Forge piece grain size prediction model training method, prediction method and related device

Abstract

The invention relates to a forging grain size prediction model training method, a forging grain size prediction method and a related device. The forging grain size prediction model training method comprises the following steps that a training set and a verification set formed by preprocessed and standardized process data are used for training a plurality of heterogeneous base regression devices; continuously acquiring test set data formed by the preprocessed and standardized process data, and using the test set data to calculate collinearity values among the basis regression devices in real time; and dynamically determining a plurality of basis regression devices with the highest collinearity based on the collinearity evaluation result, dynamically allocating weights based on the real-time performance of each basis regression device, and combining the determined basis regression devices into a weighted summation integrated model. Through combination of recursive feature elimination and collinear analysis, dual optimization of the features and the base model is realized, meanwhile, a weighted summation integration strategy is put forward, the weight of the base model is dynamically allocated, the advantages of the heterogeneous model are effectively fused, and the prediction precision is remarkably improved.

Description

Technical Field

[0001] The present application relates to the field of superalloy forging, and particularly to a training method, a prediction method and related devices for a prediction model of the grain size of forgings. Background Art

[0002] With the continuous improvement of the requirements for material properties in high-end manufacturing industries such as aerospace and automotive, forgings, as an important metal forming process, are widely used in the production of components with special property requirements such as high strength, high toughness and fatigue resistance. Grain size, as one of the key factors affecting the properties of forgings, its control and optimization are of crucial significance in the forging process. Accurately predicting the grain size of forgings can not only improve the performance stability of products, but also reduce production costs and enhance the overall manufacturing efficiency.

[0003] However, the prediction of the grain size of complex forgings still faces many challenges. Specifically, there is a complex non-linear relationship between the grain size of forgings and various process parameters, and these parameters are dynamically affected by various factors during the actual production process, such as temperature, pressure, deformation rate, etc. The interaction and change of these factors make the traditional machine learning prediction methods often unable to accurately reflect the relationship between the grain size and the process parameters, thus affecting the accuracy and reliability of the prediction. Therefore, how to improve the accuracy of the grain size prediction of complex forgings has become an urgent problem to be solved in the current forging industry.

[0004] Existing research mainly focuses on predicting the grain size through the fitting of mathematical models and experimental data. Traditional prediction methods such as linear regression and neural networks, although achieving certain effects in some cases, still have difficulty in dealing with the non-linear problems in high-complexity processes. In recent years, with the rapid development of artificial intelligence and machine learning technologies, researchers have tried to apply these advanced technologies to the grain size prediction to improve the accuracy and robustness of the model. and The latest techniques for helping to construct an integrated classifier with stacking are discussed. The performance of this method is analyzed for correlation with existing stacking methods, and the best classifier is selected through cross-validation. Baba et al. reviewed various integration techniques and the difficulties involved in selecting the best base classifier to construct an integrated model. The results showed that optimization is the better of the two methods. Pavlyshenko et al. analyzed the performance of the stacking method and logistic regression in time series prediction. The results showed that the integration technique is superior to the single method and improves the prediction performance. Pavani and Beulet analyzed the performance of the knn algorithm based on environmental parameters in predicting crop yields, and the data were from each district in Telangana. Viviliya and Vaidhehi proposed a hybrid model to recommend crops for suitable regions based on geographical and climatic parameters. The dataset used for this purpose was taken from the National Portal of India, Agriculture, Government of India. This technique was constructed using NB, J48, and association rule mining algorithms. Kevin et al. used kNN, DT, cross-validated kNN, as well as nb and SVM techniques to predict suitable crops based on the soil. The results showed that cross-validated knn has the highest prediction accuracy. Hsieh et al. and Guyon et al. constructed the rfe1 technique for selecting significant features from the dataset. Among all the methods used to select features, RFE is a new method for selecting features for small-sample classification problems. These methods have, to a certain extent, addressed the deficiencies of traditional methods, but there are still problems with poor adaptability to complex processes and dynamic changes. Summary of the Invention

[0005] In order to improve the adaptability to complex processes and dynamic changes, the present application provides a training method, a prediction method, and related devices for a forging grain size prediction model.

[0006] In a first aspect, a training method for a forging grain size prediction model provided by the present application adopts the following technical solution: A training method for a forging grain size prediction model includes the following steps: Train several heterogeneous base regressors using a training set and a validation set formed by preprocessed and standardized process data, and determine the internal parameters of each base regressor; Continuously obtain test set data formed by preprocessed and standardized process data, and use the test set data to calculate the collinearity value between each base regressor in real time to dynamically evaluate the similarity between the base regressors; Based on the collinearity evaluation result, dynamically determine several base regressors with the highest collinearity, and dynamically assign weights based on the real-time performance of each base regressor, and combine the determined base regressors into a weighted sum integration model to be used as the grain size prediction model.

