Low alloy steel flow stress prediction method and system based on fusion constitutive parameters and GA-BP neural network

Abstract

The invention provides a low alloy steel flow stress prediction method and system based on fusion of constitutive parameters and a GA-BP neural network. The method comprises the following steps: firstly, acquiring experimental data of the low alloy steel under different thermal deformation conditions and preprocessing the experimental data; then, based on an Arrhenius constitutive relation model and a dynamic material model theory, constitutive characteristic parameters of the low alloy steel under different thermal deformation conditions are obtained through calculation; combining the preprocessed experimental data with the constitutive characteristic parameters obtained by calculation, and constructing a training data set containing intrinsic parameters, process parameters and constitutive characteristic parameters; using the training data set to train a GA-BP neural network model of a back propagation neural network optimized by a genetic algorithm to obtain a flow stress prediction model; and finally, inputting parameters under a to-be-predicted working condition into the flow stress prediction model, and outputting to obtain a predicted flow stress value.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of materials processing engineering and computational intelligence, specifically relating to a method and system for predicting the rheological stress of low alloy steel based on the fusion of constitutive parameters and GA-BP neural network, which is used to predict the rheological behavior of low alloy steel during hot working with high accuracy. Background Technology

[0002] In the hot forming process of materials, various metallurgical phenomena with mutual coupling and synergistic effects are involved, mainly including work hardening, dynamic recovery, and dynamic recrystallization. These hardening and softening mechanisms are dynamically competitive and are significantly influenced by deformation temperature and strain rate. For example, high temperatures promote thermal activation processes, which are beneficial to softening mechanisms; high strain rates enhance the work hardening effect. Therefore, the flow behavior of materials is inherently quite complex. Svante Arrhenius established the Arrhenius equation in the 19th century, providing the core theoretical foundation for describing thermal activation processes. Subsequent researchers applied this theoretical framework to establish a class of material constitutive models, which can successfully predict the rheological stress behavior of materials at high temperatures by inputting parameters such as strain rate and temperature. Researchers have also established the Johnson-Cook (JC) empirical constitutive model and the Zerilli-Armstrong (ZA) constitutive model, which successfully describe the dynamic mechanical response of metallic materials under large deformation, high strain rate, and high temperature conditions by introducing the product term of strain, strain rate, and temperature. However, due to the low accuracy and poor flexibility of the constitutive equations, they are not effective in characterizing complex constitutive relationships.

[0003] With the rapid development of artificial intelligence technology and the exponential growth of materials data, machine learning is becoming the primary method for revealing the complex relationships between material characteristics and related properties. In recent years, artificial neural networks have been widely used to predict the rheological behavior of metallic materials. Lin Yongcheng et al. from Central South University conducted hot compression experiments on 42CrMo steel and constructed a predictive model for rheological stress using a feedforward BP neural network. Ahmadi et al. from Islamic Azad University compared and analyzed the predictive performance of the JC phenomenological model, the ZA physical model, and the BP neural network model through hot compression experiments, finding that the BP neural network model has a significant advantage in characterizing the complex rheological behavior of this material. Although the above methods have made progress in characterizing the hot deformation behavior of single steel grades, the transfer and generalization ability between different types of steel is fundamentally limited because the experimental data relied upon for model training usually comes from the same steel grade.

[0004] Therefore, how to achieve high-precision prediction of rheological stress in multiple steel grades is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method and system for predicting the rheological stress of low alloy steel based on the fusion of constitutive parameters and GA-BP neural network. The aim is to improve the calculation efficiency of the constitutive equation and hot working diagram parameters of low alloy steel, and to predict the rheological stress value of low alloy steel under different working conditions with high accuracy.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for predicting the rheological stress of low-alloy steel based on the fusion of constitutive parameters and GA-BP neural network, the method comprising: The raw data of hot compression experiments on low alloy steel under different hot deformation conditions were obtained and preprocessed to obtain a standardized dataset. Based on standardized datasets, combined with the Arrhenius constitutive relation model and DMM theory, machine learning algorithms were used to obtain constitutive characteristic parameters of low alloy steel under different hot deformation conditions. A training dataset is constructed based on a standardized dataset and constitutive feature parameters; A GA-BP neural network model was constructed by optimizing the backpropagation neural network using a genetic algorithm and trained on a training dataset to obtain a rheological stress prediction model. Based on the rheological stress prediction model, the rheological stress of the low alloy steel to be predicted is predicted.

[0007] Preferred methods for obtaining constitutive characteristic parameters of low-alloy steel under different hot deformation conditions, based on standardized datasets, combined with the Arrhenius constitutive relation model and DMM theory, and employing machine learning algorithms, include: For each strain condition, the standardized dataset is grouped according to the deformation temperature; Based on the grouped data, calculate the parameters included in the Arrhenius constitutive relation model; Based on the grouped data, a heat treatment map is constructed using DMM theory combined with machine learning algorithms, and the various parameters contained in the heat treatment map are calculated.

[0008] Preferred methods for calculating the parameters included in the Arrhenius constitutive relation model based on grouped data include: Calculate the temperature-compensated deformation rate factor Z parameter: ; in, For strain rate, Q The activation energy for thermal deformation, R The molar gas constant, T This refers to absolute temperature.

[0009] Preferably, based on the grouped data, a method for quickly constructing a heat treatment map using DMM theory combined with machine learning algorithms, and calculating the various parameters included in the heat treatment map, includes: The strain rate sensitivity index was calculated using the finite difference method. m : ; in, For dissipative covariance, G Power dissipation, For rheological stress; Power dissipation factor or : ; in, This represents the maximum value of the dissipation covariance. Instability Criteria g : .

