Multi-component material true stress prediction method based on machine self-learning algorithm

Abstract

The invention relates to the field of material science and engineering application, in particular to a multi-component material true stress prediction method based on a machine self-learning algorithm. According to the method, firstly, true stress strain test data of a multi-component target material under the characteristic temperature and the characteristic strain condition are collected, through data cleaning and preprocessing, analysis is conducted through various machine self-learning algorithms, an optimal true stress prediction algorithm model is constructed, and true stress is reasonably and accurately predicted through the model. According to the method, mechanical property test experiment processes of a large number of multi-component materials are combined, true stress prediction is carried out in a mode of combining experiments and machine learning, and meanwhile, follow-up optimization processing is carried out on different scenes of different component materials, so that the precision is guaranteed. According to the method, the availability and reliability of experimental data can be effectively improved and verified, the labor and time cost is greatly reduced, and the method has great engineering application value.

Description

Technical Field

[0001] This invention relates to the field of materials science and engineering applications, specifically to a method for predicting the true stress of multi-component materials based on a machine self-learning algorithm. Background Technology

[0002] The high-temperature flow stress during the hot rolling process of multi-component alloys is closely related to the deformation temperature, strain rate, strain degree, alloy composition and microstructure. The testing and acquisition of true stress and strain under the characteristic temperature and characteristic strain conditions of multi-component materials is cumbersome, time-consuming and labor-intensive. If the true stress of multi-component materials can be predicted reasonably and accurately, it will improve and verify the usability and reliability of experimental data, which is of great significance for greatly reducing labor and time costs in the future.

[0003] Machine learning methods are widely used in metal research. Compared to traditional Arrhenius models and many improved Arrhenius models, machine learning algorithms have a large optimization space, allowing for adjustments to accuracy as needed. Different machine learning algorithms can be used to meet different accuracy requirements for different samples. Among these, the Backpropagation Neural Network (BPNN) algorithm can handle problems such as composition prediction and service performance prediction, and is suitable for situations with many components and complex matrices. The Support Vector Machine (SVM) algorithm uses a hinge loss function to calculate empirical risk and adds a regularization term to optimize structural risk, making it well-suited for solving small sample problems and complementing the BPNN algorithm. Therefore, a machine self-learning algorithm is used to train and construct a true stress prediction model for multi-component target materials. Summary of the Invention

[0004] In view of at least one of the above-mentioned technical problems, this invention provides a method for predicting the true stress of multi-component materials based on machine self-learning algorithms. Machine self-learning algorithms such as SVM and BPNN are used to calculate the prediction score of the input samples to obtain the predicted true stress of the multi-component materials. Support vector machines are used to predict the true stress of the multi-component materials under characteristic temperature and strain conditions, thereby overcoming the complexity and limitations of traditional numerical analysis and processing methods in hot working research.

[0005] The technical solution of this invention is:

[0006] A method for predicting true stress in multi-component materials based on a machine self-learning algorithm includes the following steps:

[0007] 1) Select multiple target materials and conduct true stress-strain tests under characteristic temperature and characteristic strain conditions;

[0008] 2) Collect experimental data under characteristic temperature and characteristic strain conditions, and preprocess the collected data;

[0009] 3) Machine self-learning algorithms, such as Support Vector Machine (SVM) or Backpropagation Neural Network (BPNN), are used to predict the true stress of multi-component materials under non-feature conditions.

[0010] In the aforementioned method for predicting the true stress of multi-component materials based on machine self-learning algorithms, step 1) requires that the multi-component material contain no less than two principal components, each with a mass fraction of no less than 0.1%; the characteristic temperature should include at least three temperature conditions with a temperature interval of no less than 50°C; and the characteristic strain conditions should include at least three different strain rates with a ratio between strain rates of no less than 5, i.e., each strain rate should differ by at least 5 times.

[0011] In the multi-component material true stress prediction method based on machine self-learning algorithm, step 2) involves preprocessing the collected data, including but not limited to the following processing methods: establishing the relationship between true stress, temperature, true strain, and strain rate under characteristic temperature and characteristic strain conditions.

[0012] In the multi-component material true stress prediction method based on machine self-learning algorithm, step 3) refers to non-characteristic conditions that are different from characteristic temperature and characteristic strain conditions. A suitable machine self-learning algorithm is selected according to the number of components.

