Large-size complex forging forming load prediction method based on deep neural network model

Abstract

The application relates to a large-size complex forging forming load prediction method based on a deep neural network model, which comprises the following steps: firstly, obtaining the forming load of a large-size forging under different forging process parameters; secondly, performing standard deviation standardization processing on a sample data set composed of the forging process parameters and corresponding forming loads to obtain a standardized sample data set; finally, establishing a hyperparameter search space of a deep neural network candidate model Ω , using the standardized sample data set to perform model training on different candidate hyperparameter combinations in the hyperparameter search space, thereby establishing a forging forming load prediction model with the forging process parameters as input and the forming load as output, and realizing large-size forging forming load prediction; the application realizes rapid and accurate prediction of the forming load by constructing the deep neural network model, and effectively overcomes the limitations of large calculation resource consumption and low optimization efficiency of traditional finite element simulation in a complex forging process.

Description

Technical Field

[0001] This invention relates to the field of plastic forming of metal components. Background Technology

[0002] Large-sized complex forgings play an irreplaceable role in high-end manufacturing. However, during the forging process, these forgings are subjected to non-uniform thermo-mechanical coupling fields, making them highly susceptible to plastic flow instability, uneven deformation, and internal damage. These problems not only affect the forming quality of the forgings but also weaken their mechanical properties and service life. While traditional macroscopic finite element simulation can predict forming behavior, it suffers from limitations such as huge computational resource consumption and lengthy processing times when dealing with complex forging processes involving multiple steps and long flows. To address this, we introduce deep neural networks (DNNs), a machine learning algorithm, to construct an efficient data-driven surrogate model using data obtained from finite element simulations. This model accurately reflects the nonlinear relationship between complex process parameters and forming parameters, enabling millisecond-level rapid calculations while ensuring prediction accuracy. This provides an efficient new approach for optimizing the process parameters of large-sized components.

[0003] Reference 1, "Jia Dewen, Sun Yan, Deng Wei, et al. Study on the Influencing Factors of Connecting Rod Forging Load Based on Response Surface Method [J]. Precision Forming Engineering, 2023, 15(5): 193-201," studies the pre-forging process of 42CrMo connecting rods. Addressing the problem of high mold damage risk and low service life due to large forming loads, a finite element simulation model for connecting rod pre-forging was established, and the influence of billet temperature, mold temperature, forging speed, and friction factor on the forming load was analyzed. This reference uses the response surface method to establish a quadratic regression model for calculating the forming load and combines it with a multi-island genetic algorithm for process parameter optimization. The results show that the significance of each process parameter on the forming load, from largest to smallest, is billet temperature, friction factor, forging speed, and mold temperature; the coefficient of determination R² of its response surface model is 0.965, indicating that the model has a certain fitting accuracy and predictive ability. Subsequently, this paper uses the minimum forming load as the optimization objective and combines a multi-island genetic algorithm to globally optimize the process parameters. The optimized process parameters obtained are: billet temperature 1190℃, die temperature 350.7℃, forging speed 532.5mm / s, and friction coefficient 0.1. Under this parameter combination, the forming load is 3226.6kN, which is 50.74% lower than the original process. However, the technical solution in paper 1 mainly focuses on the optimization of process parameters for 42CrMo connecting rod pre-forgings, which is essentially a load fitting and optimization method based on response surface statistical regression. Compared with connecting rod pre-forgings, large-size complex forgings have characteristics such as large geometric dimensions, large cross-sectional thickness, long heat conduction paths, and uneven distribution of deformation areas. There are more complex nonlinear relationships between forming load and deformation temperature, friction state, pressing speed, and deformation amount. Paper 1 does not use deformation amount as an input parameter affecting forming load, nor does it establish a data-driven deep neural network prediction model for large-size complex forgings. Therefore, although Reference 1 can reduce the forming load within a specific connecting rod pre-forging process range, it is difficult to achieve the rapid prediction of forming load for large-size complex forgings required by this invention.