[0010] Preferably, the method for constructing a GA-BP neural network model by optimizing the backpropagation neural network based on a genetic algorithm and training it on a training dataset to obtain a rheological stress prediction model includes: Building a basic network architecture based on the MLP model; The optimal combination of neurons in the hidden layer of the basic network architecture is obtained through Bayesian optimization, thus achieving the optimal network topology. A genetic algorithm is used to globally optimize the initial weights, thresholds, and key hyperparameters of the optimal network topology to obtain the optimized network architecture. The optimized network architecture was trained using the training dataset to obtain a rheological stress prediction model.

[0011] The present invention also provides a low-alloy steel rheological stress prediction system based on the fusion of constitutive parameters and GA-BP neural network. The system is used to implement the aforementioned method and includes: a standardized dataset acquisition module, a constitutive feature parameter acquisition module, a training dataset acquisition module, a model building module, and a prediction module. The standardized dataset acquisition module is used to acquire and preprocess the raw data of hot compression experiments of low alloy steel under different hot deformation conditions to obtain a standardized dataset. The constitutive feature parameter acquisition module is used to obtain the constitutive feature parameters of low alloy steel under different hot deformation conditions based on a standardized dataset, combined with the Arrhenius constitutive relation model and DMM theory, and using machine learning algorithms. The training dataset acquisition module is used to construct a training dataset based on the standardized dataset and constitutive feature parameters; The model building module is used to optimize the backpropagation neural network based on the genetic algorithm, build the GA-BP neural network model, and train it based on the training dataset to obtain the rheological stress prediction model. The prediction module is used to predict the flow stress of low alloy steel based on the flow stress prediction model.

[0012] Preferably, the constitutive feature parameter acquisition module includes: a grouping unit, a first parameter calculation unit, and a second parameter calculation unit; Grouping units are used to group the standardized dataset by deformation temperature for each strain condition; The first parameter calculation unit is used to calculate the parameters contained in the Arrhenius constitutive relation model based on the grouped data. The second parameter calculation unit is used to construct a heat treatment map based on the grouped data, using DMM theory combined with machine learning algorithms, and to calculate the various parameters contained in the heat treatment map.

[0013] Preferred methods for calculating the parameters included in the Arrhenius constitutive relation model based on grouped data include: Calculate the temperature-compensated deformation rate factor Z parameter: ; in, For strain rate, Q The activation energy for thermal deformation, R The molar gas constant, T This refers to absolute temperature.

[0014] Preferably, based on the grouped data, a method for quickly constructing a heat treatment map using DMM theory combined with machine learning algorithms, and calculating the various parameters included in the heat treatment map, includes: The strain rate sensitivity index was calculated using the finite difference method. m : ; in, For dissipative covariance, G Power dissipation, For rheological stress; Power dissipation factor or : ; in, This represents the maximum value of the dissipation covariance. Instability Criteria g : .

[0015] Preferably, the model building module includes: a basic architecture unit, a first optimization unit, a second optimization unit, and a training unit; Infrastructure unit, used to build basic network architecture based on MLP model; The first optimization unit is used to obtain the optimal combination of the number of neurons in the hidden layer of the basic network architecture through Bayesian optimization, and to obtain the optimal network topology. The second optimization unit is used to globally optimize the initial weights, thresholds and key hyperparameters of the optimal network topology using a genetic algorithm to obtain the optimized network architecture. The training unit is used to train the optimized network architecture using the training dataset to obtain a rheological stress prediction model.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: The core innovation of this invention lies in its modeling strategy that integrates physical mechanisms with data-driven approaches. By using constitutive feature parameters along with conventional chemical composition and process parameters as inputs to a neural network, the data-driven prediction model is guided by physical rules, enabling it to not only learn data representations but also understand the underlying physical mechanisms of thermal deformation. Simultaneously, a two-level intelligent optimization strategy combining Bayesian optimization and genetic algorithms is employed to systematically optimize the structure and hyperparameters of the neural network, overcoming the blindness of traditional manual parameter tuning. Furthermore, L2 regularization, Dropout, and early stopping strategies are integrated into the prediction model construction to effectively prevent overfitting and ensure the model's generalization ability. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the flow chart of the low alloy steel rheological stress prediction method based on the fusion of constitutive parameters and GA-BP neural network in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the process of obtaining a standardized dataset according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the process of obtaining constitutive feature parameters according to an embodiment of the present invention; Figure 4 The figures show scatter plots and their fitted lines under various strain conditions according to embodiments of the present invention, where (a)-(h) represent strains of 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7 and peak strain, respectively. - s Relationship curve; Figure 5 This is a schematic diagram of the GA-BP network model and rheological stress prediction process in an embodiment of the present invention; Figure 6 This is a schematic diagram of the predicted rheological stress of low alloy steel according to an embodiment of the present invention, wherein (a)-(d) are schematic diagrams of the predicted rheological stress at 0.01s-1, 0.1s-1, 1s-1 and 10s-1, respectively. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] As can be seen from the background technology, existing models for predicting the rheological stress of low alloy steel mainly face the following bottlenecks: traditional constitutive equations have limited accuracy and flexibility in characterizing complex hot deformation behavior, making it difficult to accurately describe the rheological response of multiple steel grades under different working conditions; at the same time, most current machine learning models only use deformation temperature, strain rate and strain as input descriptors, lacking in-depth integration of material physical mechanisms, resulting in the models exhibiting obvious "black box" properties, limited predictive ability and poor generalization performance.