[0013] In the multi-component material true stress prediction method based on machine self-learning algorithm, in step 3), the support vector machine (SVM) algorithm establishes a linear model through the relationship between vectors and mechanical properties, and seeks a hyperplane to minimize the preset positive parameter tolerance error.

[0014] The aforementioned method for predicting true stress in multi-component materials based on machine self-learning algorithms uses the Support Vector Machine (SVM) algorithm to determine the parameters in the linear model by minimizing the regularization risk function. Cross-validation is used to study and select the values ​​of the kernel function and regularization parameters. The studied kernel functions include, but are not limited to, linear kernel functions, polynomial kernel functions, and Gaussian radial basis functions. Combinations of kernel function and regularization parameters are listed, along with a score table that evaluates the performance of each combination. The values ​​of the kernel function and regularization parameters are selected based on the final score.

[0015] The multi-element material true stress prediction method based on machine self-learning algorithm uses the following function as the prediction function in the Support Vector Machine (SVM) algorithm:

[0016]

[0017] In the formula, W i For the weighting parameters; φ(X) i ) is the feature mapping function; b is the bias term;

[0018] The selection of the regularization parameter C in the prediction function should satisfy, but is not limited to, avoiding overfitting or using cross-validation. After multiple iterations of training with different C values, the performance should be verified, and the optimal value should be selected after balancing performance and computational efficiency.

[0019] In the multi-component material true stress prediction method based on machine self-learning algorithm, in step 3), the backpropagation neural network (BPNN) algorithm uses the material's deformation temperature, strain rate, and true strain as input nodes and the true stress as output node; based on the Levenberg-Marquardt algorithm, single-layer or double-layer hidden layers are used for data training, and the hidden layer and output layer use the hyperbolic tangent activation function.

[0020] An initial iteration rate is given and adjusted as iterations proceed and weights change; the computational accuracy (convergence threshold) of the model is set in advance during training. When its prediction accuracy and number of iterations meet the given threshold, the iteration stops and the prediction model is output.

[0021] The multi-component material true stress prediction method based on machine self-learning algorithm uses the PSO algorithm to further improve training accuracy and performance. The data is optimized when the intermediate layer of the backpropagation neural network (BPNN) algorithm outputs the results, and then inputs them into the next layer to reduce the amount of computation and improve the calculation speed.

[0022] In the multi-component material true stress prediction method based on machine self-learning algorithm, in step 3), the support vector machine (SVM) algorithm or backpropagation neural network (BPNN) algorithm uses the correlation coefficient R to represent the accuracy of the multi-component material true stress prediction result.

[0023] The design concept of this invention is:

[0024] This invention first collects true stress-strain test data of multi-component target materials under characteristic temperature and strain conditions. After data cleaning and preprocessing, it utilizes various machine learning algorithms for analysis. The SVM algorithm excels at handling small samples and nonlinear problems, while the BPNN algorithm performs well in complex pattern recognition and nonlinear mapping. By integrating the prediction results of these algorithms, not only is the robustness and generalization ability of the prediction improved, but the complementary effect between the algorithms also uncovers deeper patterns in the data, enhancing the model's prediction accuracy and constructing an optimal true stress prediction algorithm model. This model is then used to make reasonable and accurate predictions of true stress. This method combines extensive mechanical property testing experiments of multi-component materials, conducting true stress prediction through a combination of experiments and machine learning. Subsequent optimization processing is performed for different scenarios with different component materials to ensure accuracy. Compared with traditional numerical analysis, this invention significantly reduces processing complexity and improves prediction efficiency by utilizing machine learning methods. The machine learning model can handle high-dimensional data and nonlinear relationships, compensating for the shortcomings of traditional methods in handling nonlinearity and multi-factor interactions, providing a new solution for the study of complex problems such as thermal processing.

[0025] This invention constructs an efficient and accurate true stress prediction system for multi-component materials through the interaction of multiple innovative aspects, including multi-algorithm collaboration, deep integration of data and experiments, feature engineering optimization, and iterative model updates. This provides a powerful prediction tool and theoretical support for the fields of materials science and thermal processing.

[0026] The advantages and beneficial effects of this invention are as follows:

[0027] This invention utilizes true stress-strain experiments on multi-component materials under characteristic temperature and strain conditions. SVM and BPNN algorithms are then employed to establish true stress prediction models for different application scenarios, enabling the prediction of true stress under non-characteristic conditions in multi-component materials. This method addresses the limitation of existing research, which is primarily focused on limited temperature and strain rate ranges due to economic factors and equipment testing capabilities. It generates highly accurate prediction models using only a limited dataset, achieving high precision and significantly saving time, manpower, and experimental costs. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart for data analysis and model data processing.