[0004] Reference 2, "Tang Xuefeng, Wang Zhizhou, Feng Yi, et al. Intelligent optimization of forging process of TC18 titanium alloy landing gear based on machine learning [J]. Journal of Plasticity Engineering, 2024, 31(4):254-266," addresses the challenges of narrow processing window, difficulty in predicting part forming quality and performance, and optimization of process parameters for TC18 titanium alloy. First, hot compression tests were conducted to analyze the hot deformation behavior and microstructure evolution of TC18 titanium alloy. Then, a flow stress DNN model and a microstructure DNN model were optimized based on the Bayesian algorithm. The flow stress model takes deformation temperature, strain rate, and strain as inputs and flow stress as output, with an optimal topology of 3-13-13-1-tanh. The correlation coefficient between the actual flow stress and the predicted value in the test set was 99.97%. The microstructure model takes deformation temperature, strain rate, and strain as inputs and primary α phase content and average α phase grain size as outputs, with an optimal topology of 3-8-8-2-tanh. This paper also utilizes secondary development with UG and Deform software to achieve automatic modeling, assembly, and batch simulation of landing gear pre-forgings. It establishes a pre-forging size-process-quality database and constructs a size-process-quality DNN model with pre-forging size parameters, billet temperature, and pressing rate as inputs, and maximum forming force and primary α-phase content of the target cross-section as outputs. Further optimization using GA, PSO, and NSGA2 algorithms results in a 40.6% reduction in maximum forming force and an average primary α-phase content of 20.7% in the target cross-section. However, paper 2 primarily addresses the shape design and microstructure control of TC18 titanium alloy landing gear pre-forgings. It does not establish a deep neural network prediction model for large-size complex forgings with deformation temperature, friction factor, upper die pressing speed, and deformation amount as inputs, and forming load as output. Furthermore, it does not construct a candidate training configuration set consisting of hidden layer topology, training parameters, and maximum iteration count for rapid prediction of forming loads in large-size complex forgings. Therefore, paper 2 cannot solve the problem of rapid prediction of forming loads and automatic model structure optimization for large-size complex forgings proposed in this invention. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for predicting the forming load of large-size complex forgings based on a deep neural network model. This method comprehensively considers the influence of forging temperature, friction factor, forging speed and deformation on the forming load of large-size complex forgings, can accurately predict the forming load of large-size complex forgings, and realize the effective control of the forming load of forgings.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: a method for predicting forming loads of large-size complex forgings based on a deep neural network model, comprising the following steps: Step 1: Under different forging process parameters, perform finite element simulation on large-size forgings and extract the load-stroke curve of the upper die during the pressing process to obtain the forming load of large-size forgings under different forging process parameters. Among them, the forging process parameters include the forging temperature, friction factor, forging speed and deformation amount of large-size forgings; Large-size forgings are axisymmetric disc-shaped forgings. The diameter of the disc-shaped billet for the axisymmetric disc-shaped forging is 1600mm to 2000mm, the height H is 120mm to 180mm, the target deformation is 0.70 to 0.90, and the section ratio is greater than 9. Forming load refers to the load that a large forging bears in the forming direction when the deformation reaches the target value; Step 2: The sample dataset consisting of forging process parameters and corresponding forming loads is processed using the standard deviation standardization method to eliminate differences in data magnitude and obtain a standardized sample dataset. Step 3: Establish the hyperparameter search space for deep neural network candidate models Ω The model is trained on different candidate hyperparameter combinations in the hyperparameter search space to establish a forging forming load prediction model with forging process parameters as input and forming load as output, so as to realize the prediction of forming load of large-size forgings. Among them, the hyperparameter search space is used to solve the problem that traditional neural network structures rely on manual experience settings; the determination coefficient of the test set of the forging forming load prediction model is not less than 0.97, and the root mean square error is not greater than 30MN; Candidate hyperparameters include a candidate set of hidden layer topologies. H Candidate set of training parameters P and the candidate set of maximum iteration count E .

[0007] Furthermore, in step one, the forging temperature in the forging process parameters is 350℃~550℃, and the forging speed is 0.1mm·s. -1 ~5mm·s -1 The friction factor is 0.1 to 0.7, and the deformation is 0.7 to 0.9.

[0008] Furthermore, in step two, the standard deviation of the sample dataset is calculated using the following formula: , In the formula, Let be the mean of any set of sample datasets. Let be the standard deviation of any set of sample datasets.

[0009] Furthermore, in step three, the candidate set of hidden layer topologies...H Candidate set of training parameters P and the candidate set of maximum iteration count E The hyperparameter search space Ω formed by them are respectively expressed as:

[0010] In the formula, H Each element in the table represents a sequence of the number of neurons in a hidden layer; p j This represents the j-th candidate value for the training parameter; e k This represents the k-th candidate value for the maximum number of iterations. Among them, the hidden layer topology is used to limit the number of hidden layers and the number of neurons in each hidden layer, the training parameters are used to adjust the iterative update capability of the Bayesian regularized backpropagation algorithm, and the maximum number of iterations is used to limit the termination condition of model training.

[0011] Furthermore, in step three, a full combinatorial traversal is used to search for all hyperparameter combinations in the hyperparameter search space, and a Bayesian regularization algorithm is used for model training to obtain the modified objective function representation: , In the formula, n This represents the number of training samples; y i For the first i The true value of each sample; These are the model's predicted values; w j These are the weight parameters in the network; m This represents the total number of weight parameters in the network. α and β These are performance parameters of the objective function, used to automatically adjust and balance the proportion of mean square error and weighted sum of squares in the optimization process; By comparing the root mean square error values ​​of the validation set under all combinations and experimental sequences, the model with the smallest root mean square error value is selected as the final forging forming load prediction model, and its network weights and bias matrices are saved. Among them, the Bayesian regularization algorithm effectively avoids overfitting by modifying the objective function to minimize the mean square error while constraining the sum of squares of the network weights, and is especially suitable for regression prediction of small samples or high-dimensional data.