[0022] The aforementioned problems severely limit the application value of the model in real-world industrial scenarios. Due to the lack of a coupled expression of the intrinsic properties of materials and the physical mechanisms of deformation, the model struggles to capture the influence of alloying elements and process parameters on flow stress, resulting in insufficient reliability of its predictions across compositional and process conditions. In particular, for materials like low-alloy steel, which are sensitive to composition-process-performance relationships, the prediction bias of existing models may further amplify process design errors, affecting the microstructure and performance control of the final product.

[0023] Example 1 This invention provides a method for predicting the rheological stress of low-alloy steel based on the fusion of constitutive parameters and GA-BP neural network, comprising: The raw data of hot compression experiments on low alloy steel under different hot deformation conditions were obtained and preprocessed to obtain a standardized dataset. Based on standardized datasets, combined with the Arrhenius constitutive relation model and DMM theory, machine learning algorithms were used to obtain constitutive characteristic parameters of low alloy steel under different hot deformation conditions. A training dataset is constructed based on a standardized dataset and constitutive feature parameters; A GA-BP neural network model was constructed by optimizing the backpropagation neural network using a genetic algorithm and trained on a training dataset to obtain a rheological stress prediction model. Based on the rheological stress prediction model, the rheological stress of the low alloy steel to be predicted is predicted.

[0024] like Figure 1 As shown, the specific implementation process of the present invention is as follows: Step 1: Data Acquisition and Preprocessing Raw data of hot compression experiments on low alloy steel under different hot deformation conditions were obtained. The raw data of hot compression experiments included process parameters (deformation temperature, strain rate, strain and corresponding rheological stress values) and intrinsic material parameters (chemical composition). The raw data of hot compression experiments were preprocessed to form a standardized dataset for subsequent analysis.

[0025] like Figure 2 As shown, the present invention can also conduct hot deformation experiments on five types of low-alloy test steels to obtain corresponding experimental data; on the other hand, by consulting literature, hot deformation experimental data of different types of low-alloy test steels can be collected, and then these data can be integrated to obtain raw data of hot compression experiments. Then, the raw data of hot compression experiments can be preprocessed to obtain a standardized dataset.

[0026] Step 2: Calculation of constitutive model and thermal processing diagram parameters Based on standardized datasets, and combining the Arrhenius constitutive relation model and DMM theory, machine learning algorithms are used to obtain constitutive characteristic parameters of low-alloy steel under different hot deformation conditions. The methods include: For each strain condition, the standardized dataset is grouped according to the deformation temperature; Based on the grouped data, calculate the parameters included in the Arrhenius constitutive relation model; Based on the grouped data, a heat treatment map is constructed using DMM theory combined with machine learning algorithms, and the various parameters contained in the heat treatment map are calculated.

[0027] Specifically: Based on the Arrhenius constitutive relation model and the Dynamic Material Model (DMM) theory, computational analysis is performed on a standardized dataset to solve for key constitutive characteristic parameters, including the thermal deformation activation energy. Q Zener-Hollomon parametersZ natural logarithm lnZ Power dissipation efficiency or By using numerical difference method and linear regression fitting, the constitutive equation parameters and hot working diagram parameters of low alloy steel under different strain conditions were obtained.

[0028] 1) Calculate the parameters included in the Arrhenius constitutive relation model. The process parameters (deformation temperature, strain rate, strain and corresponding rheological stress values) in the raw data of hot compression experiments obtained from literature and experiments are organized into an Excel spreadsheet through the program interface (Python's pandas library). The raw data of hot compression experiments of low alloy steel and its corresponding values ​​are then read from the Excel spreadsheet. For the raw data of the hot compression experiment to be processed, specify each strain condition, including 0.1-0.7 (interval of 0.1) strain and peak strain, and group the data according to the deformation temperature using the following formulas (1)-(4), and ensure that each temperature group contains at least two data points at different strain rates to meet the basic conditions for subsequent linear regression analysis. To solve for material constants β Calculate the regression results for each temperature group β The coefficient of determination R of the value 2 (Parameters used to evaluate the goodness of fit to the data, not constitutive equation parameters), to assess the quality of linear fit. Then, R is selected. 2 The highest front k Temperature groups (of which) k = Total number of temperature groups - 1), calculate its β Weighted average of values ​​(weights can be based on R) 2 (or the number of data points), and will also calculate all temperature groups under that strain. β The arithmetic mean of the values, used as a comparison and alternative; Adopting the above β Similar modeling and regression strategies are used to calculate the stress index sequentially. n 1 and n Stress level parameters ( ), thermal deformation activation energy Q And the intercept lnA; All calculated parameters, including those for each temperature group β , n value, R 2 and the optimized average β ,average n 1. Average n ,average α , QThe values ​​of Z, lnA, and Z are automatically saved to a specified Excel file, achieving structured storage of the calculation results. The constitutive equation parameters include: β Stress level parameters Stress index n and n 1. Activation energy of thermal deformation Q The natural logarithm of the temperature-compensated strain rate factor lnZ .