[0030] Figure 2 This is a diagram of the BPNN algorithm structure.

[0031] Figure 3 The figure shows the correlation coefficients of the BPNN algorithm model under different temperatures and a certain strain (Example 1). In the figure, the horizontal axis Target represents the normalized target value, and the vertical axis Output represents the normalized output value. The approximate fitting relationship between the two is Output≈0.96*Target+0.028.

[0032] Figure 4 The figure shows the prediction results of the SVM algorithm model under a certain strain at different temperatures (Example 2). In the figure, the horizontal axis represents the actual daily yield, the vertical axis represents the predicted daily yield, the scatter plot of actual vs. predicted values ​​shows the comparison between the actual and predicted values, and the diagonal line represents the diagonal line that measures the difference between the actual and predicted values. The closer the actual and predicted values ​​are, the closer the scatter plot will be to or the same as the diagonal line. Detailed Implementation

[0033] In specific implementation, according to a first aspect of the present invention, a method for predicting the true stress of multi-component materials under characteristic temperature and characteristic strain conditions based on SVM and BPNN algorithms is provided, comprising the following steps:

[0034] (1) Select multiple target materials to carry out true stress-strain tests under characteristic temperature and characteristic strain conditions;

[0035] (2) Collect experimental data under characteristic temperature and characteristic strain conditions, and preprocess the collected data;

[0036] (3) Establish the relationship between true stress, temperature, true strain, and strain rate under characteristic temperature and characteristic strain conditions;

[0037] (4) When there are fewer components (≤3) and the matrix is ​​simple, the Support Vector Machine (SVM) algorithm is used to solve the problem; when there are more components (≥4) and the matrix is ​​complex, the Backpropagation Neural Network (BPNN) algorithm is used to solve the problem. The two algorithms complement each other.

[0038] (5) When using the SVM algorithm, the parameters in the linear model are determined by minimizing the regularization risk function;

[0039] (6) Use cross-validation to study and select the values ​​of kernel function and regularization parameter;

[0040] (7) List the combinations of kernel function and regularization parameter and the score table that can evaluate the performance of each combination, and select the values ​​of kernel function and regularization parameter based on the final score;

[0041] (8) Introduce a loss function to penalize the error and prevent the prediction function from overfitting or underfitting;

[0042] (9) Prediction function L ε (f(X i )-y i ) is an ε-intensitive function, which can be described as:

[0043] It penalizes errors exceeding ε;

[0044] In the formula, ε is a positive threshold parameter that determines the strength of the soft thresholding operation; |z| is used to measure the magnitude of the prediction error, regardless of direction; f(X) i ) is the model's input X i The predicted output on; y i The true target value or label of the i-th sample;

[0045] (10) The correlation coefficient R is used to represent the accuracy of the true stress prediction results of the multi-element material, and the final SVM algorithm prediction function is determined as follows:

[0046]

[0047] In the formula, C is a constant parameter; Lε(f(X) i )-y i ) is the ε-insensitive loss function; W is the model's weight vector;

[0048] (11) When using the BPNN algorithm: Design a network structure consisting of three layers: input layer, hidden layer and output layer;

[0049] (12) The deformation temperature, strain rate, and true strain of the material are used as input nodes, and the true stress is used as the output node. Based on the Levenberg-Marquardt algorithm, a single hidden layer is used for data training;

[0050] (13) Pre-set the computational precision. Input the input matrix into the network and repeat the forward propagation, error calculation, back propagation, and weight and threshold update steps for training;

[0051] (14) When the intermediate results are output in the first hidden layer, the Particle Swarm Optimization (PSO) algorithm is used to optimize the intermediate results, so as to avoid the situation that the process will not end for a long time due to divergence and affect the optimization effect.

[0052] In some embodiments of the present invention, the multi-component material should contain no less than two main component elements, the content of which is not less than 0.1%; the characteristic temperature should include at least three temperature conditions with a temperature interval of not less than 50°C; the characteristic strain condition should include at least three different strain rates with a ratio between strain rates of not less than 5, that is, each strain rate should differ by at least 5 times.

