 真实 - F 预测 ) 2

[0130] Through step (1), screen out the key factors affecting the forming force by integrating the adaptive evolution and key factor perception database, and improve the data quality through evolutionary optimization.

[0131] Use the dynamic parameter optimization of step (3) for adversarial training and reinforcement exploration to predict the forming force F.

[0132] The GNN updates the node information using a message passing mechanism:

[0133]

[0134] Finally, obtain the global processing parameter feature representation: H = GNN(G, X)

[0135] Adopt residual learning to correct the forming force prediction error:

[0136] Train a lightweight network for prediction error:

[0137] Finally, correct the prediction:

[0138] Final optimization goal:

[0139] (5) Wide-area parameter prediction driven by adaptive regulation and optimization strategy

[0140] For the machining of machine tools with different materials and different process requirements, this step conducts intelligent prediction and optimization in a wider parameter space to ensure that the machining parameters can adapt to various working conditions and improve the generalization ability and accuracy of prediction. Through adaptive regulation, deep learning modeling and optimization strategies, high-precision prediction of any machining parameters is achieved to adapt to different materials and process conditions.

[0141] Let the machining parameter set: X = {x1, x2,... x N} ​

[0142] Among them, x i represents a certain processing parameter, such as laser power, cutting speed, auxiliary gas pressure, forming force, etc.

[0143] The goal is to construct a non - linear mapping: y = f(X)

[0144] where y is the processing quality index (such as surface roughness, hardness, cutting width, etc.), and optimize the parameters:

[0145] min θ E[(y 目标 - y 预测 ) 2

[0146] Based on the processing parameter features extracted by GNN, use GAN for data augmentation, and construct a prediction model to achieve the prediction of any processing parameter.

[0147] Through adversarial training, make the processing parameters generated by GAN closer to the real data distribution, and enhance the adaptability of the model in the wide - area parameter space.

[0148] Use MLP (Multi - Layer Perceptron) to predict any processing parameter:

[0149] h1 = σ(W1H + b1)

[0150] h2 = σ(W2H + b2)

[0151] y 预测 = W3h3 + b3

[0152] Use mean squared error (MSE) for error optimization:

[0153] In order to improve the optimization efficiency, use Bayesian optimization for global parameter search: Use expected improvement (EI) to select the optimal parameters: EI(X) = (μ(X) - y * )Φ(Z)+σ(X)φ(Z).

[0154] Achieve high - precision processing parameter prediction in the wide - area parameter space. This method can adapt to the complex working conditions of different materials and different processing methods, improve the applicability of the model, ensure the optimality of processing parameters in different environments, and achieve the goal of intelligent manufacturing.​

Claims

1. A method for determining processing parameters of a metal sheet forming device based on a graph adversarial-driven self-learning framework, characterized in that, It includes the following steps: (1) Establishment of a database integrating adaptive evolution and key factor perception; (2) Establishment of a theoretical model for a graph adversarial-driven self-learning framework; (3) Dynamic parameter optimization through adversarial training and reinforcement exploration; (4) High-precision prediction of fixed machining parameters based on a collaborative framework; (5) Wide-area parameter prediction driven by an adaptive regulation and optimization strategy.

2. The method for determining the processing parameters of a metal sheet forming device based on a graph adversarial-driven self-learning framework according to claim 1, wherein The step (1) includes: Model the non-linear relationship between input features such as temperature, stress, and displacement and machining quality. The Kriging model uses the Gaussian process regression method and can predict machining quality under different cutting process conditions by backpropagation and adaptively adjusting the model's parameters.

3. The method for determining the processing parameters of a metal sheet forming device based on the graph adversarial driven self-learning framework according to claim 2, wherein In the step (1), the Kriging model is approximately expressed as the sum of an arbitrary bulk function and a polynomial: Among them, g(x) is an unknown Kriging model; x = [x1, x2,..., x n is the design variable vector; f(x) = [f1(x), f2(x),..., f p (x)] is a known function of x, providing a global regression model; β = [β1, β2,..., β p are the undetermined coefficients of the regression function; z(x) represents a random process that follows a normal distribution with a mean of zero and a standard deviation of Its covariance matrix can be expressed as: cov[z(x (i) ),z(x (j) )]=σ 2 R[R(x (i) ),R(x (j) )]i,j=1,2,…,m Among them, R is the correlation matrix; R(x (i) , x (j) ) is the correlation function between any two sample points; M is the number of sample points. R(x (i) , x (j) ) can take various forms. In this study, the Gaussian correlation function is selected, and its expression is: where θ k (k = 1, 2,... m) are unknown correlation parameters, y pred is the predicted value of the machining quality, f(x) is the machining quality predicted by the Kriging model, and ε is the error term; Calculate the main effect, interaction effect, and high-order effect of each input feature, evaluate their impact on machining quality, obtain the sensitivity index of each input parameter, and determine the most important influencing factors.