[0029] Furthermore, the Arrhenius constitutive relation model can accurately describe the relationship between flow stress, deformation temperature, and strain rate during high-temperature deformation of materials. Equations (1)-(3) represent different stress level conditions: (1) (2) (3) In the formula, S is the strain rate, s -1 ; T The absolute temperature (deformation temperature), K; R The molar gas constant is 8.314 J / (mol·K); Rheological stress, MPa; Q The activation energy for thermal deformation is kJ / mol. A , A 1. A 2. n 1. n , and These are all material constants. In fact, the activation energy of thermal deformation... Q It is an important parameter characterizing atomic diffusion ability. The greater the activation energy, the greater the hindrance to atomic diffusion, and the more difficult diffusion is to proceed. In addition, the Zener-Hollomon parameter, as a temperature-compensated deformation rate factor, can be used to represent the influence of deformation temperature and strain rate on the rheological stress curve. This is of great significance for the formulation of hot working process conditions. Its expression is shown in formula (4): (4) From formula (4), we can see that, Z The parameter values ​​are closely related to the thermal deformation parameters of the specimen. The higher the deformation temperature, the better. Z The smaller the parameter, the higher the strain rate. Z The larger the parameter, the better. From the relationship between deformation temperature, strain rate, and dynamic recrystallization, it can be seen that softening behaviors such as dynamic recrystallization and dynamic recovery typically occur at low temperatures. Z Under parameter values, while when ZWhen the parameter values ​​are high, the hot deformation process is dominated by work hardening.

[0030] To calculate 1w of steel β Taking the values ​​as an example, the Arrhenius constitutive relation model first reads the raw experimental data from Excel, extracting the specific values ​​of deformation temperature, strain rate, and rheological stress. Then, for each strain condition, it groups the data by temperature, ensuring that each group contains at least two data points (to meet the basic conditions for linear regression). Next, linear regression is used to perform least squares analysis on the data within each temperature group, combined with the Arrhenius constitutive relation. Finally, the coefficient of determination R for each temperature group is calculated. 2 To evaluate the fit quality of the linear model and select R0 2 The highest front k Temperature groups ( k = n -1, where n (Total number of temperature groups), calculate its β The weighted average of the values ​​was calculated, and all temperature groups under this strain were also calculated. β The arithmetic mean of the values. To visually verify the reliability of the linear regression analysis, multiple scatter plots under different strain conditions were generated and their fitted lines were compared, as shown below. Figure 4 As shown, Figure 4 The paper presents the linear regression results for strains of 0.1–0.7 (intervals of 0.1) and peak strain, and saves the detailed calculation results, including those for each temperature group. β value, R 2 Intercept and optimized average β value.

[0031] In addition, the stress index was successfully obtained using the same method. n 1 and n Activation energy of thermal deformation Q Values ​​and test steel lnA Value. Based on β and n 1. The calculation results of these two key parameters are based on The system automatically calculates and saves data at different strain levels. Values ​​are transferred to Excel.

[0032] 2) Calculate the parameters included in the heat treatment drawing based on DMM theory. For the raw data to be processed, specify each strain condition, and the program will group the data by strain and temperature, ensuring that each temperature group contains at least two data points at different strain rates to meet the basic conditions for subsequent linear regression analysis. For data under each strain and temperature condition, the program uses the numerical difference method to calculate the strain rate sensitivity index (SFR).m For points in the middle of the data sequence, the central difference method is used to improve accuracy; for points at the beginning and end of the data sequence, the forward difference method and the backward difference method are used respectively. The above calculations yielded m The value is used to further deduce the power dissipation efficiency. or and rheological instability parameter ζ; All calculated raw values ​​and interpolation results, including m , or , g All of them were automatically saved to the corresponding Excel files. m , or , g These are the parameters of the heat treatment diagram.

[0033] Furthermore, the thermal processing diagram is rapidly constructed using DMM theory combined with machine learning algorithms. Based on this theory, the thermal deformation process of metallic materials is considered an irreversible process far removed from thermodynamic equilibrium. Under isothermal and constant strain conditions, the dynamic response of the material to the strain rate can be expressed by formula (5): (5) in, K It is a material constant that depends on the conditions. m is the strain rate sensitivity coefficient.

[0034] According to the principle of energy conservation, the total power P absorbed by the material during deformation can be decomposed into the energy consumed by the material's plastic deformation and the energy consumed by the evolution of the microstructure, as shown in formula (6): (6) in, P The total power absorbed by the material per unit volume, representing the power (W) absorbed by the deformable material; G Power dissipation refers to the energy (W) consumed by a material during plastic deformation per unit time. J Let W be the dissipative covariance.

[0035] m The strain rate sensitivity index represents the ratio of energy consumed by tissue evolution to energy consumed by plastic deformation, as shown in formula (7): (7) in, For dissipative covariance, G Power dissipation, For rheological stress; Power dissipation factor ( or Formula (8) can reflect the power dissipation characteristics of materials under different deformation conditions. orThe expression: (8) in, This represents the maximum value of the dissipation covariance. Due to softening behaviors such as dynamic recovery and recrystallization within metallic materials, a significant amount of energy can be released. To effectively determine the optimal processing region, analysis is needed based on the power dissipation diagram and the processing instability diagram. This invention applies the Prasad instability criterion, established based on Ziegler rheological theory, to plot the instability diagram. g The calculation method is shown in formula (9): (9) According to the above formulas (5)-(9), the model for quickly calculating the parameters of the hot working diagram first reads the original experimental data through Excel, extracts the specific values ​​of deformation temperature, strain rate, and rheological stress, and then groups them according to temperature for each strain condition, ensuring that each group contains at least two data points. Then, the strain rate sensitivity index is calculated using the difference method (forward / backward difference for the first and last points, and central difference for the middle points). m ), and based on this, we can deduce that or and g Secondly, cubic spline interpolation is used to construct... or and g The continuous distribution field is presented as a 2D surface map. or Spatial distribution (color gradient table dissipation efficiency), marked with gray areas. g Add key contour lines to the unstable areas where the value is <0. Finally, m , or , g The original values ​​and calculation results are saved to the corresponding Excel file.