[0053] This invention has two embodiments to verify the accuracy and practicality of the method. Embodiment 1 is described first:

[0054] Example 1

[0055] like Figure 1 As shown in the figure, this embodiment proposes a method for predicting the true stress of multi-component materials based on a machine self-learning algorithm. The method specifically includes the following steps:

[0056] (1) Obtain the mechanical characteristic parameters of the material through mechanical property testing. Mechanical property experiments include, but are not limited to, mechanical testing methods such as tensile test, compression test or bending test, measuring the deformation of the alloy during the mechanical testing process, measuring the stress response of the alloy under different conditions, etc.

[0057] (2) The experimental data collection process included, but was not limited to, casting, forging, rolling, spinning, extrusion, drawing, hot isostatic pressing, welding, or stamping processes under different temperature conditions. A high-strength, high-elasticity copper alloy with a composition of Cu-2.98Ti was prepared, and the compression specimens were cylindrical with a fixed height and diameter. The experiments were conducted using a Gleeble-3800 instrument, and the experimental conditions included different temperatures (500℃, 600℃, 700℃, and 800℃) and strain rates (0.01 s⁻¹). -1 0.1s -1 1s -1 and 10s -1 ).

[0058] (3) Preprocessing of data obtained from numerous experiments, including but not limited to data cleaning, data transformation, or data storage. Use Spark's dropDuplicates() function to remove duplicate rows. Use statistical methods (such as standard deviation or quartiles) or business logic to identify outliers. Perform data cleaning quickly by setting thresholds and using map operations, and then use reduce operations to aggregate the data cleaning results.

[0059] (4) There are m samples collected. Among them, m1 samples are used as the training set, m2 samples are used as the validation set, and the remaining samples m3 are used as the test set.

[0060] (5) The collected data has n variables. It is converted into a vector in the format x = [x1, x2, x3...] as the input of the SVM algorithm.

[0061] (6) The value of each feature is linearly scaled to the interval [0,1] by calculating the minimum and maximum values ​​of each feature. This is done to adjust the data to a uniform scale, making different features comparable.

[0062] (7) By adding the L1 norm of the parameters as a regularization term to the objective function, the parameters of the model tend to be sparse. During training, the weights of some features are set to zero, making the model simpler and more interpretable.

[0063] (8) The stochastic gradient descent (SGD) algorithm is introduced to optimize the loss function in the original prediction function. In this embodiment of the invention, the loss function W is the model parameter to be optimized. In order to find a suitable set of values ​​that minimizes the value of W, a sample (x_i, y_i) is selected in each iteration to calculate the gradient. The gradient of W on the sample can be expressed as:

[0064]

[0065] The update formula for stochastic gradient descent is as follows:

[0066]

[0067] Where η represents the learning rate, which is used to control the step size of each update.

[0068] During the iterative process of stochastic gradient descent, the following steps are repeated continuously:

[0069] Randomly select a sample (x_i, y_i).

[0070] Calculate the gradient of the loss function f(w) on the sample (x_i, y_i).

[0071] Update W according to the update formula.

[0072] Repeat steps 1-3 until the stopping condition is met (such as reaching the preset number of iterations or the loss function value being less than a certain threshold).

[0073] (9) Since SGD only uses one sample for each update, its update process may be quite volatile. This embodiment improves the implementation process of SGD by using more than one mini-batch of samples each time a sample is selected.

[0074] The prediction function after stochastic gradient optimization can be expressed as:

[0075]

[0076] In the formula, C is a constant parameter; Lε(f(X) i )-y i ) is the ε-insensitive loss function; W is the model's weight vector;

[0077] (10)Lε(f(X i )-y i ) is an ε-insensitive loss function that penalizes all errors exceeding ε.

[0078] (11) Select the kernel function and constant parameter C. List all preset parameter values ​​and verify the performance of each parameter combination using a test set.

[0079] (12) In this embodiment of the invention, the preset value of parameter C is adjusted in increments of one order of magnitude. The constant parameter C is used to adjust the fitting ability of the model. The larger the C, the higher the tolerance for error, but it may cause overfitting. Conversely, the smaller the C, the more likely it is to underfit.

[0080] (13) This invention experimented with three kernel functions: linear kernel, Gaussian kernel, and polynomial kernel, and listed the prediction scores of each kernel function. The linear kernel is simple and fast to compute and performs well on linear problems; the polynomial kernel can handle nonlinear problems, and the model complexity can be controlled by adjusting the number of polynomial terms; the Gaussian kernel handles a wide range of problem spaces and is suitable for most nonlinear problems.