4. The method for determining the processing parameters of a metal sheet forming device based on a graph adversarial driven self-learning framework according to claim 1, wherein The step (2) constructs a self-learning framework based on graph adversarial learning, uses a graph neural network to express the correlation between machining parameters, each node represents a process parameter, and the weight of the edge represents their contribution to the final forming quality.

5. The method for determining the processing parameters of a metal sheet forming device based on a graph adversarial driven self-learning framework according to claim 4, characterized in that, The process of constructing the self-learning framework based on graph adversarial learning includes: Let the graph G=(V, E) composed of machining parameters, where: γ = {v1, v2,..., v N} is a set of N processing parameters, and each node v i represents a parameter. E = {e ij} is the connection relationship between the parameters, and e ij represents the influence degree between the parameters v i and v j . At the t-th round of iteration, for node v i the eigenvector representation is Then the GNN updates the parameter relationship using the message passing mechanism: where W1, W2 are trainable weight matrices, σ is a non-linear activation function, is the set of neighbors of node v i , and α ij is the dynamic weight of the adjacency matrix, defined as: Data augmentation and parameter optimization are carried out using a generative adversarial network (GAN), and the generator G θ is responsible for generating combinations of processing parameters that conform to physical laws, and the discriminator D φ is responsible for distinguishing between real processing data and the data generated by the generator GAN adversarial training objective: Through adversarial training, enhance the model's learning ability under scarce data and complex working conditions.

6. The method for determining the processing parameters of a metal sheet forming device based on a graph adversarial driven self-learning framework according to claim 1, wherein The step (3) includes: Use adversarial training and reinforcement exploration for precise optimization, the optimization objectives of the generator and discriminator: At the same time, add KL divergence regularization to constrain the generator parameters to approach the true parameter distribution: Final generator optimization objective: Among them, λ controls the weight of KL divergence regularization; DQN uses Q-learning for parameter optimization: Q(S t , A t ) = R t + γmax A′ Q(S t+1 , A′) where: γ is the discount factor, max A′ Q(S t+1 , A′) controls the importance of future rewards and selects the optimal future action; DQN training loss function:

7. The method for determining the processing parameters of a metal sheet forming device based on a graph adversarial-driven self-learning framework according to claim 1, wherein The step (4) includes: Establishing a non-linear mapping using an adaptive database and a graph adversarial learning framework: F = g(σ y , E, v, d, T, H) And through learning optimization, minimize the prediction error: min θ E[(F 真实 -F 预测 ) 2 ​ Screen out the key factors affecting the forming force through the establishment of a database integrating adaptive evolution and key factor perception in step (1), and improve the data quality through evolutionary optimization.

8. The method for determining the processing parameters of a metal sheet forming device based on the graph adversarial driven self-learning framework according to claim 1, characterized in that In the step (4), the forming force F is an important parameter of the metal sheet forming machine tool. Use the dynamic parameter optimization of adversarial training and reinforcement exploration in step (3) to predict the forming force F. The GNN updates node information using a message passing mechanism: Finally, obtain the global machining parameter feature representation: H = GNN(G, X). Using residual learning to correct the prediction error of the forming force: Train a lightweight network for prediction error: Final corrected prediction: Final optimization goal:

9. The method for determining the processing parameters of a metal sheet forming device based on a graph adversarial driven self-learning framework according to claim 1, characterized in that The step (5) includes: Through adaptive regulation, deep learning modeling, and optimization strategies, achieve high-precision prediction of any machining parameter to adapt to different materials and process conditions. Let the set of processing parameters: X = {x1, x2,... x N} where x i represents a certain processing parameter, such as laser power, cutting speed, auxiliary gas pressure, forming force The goal is to construct a non-linear mapping: y = f(X) Among them, y is the machining quality index, and optimize the parameters: min θ E[(y 目标 -y 预测 ) 2 ​ Based on the machining parameter features extracted by the GNN, use GAN for data augmentation and construct a prediction model to achieve the prediction of any machining parameter.

10. The method for determining the processing parameters of a metal sheet forming device based on the graph adversarial driven self-learning framework according to claim 9, characterized in that, The data augmentation using GAN in the step (5) includes: Through adversarial training, make the machining parameters generated by GAN closer to the true data distribution and enhance the model's adaptability in the wide-area parameter space. Use a multi-layer perceptron to predict any machining parameter: h1 = σ(W1H + b1) h2 = σ(W2H + b2) y 预测 = W3h3 + b3 The mean squared error (MSE) is used for error optimization: Global parameter search using Bayesian optimization: Use expected improvement EI to select the optimal parameters: EI(X) = (μ(X) - y * )Φ(Z) + σ(X)φ(Z).

Images