[0012] Furthermore, in step three, the forging forming load prediction model is a feedforward multilayer perceptron model, including an input layer, a hidden layer, and an output layer. The number of nodes in the input layer corresponds to the standardized forging temperature, friction factor, forging speed, and deformation process parameters; the nodes in the output layer correspond to the forming load prediction target. Hidden layers include L Hidden layers L The number of neurons in each hidden layer is an integer greater than or equal to 2, determined by the candidate set of hidden layer topologies. H The sequence of the number of neurons in any hidden layer is determined; a candidate set of hidden layer topologies is established. H This includes hidden layer sequences with decreasing width and hidden layer sequences with equal width; Training parameters are derived from the candidate set of training parameters. P The maximum number of iterations is arbitrarily selected from the candidate set of maximum iterations. E Choose any one from the options.

[0013] Furthermore, the hyperbolic tangent function is selected as the activation function of the forging forming load prediction model, and the expression of the activation function is: , In the formula, x is the input value of the forging forming load prediction model; e represents the natural constant; The expression for the large-size forging forming load prediction model established based on deep neural networks is as follows: , In the formula, W i Let be the weight matrix of the i-th layer; b i Let i be the bias vector of the i-th layer; Substituting the forging process parameters from the standardized sample dataset into the above formula, the expression for the forming load prediction model based on deep neural networks for optimizing the forging process of large-size forgings is as follows: , In the formula, This is the input vector for forging process parameters after standard deviation standardization. This represents the average value of the forming load; This represents the standard deviation of the forming load.

[0014] Furthermore, the standardized sample dataset is randomly divided into a training set, a validation set, and a test set according to a preset ratio. The training set is used for iterative updates of the weights of the forging forming load prediction model; the validation set is used to monitor the performance of the forging forming load prediction model and prevent overfitting during the training process; and the test set is used to finally evaluate the generalization prediction accuracy of the forging forming load prediction model.

[0015] Furthermore, multiple independent training experiments were conducted for each of the different hyperparameter combinations. After training, the root mean square error (RMSE) and coefficient of determination (R²) of each model on the validation set were calculated. 2 for: , , In the formula, N The total number of samples, y pred The predicted value of the model. y true For the true value, The average of the true values; root mean square error (RMSE) and coefficient of determination (R²) 2 Quantitative analysis of the prediction results is performed to verify the model's prediction accuracy and generalization ability.

[0016] Furthermore, the load-stroke curve of the upper die during the pressing process in step one is extracted by the post-processing module of the finite element analysis software; at the same time, the finite element simulation includes establishing the geometric model of the large-size forging die and the mesh generation of the billet numerical simulation in the finite element simulation software.

[0017] The beneficial effects of this invention are: the method provided by this invention comprehensively considers the influence of forging temperature, friction factor, forging speed and deformation on the forming load of large-size complex forgings. By constructing a deep neural network (DNN) model, it realizes rapid and accurate prediction of forming load, effectively overcoming the limitations of traditional finite element simulation in complex forging processes, which consumes a lot of computational resources and has low optimization efficiency.

[0018] The prediction model established in this invention has extremely high accuracy, with a coefficient of determination of 0.996, root mean square error of 13.84MN, and mean error of 5.4% on the validation set, and a coefficient of determination of 0.992, root mean square error of 25.03MN, and mean error of 8.8% on the test set. It can achieve rapid prediction of forming loads within milliseconds, providing an efficient and reliable digital optimization method for predicting forming loads of large-sized complex forgings. Attached Figure Description

[0019] Figure 1 This is a bar chart showing the predicted and simulated values ​​of the full-temperature-range forming load verification set for complex forgings implemented by this invention; Figure 2 This is a bar chart showing the predicted and simulated values ​​of the full-temperature-range forming load test set for complex forgings implemented in this invention. Detailed Implementation

[0020] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0021] To achieve the above objectives, the present invention provides the following specific embodiments: Example 1: A method for predicting forming loads of large-size complex forgings based on a deep neural network model, comprising the following steps: S01. Under different forging process parameters, perform finite element simulation on large-size forgings and extract the load-stroke curve of the upper die during the pressing process to obtain the forming load of large-size forgings under different forging process parameters. Among them, the load-stroke curve of the upper die during the pressing process is extracted by the post-processing module of the finite element analysis software; at the same time, the finite element simulation includes establishing the geometric model of the large-size forging die and the mesh generation of the billet numerical simulation in the finite element simulation software.

[0022] Forging process parameters include forging temperature, friction factor, forging speed, and deformation amount of large forgings; Large-size forgings are axisymmetric disc-shaped forgings. The diameter of the disc-shaped billet for the axisymmetric disc-shaped forging is 1600mm to 2000mm, the height H is 120mm to 180mm, the target deformation is 0.70 to 0.90, and the section ratio is greater than 9. Forming load refers to the load that a large forging bears in the forming direction when the deformation reaches the target value; The forging process parameters include a forging temperature of 350℃~550℃ and a forging speed of 0.1mm·s. -1 ~5mm·s -1 The friction factor is 0.1 to 0.7, and the deformation is 0.7 to 0.9.