[0081] (14) As shown in Table 1, the prediction scores under different combinations of constant parameter C and kernel function are obtained. It can be seen that from 0.01 to 100, with a gradient of one order of magnitude, the constant parameter C, regardless of the kernel function used, although the magnitude of C is not strictly positively correlated with the prediction score under a single constraint, generally speaking, the larger the value of C, the higher the comprehensive score for both materials. Similarly, it can be seen that among the three kernel functions in the experiment, the Gaussian kernel function has a higher prediction score in the high-precision region. In this embodiment of the invention, the Gaussian radial basis function is selected as the kernel function, and 100 is selected as the value of the constant parameter C.

[0082] Table 1 Prediction scores under different combinations of constant parameter C and kernel function

[0083]

[0084]

[0085] (15) In this embodiment of the invention, the correlation coefficient R is used to represent the degree of correlation between the predicted result and the actual value. The closer the R value is to 1, the more accurate the prediction. The correlation coefficient R (test set) of the sample set using the SVM algorithm model is 0.98873. The relationship prediction of this invention is very accurate and is an effective and feasible prediction method.

[0086] (16) As shown in Table 2, the actual values ​​and predicted values ​​of the existing model are compared using a portion of the test set:

[0087] Table 2 Comparison of some actual and predicted values ​​in Example 1 (SVM algorithm)

[0088] actual value 143 412 322 210 560 Predicted value 149.00 407.98 321.981 213.22 549.97

[0089] (17) Divide the dataset according to step (4) to obtain an m1*n matrix as the input matrix.

[0090] (18) such as Figure 2 As shown, a three-layer BPNN algorithm structure was designed, including an input layer, a hidden layer, and an output layer, with the computational precision set in advance. Based on the actual situation such as the dataset and accuracy requirements, this embodiment of the invention sets the upper limit of the number of iterations to 12,000 and the convergence accuracy to 0.955.

[0091] (19) Input the input matrix into the network and repeat the following steps to begin training:

[0092] Forward propagation: Based on the current network weights and thresholds, the output of each neuron is calculated, mainly through weighted summation and activation function calculation. The neurons in the output layer will produce the network's predicted output.

[0093] Calculation error: Compare the network's actual output with the target output and calculate the error (e.g., mean squared error).

[0094] Backpropagation: Based on the calculated error, starting from the output layer, the error gradient of each neuron is calculated layer by layer in reverse. The chain rule and gradient descent algorithm are used to calculate the effect of weights and thresholds on the error (i.e., the gradient).

[0095] Update weights and thresholds: Based on the calculated gradient, adjust the weights and thresholds of each neuron according to the learning rate. The learning rate is a parameter between 0 and 1 that controls the step size for weight and threshold updates.

[0096] (20) When the intermediate results are output from the first hidden layer, the PSO algorithm is used to optimize the intermediate results. Since PSO iteratively searches for the optimal solution in the solution space, this embodiment of the invention initializes multiple particle swarms simultaneously and searches in different directions. The search termination condition is: a particle with a fitness value that meets the requirements appears, or multiple initial swarms search the same region and their movement speed slows down. This measure may yield a suboptimal solution, but it avoids the situation where the optimization effect is affected by the divergence and cannot be terminated for a long time.

[0097] (21) The iteration rate should be controlled during the iteration process. Since the training dataset is relatively large, the initial iteration rate is set to a slightly larger value (0.1) and then gradually reduced. The main considerations are the number of iterations and the current degree of convergence. The iteration step size should decrease by no more than 0.05 each time.

[0098] (22) The model trained by the BPNN algorithm is obtained. The correlation coefficient R is still used to represent the degree of correlation between the predicted result and the actual value. The closer the R value is to 1, the more accurate the prediction. The correlation coefficient R (test set) of the BPNN algorithm model on the sample set is 0.97774. The data shows that the relationship prediction of the present invention is very accurate and is an effective and feasible prediction method.

[0099] (23) As shown in Table 3, the actual values ​​and predicted values ​​of the existing model are compared using a portion of the test set:

[0100] Table 3 Comparison of some actual and predicted values ​​in Example 1 (BPNN algorithm)

[0101] actual value 200 352 87 366 190 Predicted value 189.37 343.28 93.64 356.17 181.00

[0102] (24) such as Figure 3 As shown, the correlation coefficient of the BPNN algorithm prediction model more intuitively demonstrates the accuracy of the model. The closer the scatter points are to the standard line, the better the model's performance.