[0007] By adopting the above technical solution, through real-time calculation of the collinearity between heterogeneous base regressors and dynamic adjustment of the model combination and weights, the adaptive optimization of the grain size prediction model is realized, effectively solving the problem that the prediction accuracy of the traditional fixed model decreases due to process parameter or production line changes, thereby significantly improving the robustness of the model to changes in the actual production environment and the stability of the prediction results.

[0008] Optionally, the several heterogeneous base regressors include: SVR, RF, Ridge, BP neural network.

[0009] Optionally, the calculation formula for the collinearity value between the base regressors is where c i represents the predicted value obtained after the i-th group of features in the test set is input into a regression model, represents the mean of all predicted values obtained after all groups of features in the test set are input into this regression model; d i represents the predicted value obtained after the i-th group of features in the test set is input into another regression model, represents the mean of all predicted values obtained after all groups of features in the test set are input into this regression model.

[0010] Optionally, the step of dynamically determining several base regressors with the highest collinearity based on the collinearity evaluation result, dynamically allocating weights based on the real-time performance of each base regressor, and combining the determined base regressors into a weighted sum integration model as the grain size prediction model includes the following steps: Take the two base regressors with the highest collinearity value as the best base regressors; Calculate the root mean square error of the predicted values output by the best base regressors, and calculate the weights of the two best base regressors according to the calculated root mean square error, where w i is the weight of the i-th model, and RMSE i is the root mean square error of the i-th model; Take the sum of the products of the calculated weights and the output results of the base regressors as the output of the weighted sum integration model.

[0011] In a second aspect, a method for predicting the grain size of forgings provided by the present application adopts the following technical solution: A method for predicting the grain size of forgings includes the following steps: S1. Continuously obtain several process data measured during the forging process of the forging; among them, the process data is divided into a pre-acquired part and a continuously acquired part; S2. Perform data preprocessing; S3. Split the generated dataset using stratified random sampling to generate a training set, a validation set, and a test set; among them, the continuously acquired part of the process data corresponds to the test set; S4. Use the recursive feature elimination method to screen a number of target features, and standardize the features by removing the mean and scaling to unit variance; S5. Train the training set and the validation set using the training method of the forging grain size prediction model as described above to obtain a weighted summation ensemble model; S6. Input the test set into the weighted summation ensemble model to obtain the forging grain size prediction result.

[0012] By adopting the above technical solution, a dynamically updated forging grain size prediction method is constructed, which incorporates the continuously acquired production process data into the prediction process in real time, and combines the feature recursive elimination method for feature selection and standardization. Furthermore, a weighted summation ensemble model that can be adaptively adjusted is established, enabling the prediction result to respond to process parameter changes in real time, thereby significantly improving the accuracy and stability of forging grain size prediction and effectively reducing the risk of product quality fluctuations caused by parameter fluctuations during the forging production process.

[0013] Optionally, S2 includes the following steps: S21. Delete redundant values; S22. Fill in missing values with the median; S23. Use the box plot method to identify outliers; S24. For values exceeding the upper and lower quartiles by ±1.5 times the interquartile range, truncate them to a preset range and fill the samples.

[0014] Optionally, it is used for the grain size prediction of GH4169 alloy turbine forgings.

[0015] In a third aspect, a computer device provided by the present application adopts the following technical solution: A computer device, which includes: One or more processors; A memory; One or more applications, where the one or more applications are stored in the memory and are configured to be executed by the one or more processors, and the one or more programs are configured to: Execute the above forging grain size prediction method.

[0016] In a fourth aspect, a computer-readable storage medium provided by the present application adopts the following technical solution: A computer-readable storage medium stores a computer program that can be loaded and executed by a processor to perform the above method.

[0017] The storage medium stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement: The forging grain size prediction method as described above.

[0018] In summary, the present application includes at least one of the following beneficial technical effects: 1. Through the combination of recursive feature elimination and collinearity analysis, dual optimization of features and the base model is achieved, and the robustness of the model is improved; 2. A weighted summation integration strategy is proposed to dynamically allocate the weights of the base models, effectively integrating the advantages of heterogeneous models, and significantly improving the prediction accuracy; 3. The computational efficiency is significantly improved through feature screening; 4. By predicting and guiding the adjustment of process parameters, the standard deviation of the forging grain size produced by the traditional process is effectively reduced, and the consistency of product performance is significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The flowchart of a method for training a forging grain size prediction model according to an embodiment of the present invention is shown.

[0020] Figure 2 The flowchart of a method for predicting the forging grain size according to an embodiment of the present invention is shown.