[0023] S02. The sample dataset consisting of forging process parameters and corresponding forming loads is processed using the standard deviation standardization method to eliminate differences in data magnitude and obtain a standardized sample dataset. The standardized formula for standard deviation is as follows: , In the formula, Let be the mean of any set of sample datasets. Let be the standard deviation of any set of sample datasets.

[0024] S03. The standardized sample dataset is randomly divided into a training set, a validation set, and a test set according to a preset ratio. The training set is used for iterative updates of the weights of the forging forming load prediction model; the validation set is used to monitor the performance of the forging forming load prediction model and prevent overfitting during the training process; and the test set is used to finally evaluate the generalization prediction accuracy of the forging forming load prediction model.

[0025] S04. Establish the hyperparameter search space for deep neural network candidate models. Ω The hyperparameter search space is used to address the problem that traditional neural network structures rely on manual experience for setting; candidate hyperparameters include a candidate set of hidden layer topologies. H Candidate set of training parameters P and the candidate set of maximum iteration count E， The hyperparameter search space Ω formed by them are respectively expressed as:

[0026] In the formula, H Each element in the table represents a sequence of the number of neurons in a hidden layer; p j This represents the j-th candidate value for the training parameter; e k This represents the k-th candidate value for the maximum number of iterations. Among them, the hidden layer topology is used to limit the number of hidden layers and the number of neurons in each hidden layer, the training parameters are used to adjust the iterative update capability of the Bayesian regularized backpropagation algorithm, and the maximum number of iterations is used to limit the termination condition of model training.

[0027] S05. Train the model for different candidate hyperparameter combinations in the hyperparameter search space. Use a full combinatorial traversal to search for all hyperparameter combinations in the hyperparameter search space. Use the Bayesian regularization algorithm to train the model, and obtain the modified objective function as follows: , In the formula, n This represents the number of training samples; y i For the first i The true value of each sample; These are the model's predicted values; w j These are the weight parameters in the network; m This represents the total number of weight parameters in the network. α and β These are performance parameters of the objective function, used to automatically adjust and balance the proportion of mean square error and weighted sum of squares in the optimization process; By comparing the root mean square error values ​​of the validation set under all combinations and experimental sequences, the model with the smallest root mean square error value is selected as the final forging forming load prediction model, and its network weights and bias matrices are saved. Among them, the Bayesian regularization algorithm effectively avoids overfitting by modifying the objective function to minimize the mean square error while constraining the sum of squares of the network weights, and is especially suitable for regression prediction of small samples or high-dimensional data.

[0028] S06. Establish a forging forming load prediction model with forging process parameters as input and forming load as output to realize the prediction of forming load of large-size forgings. The determination coefficient of the test set of the forging forming load prediction model is not less than 0.97 and the root mean square error is not greater than 30MN. The forging forming load prediction model is a feedforward multilayer perceptron model, including an input layer, a hidden layer, and an output layer; The number of nodes in the input layer corresponds to the standardized forging temperature, friction factor, forging speed, and deformation process parameters; the nodes in the output layer correspond to the forming load prediction target. Hidden layers include L Hidden layers L The number of neurons in each hidden layer is an integer greater than or equal to 2, determined by the candidate set of hidden layer topologies. H The sequence of the number of neurons in any hidden layer is determined; a candidate set of hidden layer topologies is established. H This includes hidden layer sequences with decreasing width and hidden layer sequences with equal width; Training parameters are derived from the candidate set of training parameters. P The maximum number of iterations is arbitrarily selected from the candidate set of maximum iterations. E Choose any one from the options.

[0029] S07. The hyperbolic tangent function is selected as the activation function for the forging forming load prediction model. The expression of the activation function is: , In the formula, x is the input value of the forging forming load prediction model; e represents the natural constant; The expression for the large-size forging forming load prediction model established based on deep neural networks is as follows: , In the formula, W i Let be the weight matrix of the i-th layer; b i Let i be the bias vector of the i-th layer; Substituting the forging process parameters from the standardized sample dataset into the above formula, the expression for the forming load prediction model based on deep neural networks for optimizing the forging process of large-size forgings is as follows: , In the formula, This is the input vector for forging process parameters after standard deviation standardization. This represents the average value of the forming load; This represents the standard deviation of the forming load.

[0030] S08. For the standardized training set, perform multiple independent training sessions for different hyperparameter combinations. After training, calculate the root mean square error and coefficient of determination R² of each candidate model on the validation set. The hyperparameter combination with the smallest root mean square error on the validation set is determined as the optimal parameter configuration of the model. Subsequently, the optimized model is applied to the test set, and the prediction accuracy and generalization ability of the model are further evaluated by the RMSE and R² on the test set. , , In the formula, N The total number of samples, y pred The predicted value of the model. y true For the true value, The average of the true values; root mean square error (RMSE) and coefficient of determination (R²) 2 Quantitative analysis of the prediction results is performed to verify the model's prediction accuracy and generalization ability.