[0103] (25) From the comparison of the results of the specific implementation of Embodiment 1 of the present invention, it can be seen that the BPNN algorithm is suitable for situations with a large number of components and complex matrix, while the SVM algorithm has excellent performance and higher accuracy when there are few components.

[0104] As a supplement, Example 2 is described below:

[0105] Example 2

[0106] In this embodiment, a method for predicting the true stress of multi-component materials based on a machine self-learning algorithm includes the following steps:

[0107] (1) Obtain the mechanical characteristic parameters of the material through mechanical property testing. Mechanical property experiments include, but are not limited to, mechanical testing methods such as tensile test, compression test or bending test, measuring the deformation of the alloy during the mechanical testing process, measuring the stress response of the alloy under different conditions, etc.

[0108] (2) The experimental data collection process included, but was not limited to, casting, forging, rolling, spinning, extrusion, drawing, hot isostatic pressing, welding, or stamping processes under different temperature conditions. A precipitation-strengthened copper alloy with the composition Cu-2.35Ni-0.59Si-0.15La was prepared, and the compression specimens had a fixed height and diameter. The experiments were conducted using a Gleeble-3800D instrument, and the experimental conditions included different temperatures (500℃, 600℃, 700℃, 800℃, and 900℃) and strain rates (0.1 s⁻¹). -1 1s -1 and 10s -1 ).

[0109] (3) Preprocessing of data obtained from numerous experiments, including but not limited to data cleaning, data transformation, and data storage. Use Spark's dropDuplicates() function to remove duplicate rows. Use statistical methods (such as standard deviation, quartiles, etc.) or business logic to identify outliers. Set thresholds, perform rapid data cleaning using map operations, and then use reduce operations to aggregate the cleaned data results.

[0110] (4) There are m samples collected. Among them, m1 samples are used as the training set, m2 samples are used as the validation set, and the remaining samples m3 are used as the test set.

[0111] (5) The collected data has n variables. It is converted into a vector in the format x = [x1, x2, x3...] as the input of the SVM algorithm.

[0112] (6) The value of each feature is linearly scaled to the interval [0,1] by calculating the minimum and maximum values ​​of each feature. This is done to adjust the data to a uniform scale, making different features comparable.

[0113] (7) By adding the L1 norm of the parameters as a regularization term to the objective function, the parameters of the model tend to be sparse. During training, the weights of some features are set to zero, making the model simpler and more interpretable.

[0114] (8) The stochastic gradient descent (SGD) algorithm is introduced to optimize the loss function in the original prediction function. In this embodiment of the invention, the loss function W is the model parameter to be optimized. In order to find a suitable set of values ​​that minimizes the value of W, a sample (x_i, y_i) is selected in each iteration to calculate the gradient. The gradient of W on the sample can be expressed as:

[0115]

[0116] The update formula for stochastic gradient descent is as follows:

[0117]

[0118] Where η represents the learning rate, which is used to control the step size of each update.

[0119] During the iterative process of stochastic gradient descent, the following steps are repeated continuously:

[0120] Randomly select a sample (x_i, y_i).

[0121] Calculate the gradient of the loss function f(w) on the sample (x_i, y_i).

[0122] Update W according to the update formula.

[0123] Repeat steps 1-3 until the stopping condition is met (such as reaching the preset number of iterations or the loss function value being less than a certain threshold).

[0124] (9) Since SGD only uses one sample for each update, its update process may be quite volatile. This embodiment improves the implementation process of SGD by using more than one mini-batch of samples each time a sample is selected.

[0125] The prediction function after stochastic gradient optimization can be expressed as:

[0126]

[0127] In the formula, C is a constant parameter; Lε(f(X) i )-y i ) is the ε-insensitive loss function; W is the model's weight vector;

[0128] (10)Lε(f(X i )-y i ) is an ε-insensitive loss function that penalizes all errors exceeding ε.

[0129] (11) Select the kernel function and constant parameter C. List all preset parameter values ​​and verify the performance of each parameter combination using a test set.

[0130] (12) In this embodiment of the invention, the preset value of parameter C is adjusted in increments of one order of magnitude. The constant parameter C is used to adjust the fitting ability of the model. The larger the C, the higher the tolerance for error, but it may cause overfitting. Conversely, the smaller the C, the more likely it is to underfit.