[0036] like Figure 3 As shown, step 2 may further include: explicitly adopting the Arrhenius constitutive relation model, and combining it with the Dynamic Material Model (DMM) theory as a foundation; inputting the raw data of rheological stress at different temperatures, strain rates, and corresponding strains (such as 0.1Y, 0.3Y, and peak strain PeakY) into a table according to the prescribed format to prepare for subsequent parameter calculations; and calculating the constitutive characteristic parameters of low-alloy steel under different hot deformation conditions, including the hot deformation activation energy, based on the above model, theory, and input raw data. Q The natural logarithm of the temperature-compensated strain rate factor lnZ Power dissipation efficiency or .

[0037] Step 3: Constitutive training dataset The pre-processed process parameters (deformation temperature, strain rate, strain) and intrinsic material parameters (chemical composition) from step 1 are compared with the constitutive characteristic parameters calculated in step 2. or , Q , lnZ The dataset is fused (the fusion method is a simple permutation; the original dataset only has process parameters and material intrinsic parameters, and the fusion adds another set of features to the model, namely constitutive feature parameters), to construct a dataset for training the machine learning model; the input features include chemical composition, deformation temperature, strain rate, strain, and constitutive feature parameters. or , Q , lnZ The output target is the corresponding rheological stress value.

[0038] Step 4: Training the GA-BP Neural Network Model The methods for obtaining a rheological stress prediction model by optimizing a backpropagation neural network using a genetic algorithm, constructing a GA-BP neural network model, and training it on a training dataset include: Building a basic network architecture based on the MLP model; The optimal combination of neurons in the hidden layer of the basic network architecture is obtained through Bayesian optimization, thus achieving the optimal network topology. A genetic algorithm is used to globally optimize the initial weights, thresholds, and key hyperparameters of the optimal network topology to obtain the optimized network architecture. The optimized network architecture was trained using the training dataset to obtain a rheological stress prediction model.

[0039] Based on the training dataset constructed in step 3, the optimal structure for the number of neurons in the hidden layer of the network is further determined through Bayesian optimization. A genetic algorithm is then used to globally optimize the initial weights, thresholds, and key hyperparameters of the BP neural network, training a GA-BP neural network rheological stress prediction model that integrates physical mechanisms and data-driven approaches. The schematic diagram of the GA-BP neural network structure and the model expression are shown below. Figure 5 As shown, specifically: This invention designs two MLP models with different complexities for comparative analysis. Model 1: Constructs an MLP model with a three-layer hidden layer structure. Model 1 adopts a descending hierarchical architecture, with the number of neurons in each hidden layer decreasing sequentially, and uniformly uses the ReLU activation function (f(x) = max(0, x)) to enhance nonlinear expressive power. To avoid overfitting, L2 regularization (weight decay coefficient λ=0.01) and a 30% Dropout mechanism are introduced in the first two hidden layers. The Adam optimization algorithm (learning rate η=0.001~0.003) is used for model training, and an early stopping strategy is implemented, automatically terminating training when the validation loss does not decrease for 20 consecutive rounds. The model is trained and evaluated using 5-fold cross-validation, and the mean absolute error (MAE) and coefficient of determination (R²) of the test set are recorded. 2 To further verify the impact of model complexity on performance, an enhanced MLP model (Model 2) was subsequently constructed. While maintaining the original three-layer structure, the number of neurons in each hidden layer was moderately increased. Experimental results show that Model 2 achieved better performance on the task, with a reduced average MAE and R... 2 The corresponding improvements will be made, using Model 2 as the basic network architecture.

[0040] After determining the basic network architecture, an automatic tuning method for the number of hidden layer neurons based on Bayesian optimization was adopted. This method included: constructing a Gaussian process surrogate model and combining it with the expectation improvement (EI) acquisition function to optimize parameters within a set search range (128-320 for the first layer, 64-224 for the second layer, and 32-128 for the third layer); using the mean square error of the validation set as the objective function, the optimal combination of neurons was determined after multiple iterations, and this optimal network structure was saved as a fixed topology for subsequent hyperparameter optimization, thus obtaining the optimal network topology.

[0041] After determining the optimal network topology, a genetic algorithm is used to systematically optimize its hyperparameters. This includes: using a hybrid encoding scheme to represent the hyperparameters, where the learning rate and L2 regularization coefficients are encoded using logarithmic scaling (10⁻⁴ to 10⁻²), the Dropout ratio uses linear encoding (0.1 to 0.4), and the batch size uses discrete integer value encoding (16 / 32 / 64 / 128); initializing a population of size 20, with a validation set R... 2 As a fitness evaluation metric, the fitness of each individual is assessed through the complete training process; tournament selection, two-point crossover and protection mechanisms are implemented to ensure parameter validity; a differentiated mutation strategy is adopted, Gaussian mutation is applied to continuous parameters and random reset mutation is applied to discrete parameters, and mutation probabilities are set; after multiple iterations of evolution, the historical best individuals are retained from the population to obtain the optimal hyperparameter combination and the optimized network architecture.