[0031] To further illustrate the technical solution and technical effects of the present invention, the following specific examples are provided: Specific Example 1: A method for predicting forming loads in large-size complex forgings across the entire temperature range based on a deep neural network model, including the following steps: Step 1: Take a plastic billet with a radius of 915mm and a height of 145mm. The forging is symmetrical along the axis of symmetry. To simplify the calculation, half of the billet is selected for simulation. Both the upper and lower dies are defined as rigid bodies. The specific steps for obtaining the numerical simulation values ​​are as follows: The billet was meshed using tetrahedral elements, and a total of about 170,000 elements were divided according to the absolute meshing method. During the simulated forging process, the lower die remained fixed and stationary, while the upper die moved downward at a fixed speed until the preset reduction amount was reached. A full factorial experimental method was used to obtain numerical simulation process schemes for the forging process by processing different combinations of forging processes. The forging process parameters include forging temperatures of 400℃, 425℃, 450℃, 475℃, and 500℃, and a forging speed of 0.1 mm / s. -1 0.2 mm·s -1 1 mm·s -1 2mm·s -1 5mm·s -1 The friction factors are 0.1, 0.3, 0.5, and 0.7, and the deformation amounts are 0.74, 0.78, 0.82, and 0.86.

[0032] Finite element simulation software was used to perform numerical simulation on the aforementioned numerical simulation process scheme, and the simulation results corresponding to each set of forging process conditions were obtained.

[0033] Step 2: Use the post-processing module of the finite element analysis software to extract the load-stroke curve of the upper mold during the pressing process, analyze the changing trend of the curve, and read the forming load under different process parameters. Table 1 shows the forming load values ​​of some forging process parameters corresponding to forgings in specific example 1: Table 1

[0034] Step 3: The forging temperature, friction factor, forging speed, deformation amount and forming load are processed using the standard deviation standardization method to eliminate differences in data magnitude; The standardized formula for calculating the standard deviation of the processing parameters is as follows: , In the formula This represents the average value of the process parameters. This represents the standard deviation of the process parameters.

[0035] Step 4: After standardizing the sample dataset, a prediction model is constructed based on a deep neural network model. This model takes forging temperature, friction factor, forging speed, and deformation as inputs and forming load as output to predict the forming load of the forging. The predicted values ​​of the forming load index of the forging are obtained through the prediction model. Specifically: 1) The standardized sample set obtained in step 3 is randomly divided into training set, validation set and test set according to a preset ratio. The training set accounts for 80% of the total sample size and is used for iterative updates of model weights; the validation set accounts for 10% of the total sample size and is used to monitor model performance and prevent overfitting during training; the test set accounts for 10% of the total sample size and is used to finally evaluate the generalization prediction accuracy of the model. 2) In order to solve the problem that traditional neural network structures rely on human experience to set, this example constructs a hyperparameter search space that includes various hidden layer topologies, training parameters and iteration times.

[0036] The hidden layer topologies include, but are not limited to, various neuron arrangements such as [32, 16, 8], [16, 8, 4], [16, 8], [20, 16, 10], [8, 4, 2], [8, 4], and [16, 16, 16], covering network structures of different depths and widths; The training parameters are set to multiple gradient values, including 0.001, 0.005, and 0.01; The maximum number of iterations is set to multiple gradient values, including 500, 1000, 2000 and 3000 times; 3) Training is performed by searching all hyperparameter combinations in the space using a full combinatorial traversal approach. To improve the model's generalization ability and robustness, the preferred training algorithm is Bayesian regularization. This algorithm modifies the objective function to minimize the mean squared error while constraining the sum of squares of the network weights, thus effectively avoiding overfitting. It is particularly suitable for regression prediction of small samples or high-dimensional data. 4) To eliminate the impact of randomness in the initialization of neural network weights, multiple independent training experiments were conducted for each hyperparameter combination. After training, the root mean square error (RMSE) and coefficient of determination (R²) of each model on the validation set were calculated. 2 .

[0037] The forming load prediction model calculated in step four adopts a 4-8-4-2-1 topology. The input layer has four nodes, corresponding to four key process parameters: forging temperature, friction factor, forging speed, and deformation. The network contains three hidden layers with 8, 4, and 2 neurons respectively. The output layer contains one node, outputting the predicted forming load value. The training parameter for this model is 0.001, and the maximum training epochs are 1000. The optimized weight matrices and bias vectors for each layer are as follows:

[0038]

[0039] The expression for the forming load prediction model is as follows:

[0040] The coefficient of determination of the forming load prediction model on the validation set was 0.996, the root mean square error was 13.84MN, and the mean error was 5.4%; on the test set, the coefficient of determination was 0.992, the root mean square error was 25.03MN, and the mean error was 8.8%.