[0131] (13) In this embodiment of the invention, the correlation coefficient R is used to represent the degree of correlation between the predicted result and the actual value. The closer the R value is to 1, the more accurate the prediction. The correlation coefficient R (test set) of the sample set using the SVM algorithm model is 0.94038. The relationship prediction of this invention is relatively accurate and is an effective and feasible prediction method.

[0132] (14) such as Figure 4 As shown in the scatter plot of the SVM algorithm prediction model, the closer the scatter points are to the standard line, the better the model's performance.

[0133] (15) As shown in Table 4, the actual values ​​and predicted values ​​of the existing model are compared using a portion of the test set:

[0134] Table 4 Comparison of Actual and Predicted Values ​​in Example 2 (SVM Algorithm)

[0135] actual value 50 33 77 396 440 Predicted value 61.98 54.16 87.90 379.89 397.86

[0136] (16) Divide the dataset according to step (4) to obtain an m1*n matrix as the input matrix.

[0137] (17) A three-layer BPNN algorithm structure was designed, including an input layer, a hidden layer, and an output layer. The computational precision was pre-set. Based on the actual situation of the dataset and precision requirements, this embodiment of the invention sets the upper limit of the number of iterations to 10,000 and the convergence precision to 0.95.

[0138] (18) Input the input matrix into the network and repeat the following steps to begin training:

[0139] Forward propagation: Based on the current network weights and thresholds, the output of each neuron is calculated. This mainly involves weighted summation and activation function calculation. The neurons in the output layer will produce the network's predicted output.

[0140] Calculation error: Compare the network's actual output with the target output and calculate the error (e.g., mean squared error).

[0141] Backpropagation: Based on the calculated error, starting from the output layer, the error gradient of each neuron is calculated layer by layer in reverse. The chain rule and gradient descent algorithm are used to calculate the effect of weights and thresholds on the error (i.e., the gradient).

[0142] Update weights and thresholds: Based on the calculated gradient, adjust the weights and thresholds of each neuron according to the learning rate. The learning rate is a parameter between 0 and 1 that controls the step size for weight and threshold updates.

[0143] (19) When the intermediate results are output from the first hidden layer, the PSO algorithm is used to optimize the intermediate results. Since PSO iteratively searches for the optimal solution in the solution space, this embodiment of the invention initializes multiple particle swarms simultaneously and searches in different directions. The search termination condition is: a particle with a fitness value that meets the requirements appears, or multiple initial swarms search the same region and their movement speed slows down. This measure may yield a suboptimal solution, but it avoids the situation where the optimization effect is affected by the divergence and cannot be terminated for a long time.

[0144] (20) The iteration rate should be controlled during the iteration process. Since the training dataset is relatively large, the initial iteration rate should be a slightly large value (0.1) and then gradually reduced. The main considerations are the number of iterations and the current degree of convergence. The iteration step size should decrease by no more than 0.05 each time.

[0145] (21) The model trained by the BPNN algorithm is obtained, and the correlation coefficient R is still used to represent the degree of correlation between the predicted result and the actual value. The correlation coefficient R (test set) of the BPNN algorithm model on the sample set is 0.96066. The data shows that the relationship prediction of the present invention is very accurate and is an effective and feasible prediction method.

[0146] (22) As shown in Table 5, the actual values ​​and predicted values ​​of the existing model are compared using a portion of the test set:

[0147] Table 5. Comparison of some actual and predicted values ​​in Example 2 (BPNN algorithm)

[0148] actual value 376 266 146 108 98 Predicted value 361.52 255.56 160.79 99.84 95.51

[0149] (23) From the comparison of the specific implementation results of Embodiment 2 of the present invention, it can be seen that the BPNN algorithm is suitable for situations with a large number of components and complex matrix, while the SVM algorithm has excellent performance and higher accuracy when there are few components.

[0150] Based on the above prediction and evaluation methods, and combined with the implementation results, the method of this invention can effectively predict the true stress of multi-component materials, greatly saving time, money, and manpower costs, and has high accuracy, making it widely applicable. This method can effectively improve and verify the usability and reliability of experimental data, significantly reducing labor and time costs, and has great engineering application value.

[0151] The above embodiments are merely illustrative of the principles and performance of the present invention and are not exhaustive. People can obtain other embodiments based on these embodiments without creative effort, and these embodiments all fall within the protection scope of the present invention.