[0021] Figure 3 The flowchart of a method for predicting the forging grain size according to an embodiment of the present invention is shown.

[0022] Figure 4 The flowchart of an ensemble learning model according to an embodiment of the present invention is shown.

[0023] Figure 5 The implementation process of feature selection according to an embodiment of the present invention is shown.

[0024] Figure 6 The schematic diagram of a computer device according to an embodiment of the present invention is shown. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] The following further describes the present application in detail with reference to the drawings. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0026] In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the inventive concept. As part of this specification, some of the drawings in the present disclosure represent structures and devices in block diagram form to avoid obscuring the disclosed principles. For clarity, not all features of actual specific implementations are necessarily described. In addition, the language used in the present disclosure has been mainly selected for readability and guidance purposes and may not have been chosen to delimit or define the subject matter of the invention, and thus reference is made to the necessary claims to determine such inventive subject matter. References in the present disclosure to “a particular implementation” or “particular implementations” mean that the specific features, structures, or characteristics described in connection with that particular implementation are included in at least one particular implementation, and multiple references to “a particular implementation” or “particular implementations” should not be construed as necessarily all referring to the same particular implementation.

[0027] Unless explicitly defined otherwise, the terms “a,” “an,” and “the” are not intended to refer to a singular entity but include the general category for which a particular example can be used for illustration. Thus, the use of the term “a” or “an” can mean any number of at least one, including “one,” “one or more,” “at least one,” and “one or more than one.” The term “or” means any of the alternatives and any combination of the alternatives, including all of the alternatives, unless the alternatives are explicitly indicated as mutually exclusive. The phrase “at least one of” when combined with a list of items refers to a single item in the list or any combination of items in the list. The phrase does not require all of the items listed, unless explicitly so defined.

[0028] The alloy turbine disk is a key engine component widely used in the core parts of aero-engines and space vehicles, especially in extremely harsh working environments such as high temperature and high stress. The alloy turbine disk usually has high temperature resistance, fatigue resistance, creep resistance, and excellent mechanical properties, and the quality of its performance directly determines the reliability and service life of the aero-engine or spacecraft.

[0029] The nickel-based superalloy turbine disk forging is a turbine disk structural part made of nickel as the matrix and adding a variety of strengthening elements through forging process. It has outstanding high temperature strength, high temperature oxidation resistance, and good thermal stability, and can operate stably for a long time in high temperature and high load environments. Therefore, it is widely used in the parts with the highest temperature and the heaviest load in aircraft engines and is a key component affecting the overall quality of the engine.

[0030] The GH4169 alloy is a typical nickel-based superalloy with excellent comprehensive properties, such as high strength, good toughness, creep resistance, and oxidation resistance. At the same time, it has good plasticity and processing performance, so it is particularly suitable for the manufacture of aeroengine turbine disk forgings. The GH4169 alloy has outstanding structural stability under high-temperature conditions and can effectively resist structural damage caused by high temperature and alternating stress.

[0031] The grain size of a forging refers to the size of the grains in the microscopic structure of the forging material, usually characterized by the average grain diameter or the grain grade. The grain size is an important microscopic factor affecting the macroscopic mechanical properties of alloy materials. The finer the grains, the better the strength, toughness, and fatigue resistance of the forging generally are. The grain size is generally measured by observing with a metallographic microscope and image analysis. By measuring the grain size, the quality level and process stability of the forging can be effectively evaluated, thereby guiding and optimizing the actual production process. Accurately and efficiently predicting and controlling the grain size of GH4169 alloy turbine disk forgings is of great significance for improving the manufacturing quality and reliability of aeroengines in China.

[0032] Therefore, the embodiments of this application disclose a training method for a forging grain size prediction model. Refer to Figure 1 , the training method of this forging grain size prediction model includes the following steps A-C.

[0033] Step A. Train several heterogeneous base regressors using the training set and validation set formed by preprocessed and standardized process data, and determine the internal parameters of each base regressor.

[0034] Process data refers to various data indicators directly related to the production process recorded during forging production. These data are usually automatically obtained by the factory digital process acquisition system to ensure the timeliness, continuity, and accuracy of data acquisition. Specifically, it includes the starting forging temperature, final forging temperature, blanking temperature, transfer time, forging time, etc. In different embodiments, it may also include important parameters such as the deformation rate, forging pressure, die temperature, cooling rate, holding time, and annealing temperature. The changes in these process data have a direct impact on the grain size and grain structure of the forging. For example, the starting forging temperature and the final forging temperature determine the recrystallization degree of the material during deformation, and the blanking temperature and holding time affect the kinetic process of grain growth. The deformation rate and forging pressure directly determine the deformation uniformity of the material and the degree of grain refinement, and the die temperature and cooling rate further determine the final grain size and microstructure uniformity of the forging.