[0041] Specific Example 2: A method for predicting low-temperature forming loads of large-size complex forgings based on a deep neural network model, which differs from Specific Example 1 in that: The forging process parameters mentioned above include forging temperatures of 400℃, 425℃, and 450℃, and a forging speed of 0.1 mm·s. -1 0.2 mm·s -1 1 mm·s -1 2mm·s -1 5mm·s -1The friction factors are 0.1, 0.3, 0.5, and 0.7, and the deformation amounts are 0.74, 0.78, 0.82, and 0.86.

[0042] Finite element simulation software was used to perform numerical simulation on the aforementioned numerical simulation process scheme, and the simulation results corresponding to each set of forging process conditions were obtained.

[0043] The load-stroke curve of the upper die during the pressing process was extracted using the post-processing module of the finite element analysis software. The trend of the curve was analyzed, and the forming load under different process parameters was read. Table 2 shows the forming load values ​​of the forgings corresponding to some forging process parameters in specific example 2. Table 2

[0044] After standardizing the sample dataset, a prediction model for predicting the forming load of forgings in the low-temperature domain is constructed based on a deep neural network model. This model takes forging temperature, friction factor, forging speed and deformation as inputs and forming load as output.

[0045] The forming load prediction model calculated in step four adopts a 4-8-4-2-1 topology. The input layer has four nodes, corresponding to four key process parameters: forging temperature, friction factor, forging speed, and deformation. The network contains three hidden layers with 8, 4, and 2 neurons respectively. The output layer contains one node, outputting the predicted forming load value. The training parameter for this model is 0.005, and the maximum training epochs are 2000. The optimized weight matrices and bias vectors for each layer are as follows:

[0046]

[0047]

[0048] The expression for the forming load prediction model is as follows:

[0049] The coefficient of determination of the low-temperature forming load prediction model for large-size forgings is 0.994, the root mean square error is 18.01MN, and the mean error is 4.16% on the validation set; the coefficient of determination is 0.976, the root mean square error is 26.41MN, and the mean error is 9.92% on the test set.

[0050] Specific Example 3: A method for predicting high-temperature forming loads of large-size forgings based on a deep neural network model, which differs from Specific Example 1 in that: The forging process parameters include forging temperatures of 450℃, 475℃, and 500℃, and a forging speed of 0.1 mm / s. -1 0.2 mm·s -1 1 mm·s -1 2mm·s -1 5mm·s -1 The friction factors are 0.1, 0.3, 0.5, and 0.7, and the deformation amounts are 0.74, 0.78, 0.82, and 0.86.

[0051] Finite element simulation software was used to perform numerical simulation on the aforementioned numerical simulation process scheme, and the simulation results corresponding to each set of forging process conditions were obtained.

[0052] The load-stroke curve of the upper die during the pressing process was extracted using the post-processing module of the finite element analysis software. The changing trend of the curve was analyzed, and the forming load under different process parameters was read. Table 3 shows the forming load values ​​of the forgings corresponding to some forging process parameters in specific example 3. Table 3

[0053] After standardizing the sample dataset, a prediction model for predicting the forming load of forgings in the high-temperature domain is constructed based on a deep neural network model. This model takes forging temperature, friction factor, forging speed and deformation as inputs and forming load as output.

[0054] The forming load prediction model calculated in step four adopts a 4-8-4-2-1 topology. The input layer has four nodes, corresponding to four key process parameters: forging temperature, friction factor, forging speed, and deformation. The network contains three hidden layers with 8, 4, and 2 neurons respectively. The output layer contains one node, outputting the predicted forming load value. The model's training parameter is 0.01, and the maximum training epochs are 3000. The optimized weight matrices and bias vectors for each layer are as follows:

[0055]

[0056] The expression for the forming load prediction model is as follows:

[0057] The coefficient of determination of the high-temperature forming load prediction model for large-size forgings is 0.999, the root mean square error is 4.64MN, and the mean error is 2.38% on the validation set; the coefficient of determination is 0.989, the root mean square error is 11.95MN, and the mean error is 7.10% on the test set.

[0058] Specific Example 4: A method for predicting forming loads in the high deformation zone of large-size forgings based on a deep neural network model, which differs from Specific Example 1 in that: The forging process parameters include forging temperatures of 400℃, 425℃, 450℃, 475℃, and 500℃, and a forging speed of 0.1 mm / s. -1 0.2 mm·s -1 1 mm·s -1 2mm·s -1 5mm·s -1 The friction factors are 0.1, 0.3, 0.5, and 0.7, and the deformation amounts are 0.82 and 0.86.

[0059] Finite element simulation software was used to perform numerical simulation on the aforementioned numerical simulation process scheme, and the simulation results corresponding to each set of forging process conditions were obtained.

[0060] The load-stroke curve of the upper die during the pressing process was extracted using the post-processing module of the finite element analysis software. The changing trend of the curve was analyzed, and the forming load under different process parameters was read. Table 4 shows the forming load values ​​of the forgings corresponding to some forging process parameters in specific example 4. Table 4

[0061] After standardizing the sample dataset, a prediction model for predicting the forming load of forgings in the high deformation zone is constructed based on a deep neural network model. This model takes forging temperature, friction factor, forging speed and deformation as inputs and forming load as output.