[0042] After determining the optimal hyperparameter combination, the optimized network architecture is trained and output to predict rheological stress. This also includes: retraining the GA-BP neural network model using the optimized network structure and the optimal hyperparameter combination. The trained GA-BP neural network model is the rheological stress prediction model. The optimal model parameters, architecture information and performance indicators after training are saved as a deployable model file, and the model training process curve and performance evaluation report are generated to provide a reliable prediction tool for rheological stress prediction.

[0043] Step 5: Rheological stress prediction The chemical composition, target process parameters, and intrinsic parameters of the low-alloy steel workpiece to be predicted are input into the trained GA-BP neural network model. The model automatically outputs the predicted rheological stress value under the corresponding working condition, providing core data support for the accurate design and intelligent control of the hot working process of low-alloy steel.

[0044] like Figure 5 As shown, process parameters (including deformation temperature, strain rate, and strain), chemical composition (12 types: C, Si, Mn, Cr, Al, Mo, Ni, V, B, Nb, P, and S), and constitutive parameters (including the hot deformation activation energy Q, the natural logarithm of the temperature-compensated strain rate factor lnZ, and the power dissipation efficiency η) are used as the input layer of the rheological stress prediction model. The rheological stress values ​​corresponding to each strain are set as the output layer. The remaining layers are used as hidden layers to mine and learn the complex nonlinear relationships between the input data, thereby capturing the inherent laws of the data and predicting the rheological stress of low-alloy steel. The rheological stress prediction results are shown below. Figure 6 As shown.

[0045] In summary, this invention provides a method for predicting the rheological stress of low-alloy steel based on the fusion of constitutive parameters and a GA-BP neural network: Experimental data of low-alloy steel under different hot deformation conditions are acquired and preprocessed; constitutive characteristic parameters of low-alloy steel are calculated based on the Arrhenius constitutive relation model and dynamic material model theory; a training dataset integrating intrinsic parameters, process parameters, and constitutive characteristic parameters is constructed; and a GA-BP neural network optimized by a genetic algorithm is used for model training to obtain a high-precision rheological stress prediction model.

[0046] This invention introduces constitutive characteristic parameters as physical guidance, enhancing the constraints of metallurgical mechanisms on the predictive model based on data-driven principles and significantly improving its ability to describe complex hot deformation behavior. Furthermore, through a system optimization strategy combining genetic algorithms and Bayesian optimization, the predictive model achieves global optimization at the structural and hyperparameter levels, effectively suppressing overfitting and enhancing generalization performance. This invention not only significantly improves the prediction accuracy of rheological stress but also overcomes the limitations of traditional "black box" models, providing reliable technical support for the precise design and intelligent control of hot working processes for low-alloy steel.

[0047] Example 2 This embodiment illustrates the implementation of the method described in the foregoing embodiments using specific experimental data: Experimental data 1: The five low-alloy experimental steels have the following main chemical components by mass percentage: C: 0.24, Si: 1.44, Mn: 2.1, Cr: 0.91, Al: 0.34, the remainder being Fe; C: 0.41, Si: 1.43, Mn: 1.73, Cr: 0.95, Al: 0.46, the remainder being Fe; C: 0.44, Si: 0.5, Mn: 1.66, Cr: 0.94, Al: 0.42, the remainder being Fe; C: 0.75, Si: 1.54, Mn: 1.6, Cr: 0.95, Al: 0.39, the remainder being Fe; C: 0.24, Si: 1.44, Mn: 2.1, Cr: 0.91, Al: 0.47, the remainder being Fe; The above-mentioned test steel was subjected to single-pass compression tests at varying temperatures and strain rates using a Gleeble-3500 thermal simulation testing machine. The specimens were heated to 1150 °C at a rate of 10 °C / s and held for 300 s to obtain a uniform austenitic structure. Subsequently, they were cooled to various deformation temperatures (850–1150 °C, in 50 °C intervals) at a rate of 5 °C / s and held for 30 s, followed by a decrease in temperature at 0.1 s. -1 1 s -1 10 s -1The strain was compressed at a rate of 50% until the true strain reached 0.7, and then immediately water-quenched. Rheological stress curves were obtained at 850–1150 °C. Then, based on steps 1 and 2, parameter values ​​for strains of 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, and 0.7 were extracted. A genetic algorithm was used to globally optimize the initial weights, thresholds, and key hyperparameters of the BP neural network. Bayesian optimization was then used to further determine the optimal structure for the number of neurons in the hidden layers of the network, training a GA-BP neural network rheological stress prediction model that integrates physical mechanisms and data-driven approaches. Process parameters (including deformation temperature, strain rate, and strain), chemical compositions (12 in total), and constitutive parameters (including the thermal deformation activation energy Q, the natural logarithm of the temperature-compensated strain rate factor lnZ, and the power dissipation efficiency η) were used as the input layer of the model, and the corresponding rheological stress values ​​at each strain were set as the output layer. The remaining layers, acting as hidden layers, mine and learn the complex nonlinear relationships between the input data to capture the inherent patterns in the data and predict the rheological stress of low-alloy steel. The model's prediction accuracy has a coefficient of determination R0. 2 The strength reached 0.945, and the MAE was 10.531 MPa.