Claims

1. A method for predicting the true stress of multi-component materials based on a machine self-learning algorithm, characterized in that, Includes the following steps: 1) Select multiple target materials and conduct true stress-strain tests under characteristic temperature and characteristic strain conditions; 2) Collect experimental data under characteristic temperature and characteristic strain conditions, and preprocess the collected data; 3) Machine self-learning algorithms, such as Support Vector Machine (SVM) or Backpropagation Neural Network (BPNN), are used to predict the true stress of multi-component materials under non-feature conditions.

2. The method for predicting the true stress of multi-component materials based on a machine self-learning algorithm according to claim 1, characterized in that, In step 1), the multi-component material should contain no less than 2 main component elements, and the mass fraction of each main component should be no less than 0.1%; the characteristic temperature should include at least 3 temperature conditions with a temperature interval of no less than 50°C; the characteristic strain condition should include at least 3 different strain rates with a ratio between strain rates of no less than 5, that is, each strain rate should differ by at least 5 times.

3. The method for predicting the true stress of multi-component materials based on a machine self-learning algorithm according to claim 1, characterized in that, In step 2), the collected data is preprocessed, including but not limited to the following processing methods: establishing the relationship between true stress, temperature, true strain, and strain rate under characteristic temperature and characteristic strain conditions.

4. The method for predicting the true stress of multi-component materials based on a machine self-learning algorithm according to claim 1, characterized in that, In step 3), non-characteristic conditions are temperature and / or strain conditions that are different from characteristic temperature and characteristic strain conditions. A suitable machine self-learning algorithm is selected according to the number of components.

5. The method for predicting the true stress of multi-component materials based on a machine self-learning algorithm according to claim 1, characterized in that, In step 3), the Support Vector Machine (SVM) algorithm establishes a linear model by analyzing the relationship between vectors and mechanical properties, and seeks a hyperplane to minimize the preset positive parameter tolerance error.

6. The method for predicting true stress of multi-component materials based on machine self-learning algorithm according to claim 1 or 5, characterized in that, The Support Vector Machine (SVM) algorithm determines the parameters in a linear model by minimizing a regularization risk function. It uses cross-validation to study and select the values ​​of the kernel function and regularization parameters. The kernel functions studied include, but are not limited to, linear kernel functions, polynomial kernel functions, and Gaussian radial basis functions. It lists the combinations of kernel function and regularization parameters and a score table that can evaluate the performance of each combination. The values ​​of the kernel function and regularization parameters are selected based on the final score.

7. The method for predicting true stress of multi-component materials based on machine self-learning algorithm according to claim 6, characterized in that, The Support Vector Machine (SVM) algorithm uses the following function as its prediction function: In the formula, W i For the weighting parameters; φ(X) i ) is the feature mapping function; b is the bias term; The selection of the regularization parameter C in the prediction function should satisfy, but is not limited to, avoiding overfitting or using cross-validation. After multiple iterations of training with different C values, the performance should be verified, and the optimal value should be selected after balancing performance and computational efficiency.

8. The method for predicting the true stress of multi-component materials based on a machine self-learning algorithm according to claim 1, characterized in that, In step 3), the backpropagation neural network (BPNN) algorithm uses the material's deformation temperature, strain rate, and true strain as input nodes and the true stress as output nodes. Based on the Levenberg-Marquardt algorithm, data training is performed using a single or two hidden layers, with the hidden and output layers using the hyperbolic tangent activation function. An initial iteration rate is given and adjusted as iterations proceed and weights change; the computational accuracy (convergence threshold) of the model is set in advance during training. When its prediction accuracy and number of iterations meet the given threshold, the iteration stops and the prediction model is output.

9. The method for predicting the true stress of multi-component materials based on a machine self-learning algorithm according to claim 1 or 8, characterized in that, To further improve training accuracy and performance, the PSO algorithm is used. The data is optimized when the intermediate layer of the backpropagation neural network (BPNN) algorithm outputs the results, and then inputs them into the next layer to reduce the amount of computation and improve the computation speed.

10. The method for predicting the true stress of multi-component materials based on a machine self-learning algorithm according to claim 1, characterized in that, In step 3), the support vector machine (SVM) algorithm or the backpropagation neural network (BPNN) algorithm uses the correlation coefficient R to represent the accuracy of the true stress prediction results of the multi-element material.

Images