[0062] The forming load prediction model calculated in step four adopts a 4-8-4-2-1 topology. The input layer has four nodes, corresponding to four key process parameters: forging temperature, friction factor, forging speed, and deformation. The network contains three hidden layers with 8, 4, and 2 neurons respectively. The output layer contains one node, outputting the predicted forming load value. The model's training parameter is 0.01, and the maximum training epochs are 3000. The optimized weight matrices and bias vectors for each layer are as follows:

[0063]

[0064]

[0065]

[0066] The expression for the forming load prediction model is as follows: The coefficient of determination of the prediction model for forming load in the high deformation zone of large forgings is 0.994, root mean square error is 13.70MN, and mean error is 2.46% on the validation set; the coefficient of determination of the model is 0.976, root mean square error is 19.17MN, and mean error is 5.34% on the test set.

[0067] Comparative Example 1: The method is the same as in Specific Example 1, except that the model uses default parameters that have not been optimized by hyperparameters.

[0068] The forming load prediction model calculated in step four adopts a 4-8-4-1 topology. The input layer has four nodes, corresponding to four key process parameters: forging temperature, friction factor, forging speed, and deformation. The network contains two hidden layers with 8 and 4 neurons respectively. The output layer contains one node, outputting the predicted forming load value. The model's training parameter is 0.001, and the maximum training epochs are 3000. The optimized weight matrices and bias vectors for each layer are as follows:

[0069]

[0070]

[0071] The expression for the forming load prediction model is as follows:

[0072] The coefficient of determination of the large-size forging forming load prediction model on the validation set is 0.995, the root mean square error is 16.25MN, and the mean error is 6.3%; the coefficient of determination on the test set is 0.9409, the root mean square error is 65.87MN, and the mean error is 15.6%.

[0073] As can be seen from the data in Comparative Example 1, without hyperparameter search optimization, the root mean square error and mean error of the model are significantly increased compared to Specific Example 1. The root mean square error of the validation set increased by 17.4%, and the mean error increased by 16.7%, while the root mean square error of the test set increased by 163.2%, and the mean error increased by 77.3%. This proves that the large-size forging forming load prediction method based on the deep neural network model described in this invention can establish a forming load prediction model with extremely high prediction accuracy, providing an efficient new approach for process prediction of large-size complex components.

[0074] from Figure 1 , Figure 2 As can be seen, under different combinations of forging temperature, friction factor, forging speed, and deformation, the model prediction values ​​in the test set and verification set are in good agreement with the finite element simulation values. Whether in the low load region, medium load region, or high load region, the predicted values ​​can follow the changes in the finite element simulation values ​​well, without significant deviation. This indicates that the forming load prediction model based on deep neural networks established in this invention can effectively reflect the variation law of forming load of large-size forgings under the coupled action of multiple process parameters, has high prediction accuracy and good generalization ability, and can achieve rapid and reliable prediction of forming load of large-size forgings.

[0075] 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, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting forming loads of large-size complex forgings based on a deep neural network model, characterized in that, Includes the following steps: Step 1: Under different forging process parameters, perform finite element simulation on large-size forgings and extract the load-stroke curve of the upper die during the pressing process to obtain the forming load of large-size forgings under different forging process parameters. Among them, the forging process parameters include the forging temperature, friction factor, forging speed and deformation amount of large-size forgings; Large-size forgings are axisymmetric disc-shaped forgings. The diameter of the disc-shaped billet for the axisymmetric disc-shaped forging is 1600mm to 2000mm, the height H is 120mm to 180mm, the target deformation is 0.70 to 0.90, and the section ratio is greater than 9. Forming load refers to the load that a large forging bears in the forming direction when the deformation reaches the target value; Step 2: The sample dataset consisting of forging process parameters and corresponding forming loads is processed using the standard deviation standardization method to eliminate differences in data magnitude and obtain a standardized sample dataset. Step 3: Establish the hyperparameter search space for deep neural network candidate models Ω The model is trained on different candidate hyperparameter combinations in the hyperparameter search space to establish a forging forming load prediction model with forging process parameters as input and forming load as output, so as to realize the prediction of forming load of large-size forgings. Among them, the hyperparameter search space is used to solve the problem that traditional neural network structures rely on manual experience settings; the determination coefficient of the test set of the forging forming load prediction model is not less than 0.97, and the root mean square error is not greater than 30MN; Candidate hyperparameters include a candidate set of hidden layer topologies. H Candidate set of training parameters P and the candidate set of maximum iteration count E .

2. The method for predicting forming loads of large-size complex forgings based on a deep neural network model as described in claim 1, characterized in that, In step one, the forging temperature in the forging process parameters is 350℃~550℃, and the forging speed is 0.1mm·s. -1 ~5mm·s -1 The friction factor is 0.1 to 0.7, and the deformation is 0.7 to 0.

9.