[0048] Experimental data 2: The four low-alloy experimental steels, with the following mass percentages of main chemical components in their matrix: C: 0.2, Si: 1.33, Mn: 1.63, Cr: 0.97, Al: 0.02, Mo: 0.19, Ni: 0.17, V: 0.086, P: 0.003, S: 0.004, with the remainder being Fe; C: 0.2, Si: 1.31, Mn: 1.61, Cr: 0.95, Al: 0.24, Mo: 0.19, Ni: 0.17, V: 0.085, P: 0.003, S: 0.004, with the remainder being Fe; C: 0.19, Si: 1.31, Mn: 1.6, Cr: 0.95, Al: 0.46, Mo: 0.19, Ni: 0.17, V: 0.086, P: 0.003, S: 0.004, with the remainder being Fe; C: 0.21, Si: 1.33, Mn: 1.62, Cr: 0.96, Al: 0.67, Mo: 0.19, Ni: 0.17, V: 0.085, P: 0.003, S: 0.004, with the remainder being Fe; The experimental steel used was based on data collected from literature. The process involved single-pass compression tests at varying temperatures and strain rates using a Gleeble-3800 thermal simulation testing machine. The specimens were heated to 1100 °C at a rate of 10 °C / s and held for 300 s to obtain a uniform austenitic structure. Subsequently, the temperature was lowered to various deformation temperatures (850–1100 °C, at 50 °C intervals) at a rate of 5 °C / s and held for 30 s, followed by a decrease in temperature at 0.01 s. -1 0.1 s -1 1 s -1 10 s -1 The strain was compressed at a rate of 50% until the true strain reached 0.7, and immediately water-quenched after deformation. Rheological stress curves were obtained at 850–1100 °C. Then, based on steps 1 and 2, parameter values ​​for strains of 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, and 0.7 were extracted. A genetic algorithm was used to globally optimize the initial weights, thresholds, and key hyperparameters of the BP neural network. Bayesian optimization was then used to further determine the optimal structure for the number of neurons in the hidden layers of the network, training a GA-BP neural network rheological stress prediction model that integrates physical mechanisms and data-driven approaches. Process parameters (including deformation temperature, strain rate, and strain), chemical compositions (12 in total), and constitutive parameters (including the thermal deformation activation energy Q, the natural logarithm of the temperature-compensated strain rate factor lnZ, and the power dissipation efficiency η) were used as the input layer of the model, and the corresponding rheological stress values ​​at each strain were set as the output layer. The remaining layers, acting as hidden layers, mine and learn the complex nonlinear relationships between the input data to capture the inherent patterns in the data and predict the rheological stress of low-alloy steel. The model's prediction accuracy has a coefficient of determination R0. 2 The strength reached 0.96, and the MAE was 8.764 MPa.

[0049] Example 3 The present invention also provides a low-alloy steel rheological stress prediction system based on the fusion of constitutive parameters and GA-BP neural network, for implementing the method described in the foregoing embodiments. The system includes: a standardized dataset acquisition module, a constitutive feature parameter acquisition module, a training dataset acquisition module, a model building module, and a prediction module. The standardized dataset acquisition module is used to acquire and preprocess the raw data of hot compression experiments of low alloy steel under different hot deformation conditions to obtain a standardized dataset. The constitutive feature parameter acquisition module is used to obtain the constitutive feature parameters of low alloy steel under different hot deformation conditions based on a standardized dataset, combined with the Arrhenius constitutive relation model and DMM theory, and using machine learning algorithms. The training dataset acquisition module is used to construct a training dataset based on the standardized dataset and constitutive feature parameters; The model building module is used to optimize the backpropagation neural network based on the genetic algorithm, build the GA-BP neural network model, and train it based on the training dataset to obtain the rheological stress prediction model. The prediction module is used to predict the flow stress of low alloy steel based on the flow stress prediction model.

[0050] Furthermore, the constitutive feature parameter acquisition module includes: a grouping unit, a first parameter calculation unit, and a second parameter calculation unit; Grouping units are used to group the standardized dataset by deformation temperature for each strain condition; The first parameter calculation unit is used to calculate the parameters contained in the Arrhenius constitutive relation model based on the grouped data. The second parameter calculation unit is used to construct a heat treatment map based on the grouped data, using DMM theory combined with machine learning algorithms, and to calculate the various parameters contained in the heat treatment map.

[0051] Furthermore, based on the grouped data, methods for calculating the parameters included in the Arrhenius constitutive relation model include: Calculate the temperature-compensated deformation rate factor Z parameter: ; in, For strain rate, Q The activation energy for thermal deformation, R The molar gas constant, T This refers to absolute temperature.

[0052] Furthermore, based on the grouped data, a heat treatment map is quickly constructed using DMM theory combined with machine learning algorithms. The methods for calculating the various parameters included in the heat treatment map include: The strain rate sensitivity index was calculated using the finite difference method. m : ; in, For dissipative covariance, G Power dissipation, For rheological stress; Power dissipation factor or : ; in, This represents the maximum value of the dissipation covariance. Instability Criteria g : .

[0053] Furthermore, the model building module includes: an infrastructure unit, a first optimization unit, a second optimization unit, and a training unit; Infrastructure unit, used to build basic network architecture based on MLP model; The first optimization unit is used to obtain the optimal combination of the number of neurons in the hidden layer of the basic network architecture through Bayesian optimization, and to obtain the optimal network topology. The second optimization unit is used to globally optimize the initial weights, thresholds and key hyperparameters of the optimal network topology using a genetic algorithm to obtain the optimized network architecture. The training unit is used to train the optimized network architecture using the training dataset to obtain a rheological stress prediction model.

[0054] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting the rheological stress of low-alloy steel based on the fusion of constitutive parameters and GA-BP neural network, characterized in that, The method includes: The raw data of hot compression experiments on low alloy steel under different hot deformation conditions were obtained and preprocessed to obtain a standardized dataset. Based on standardized datasets, combined with the Arrhenius constitutive relation model and DMM theory, machine learning algorithms were used to obtain constitutive characteristic parameters of low alloy steel under different hot deformation conditions. A training dataset is constructed based on a standardized dataset and constitutive feature parameters; A GA-BP neural network model was constructed by optimizing the backpropagation neural network using a genetic algorithm and trained on a training dataset to obtain a rheological stress prediction model. Based on the rheological stress prediction model, the rheological stress of the low alloy steel to be predicted is predicted.

2. The method according to claim 1, characterized in that, Based on standardized datasets, and combining the Arrhenius constitutive relation model and DMM theory, machine learning algorithms are used to obtain constitutive characteristic parameters of low-alloy steel under different hot deformation conditions. The methods include: For each strain condition, the standardized dataset is grouped according to the deformation temperature; Based on the grouped data, calculate the parameters included in the Arrhenius constitutive relation model; Based on the grouped data, a heat treatment map is constructed using DMM theory combined with machine learning algorithms, and the various parameters contained in the heat treatment map are calculated.

3. The method according to claim 2, characterized in that, Based on the grouped data, methods for calculating the parameters included in the Arrhenius constitutive relation model include: Calculate the temperature-compensated deformation rate factor Z parameter: ； in, For strain rate, Q The activation energy for thermal deformation, R The molar gas constant, T This refers to absolute temperature.

4. The method according to claim 3, characterized in that, Based on the grouped data, a heat treatment map is quickly constructed using DMM theory combined with machine learning algorithms. The methods for calculating the various parameters included in the heat treatment map include: The strain rate sensitivity index was calculated using the finite difference method. m : ； in, For dissipative covariance, G Power dissipation, For rheological stress; Power dissipation factor η : ； in, This represents the maximum value of the dissipation covariance. Instability Criteria ζ : 。 5. The method according to claim 1, characterized in that, The methods for obtaining a rheological stress prediction model by optimizing a backpropagation neural network using a genetic algorithm, constructing a GA-BP neural network model, and training it on a training dataset include: Building a basic network architecture based on the MLP model; The optimal combination of neurons in the hidden layer of the basic network architecture is obtained through Bayesian optimization, thus achieving the optimal network topology. A genetic algorithm is used to globally optimize the initial weights, thresholds, and key hyperparameters of the optimal network topology to obtain the optimized network architecture. The optimized network architecture was trained using the training dataset to obtain a rheological stress prediction model.

6. A system for predicting the rheological stress of low-alloy steel based on the fusion of constitutive parameters and a GA-BP neural network, the system being used to implement the method described in any one of claims 1-5, characterized in that, The system includes: a standardized dataset acquisition module, a constitutive feature parameter acquisition module, a training dataset acquisition module, a model building module, and a prediction module; The standardized dataset acquisition module is used to acquire and preprocess the raw data of hot compression experiments of low alloy steel under different hot deformation conditions to obtain a standardized dataset. The constitutive feature parameter acquisition module is used to obtain the constitutive feature parameters of low alloy steel under different hot deformation conditions based on a standardized dataset, combined with the Arrhenius constitutive relation model and DMM theory, and using machine learning algorithms. The training dataset acquisition module is used to construct a training dataset based on the standardized dataset and constitutive feature parameters; The model building module is used to optimize the backpropagation neural network based on the genetic algorithm, build the GA-BP neural network model, and train it based on the training dataset to obtain the rheological stress prediction model. The prediction module is used to predict the flow stress of low alloy steel based on the flow stress prediction model.

7. The system according to claim 6, characterized in that, The constitutive feature parameter acquisition module includes: a grouping unit, a first parameter calculation unit, and a second parameter calculation unit; Grouping units are used to group the standardized dataset by deformation temperature for each strain condition; The first parameter calculation unit is used to calculate the parameters contained in the Arrhenius constitutive relation model based on the grouped data. The second parameter calculation unit is used to construct a heat treatment map based on the grouped data, using DMM theory combined with machine learning algorithms, and to calculate the various parameters contained in the heat treatment map.

8. The system according to claim 7, characterized in that, Based on the grouped data, methods for calculating the parameters included in the Arrhenius constitutive relation model include: Calculate the temperature-compensated deformation rate factor Z parameter: ； in, For strain rate, Q The activation energy for thermal deformation, R The molar gas constant, T This refers to absolute temperature.

9. The system according to claim 8, characterized in that, Based on the grouped data, a heat treatment map is quickly constructed using DMM theory combined with machine learning algorithms. The methods for calculating the various parameters included in the heat treatment map include: The strain rate sensitivity index was calculated using the finite difference method. m : ； in, For dissipative covariance, G Power dissipation, For rheological stress; Power dissipation factor η : ； in, This represents the maximum value of the dissipation covariance. Instability Criteria ζ : 。 10. The system according to claim 6, characterized in that, The model building module includes: basic architecture unit, first optimization unit, second optimization unit, and training unit; Infrastructure unit, used to build basic network architecture based on MLP model; The first optimization unit is used to obtain the optimal combination of the number of neurons in the hidden layer of the basic network architecture through Bayesian optimization, and to obtain the optimal network topology. The second optimization unit is used to globally optimize the initial weights, thresholds and key hyperparameters of the optimal network topology using a genetic algorithm to obtain the optimized network architecture. The training unit is used to train the optimized network architecture using the training dataset to obtain a rheological stress prediction model.

Images