3. The method for predicting forming loads of large-size complex forgings based on a deep neural network model as described in claim 1, characterized in that, In step two, the standard deviation of the sample dataset is calculated using the following formula: ， In the formula, Let be the mean of any set of sample datasets. Let be the standard deviation of any set of sample datasets.

4. The method for predicting forming loads of large-size complex forgings based on a deep neural network model as described in claim 1, characterized in that, In step three, the candidate set of hidden layer topologies H Candidate set of training parameters P and the candidate set of maximum iteration count E The hyperparameter search space Ω formed by them are respectively expressed as: In the formula, H Each element in the table represents a sequence of the number of neurons in a hidden layer; p j This represents the j-th candidate value for the training parameter; e k This represents the k-th candidate value for the maximum number of iterations. Among them, the hidden layer topology is used to limit the number of hidden layers and the number of neurons in each hidden layer, the training parameters are used to adjust the iterative update capability of the Bayesian regularized backpropagation algorithm, and the maximum number of iterations is used to limit the termination condition of model training.

5. The method for predicting forming loads of large-size complex forgings based on a deep neural network model as described in claim 1, characterized in that, In step three, a full combinatorial traversal is used to search for all hyperparameter combinations in the hyperparameter search space. Bayesian regularization is then used to train the model, resulting in the modified objective function: ， In the formula, n This represents the number of training samples; y i For the first i The true value of each sample; These are the model's predicted values; w j These are the weight parameters in the network; m This represents the total number of weight parameters in the network. α and β These are performance parameters of the objective function, used to automatically adjust and balance the proportion of mean squared error and weighted sum of squares in the optimization process; By comparing the root mean square error values ​​of the validation set under all combinations and experimental sequences, the model with the smallest root mean square error value is selected as the final forging forming load prediction model, and its network weights and bias matrices are saved. Among them, the Bayesian regularization algorithm effectively avoids overfitting by modifying the objective function to minimize the mean square error while constraining the sum of squares of the network weights, and is especially suitable for regression prediction of small samples or high-dimensional data.

6. The method for predicting forming loads of large-size complex forgings based on a deep neural network model as described in claim 1, characterized in that, In step three, the forging forming load prediction model is a feedforward multilayer perceptron model, which includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer corresponds to the standardized forging temperature, friction factor, forging speed, and deformation process parameters; the nodes in the output layer correspond to the forming load prediction target. Hidden layers include L Hidden layers L The number of neurons in each hidden layer is an integer greater than or equal to 2, determined by the candidate set of hidden layer topologies. H The sequence of the number of neurons in any hidden layer is determined; a candidate set of hidden layer topologies is established. H This includes hidden layer sequences with decreasing width and hidden layer sequences with equal width; Training parameters are derived from the candidate set of training parameters. P The maximum number of iterations is arbitrarily selected from the candidate set of maximum iterations. E Choose any one from the options.

7. The method for predicting forming loads of large-size complex forgings based on a deep neural network model as described in claim 1, characterized in that, The hyperbolic tangent function is selected as the activation function for the forging forming load prediction model. The expression of the activation function is as follows: ， In the formula, x is the input value of the forging forming load prediction model; e represents the natural constant; The expression for the large-size forging forming load prediction model established based on deep neural networks is as follows: ， In the formula, W i Let be the weight matrix of the i-th layer; b i Let i be the bias vector of the i-th layer; Substituting the forging process parameters from the standardized sample dataset into the above formula, the expression for the forming load prediction model based on deep neural networks for optimizing the forging process of large-size forgings is as follows: ， In the formula, This is the input vector for forging process parameters after standard deviation standardization. This represents the average value of the forming load; This represents the standard deviation of the forming load.

8. The method for predicting forming loads of large-size complex forgings based on a deep neural network model as described in claim 1, characterized in that, The standardized sample dataset is randomly divided into a training set, a validation set, and a test set according to a preset ratio. The training set is used for iterative updates of the weights of the forging forming load prediction model; the validation set is used to monitor the performance of the forging forming load prediction model and prevent overfitting during the training process; and the test set is used to finally evaluate the generalization prediction accuracy of the forging forming load prediction model.

9. The method for predicting forming loads of large-size complex forgings based on a deep neural network model as described in claim 4, characterized in that, For each of the different hyperparameter combinations, multiple independent training experiments were conducted. After training, the root mean square error (RMSE) and coefficient of determination (R²) of each model on the validation set were calculated. 2 for: ， ， In the formula, N The total number of samples, y pred The predicted value of the model. y true For the true value, The average of the true values; root mean square error (RMSE) and coefficient of determination (R²) 2 Quantitative analysis of the prediction results is performed to verify the model's prediction accuracy and generalization ability.

10. The method for predicting forming loads of large-size complex forgings based on a deep neural network model as described in claim 1, characterized in that, The load-stroke curve of the upper die during the pressing process in step one is extracted by the post-processing module of the finite element analysis software; at the same time, the finite element simulation includes establishing the geometric model of the large-size forging die and the mesh generation of the billet numerical simulation in the finite element simulation software.

Images