[0035] The data preprocessing step is an important step before feature selection and model construction. It improves data quality, ensures data consistency and reliability by performing necessary cleaning, correction, and transformation on the original process data. Through good data preprocessing, the model can be prevented from being negatively affected by data anomalies, missing values, or redundant values. Additionally, standardize the important features selected from the dataset to ensure that all features are within the same scale range, thus avoiding the unbalanced impact of features with different scales on the prediction model.

[0036] A base regressor refers to each individual basic regression model used alone in the process of constructing an ensemble learning model. Each base regressor independently learns and predicts the input data and generates independent prediction results. Subsequently, these results are combined through an ensemble method (such as weighted summation, averaging, etc.) to improve the prediction accuracy and stability of the overall model.

[0037] The key concept behind the proposed ensemble model is to train multiple base models and aggregate their predictions. Through this process, the proposed model becomes more robust and less prone to overfitting. The ensemble nature of the model helps improve generalization and prediction performance. In the prediction phase, each machine learning base model in the ensemble independently predicts the input data. Then, the predictions of all base models are aggregated to produce the final prediction. In different embodiments, the number and types of base regressors can vary and can be determined according to actual needs. In this study, the final output of the neural network ensemble model is the weighted average of the predictions made by each base model. As an example, to achieve this, the scikit-learn Python ML library is used. The HalvingGridSearchCV class is used to determine the optimal hyperparameter configuration for four models. The four methods are described in detail below.

[0038] a. Hyperparameter Optimization of the Grain Size Prediction Model Based on the BP Neural Network A neural network can have multiple fully connected layers as hidden layers, with each input neuron connected to each output neuron. The hyperparameters considered include the number of hidden layers, the number of neurons in each hidden layer, and the type of activation function. Throughout the process, the training dataset is used to train each BP neural network. Early stopping is used to prevent the BP neural network from overfitting by training too many epochs; the validation dataset is used to detect and prevent overfitting. For each hyperparameter configuration, the mean squared error (MSE) of the validation dataset is calculated. Based on the hyperparameter configuration details of the BP neural network in existing research, this study implemented the BP neural network using the TensorFlow library and used it as a benchmark. Table 1 lists the hyperparameter combinations of the BP neural network.

[0039] Table 1 Hyperparameter Configuration of BP Neural Network b. Hyperparameter Optimization of Grain Size Prediction Model Based on Support Vector Regression (SVR) Support vector regression is used for grain size prediction. In this study, the SVR class in the scikit - learn library is used to implement support vector regression. The SVR class is an implementation of support vector regression, including various hyperparameters such as kernel type, penalty coefficient, and gamma parameter. Table 2 shows the optimal hyperparameter combination of SVR determined through the hyperparameter optimization process. Table 2 Hyperparameter Configuration of SVR c. Hyperparameter Optimization of Grain Size Prediction Model Based on Random Forest (RF) Random forest technology is used for grain size prediction. In this study, it is implemented using the RandomForestRegressor class in the scikit - learn library. This class includes various hyperparameters, such as the number of decision trees in the model, the type of function used to measure the split quality, and the maximum depth of the tree. Table 3 shows the optimal hyperparameter combination of RF determined through the hyperparameter optimization process.

[0040] Table 3 Hyperparameter Configuration of RF d. Hyperparameter Optimization of Grain Size Prediction Model Based on Ridge Regression Ridge regression, also known as Tikhonov regularization, is a technique that uses linear regression to address multicollinearity and prevent overfitting. Like LASSO, Ridge regression introduces a penalty term, denoted as L2, which is proportional to the square of the coefficients, into the traditional linear regression objective function. Different from LASSO, Ridge regression does not force the coefficients to be zero. Instead, it shrinks the coefficients towards zero, thereby reducing their magnitudes. Ridge regression can handle multicollinearity, which occurs when the independent variables in a regression model are highly correlated. Multicollinearity also leads to unstable and unreliable coefficient estimates. Ridge regression adds a penalty term to the objective function to mitigate the impact of multicollinearity by suppressing overly large coefficients. For simplicity, the linear regression model trained with L2 is named Ridge. It is implemented using the Ridge class in the scikit - learn library and is used as a benchmark method. Table 4 shows the optimal hyperparameter combination of MLR determined through the hyperparameter optimization process.

[0041] Table 4 Hyperparameter Configuration of Ridge Regression Step B. Continuously obtain the test set data formed by using the pre - processed and standardized process data, and use the test set data to calculate the collinearity values between the base regressors in real - time to dynamically evaluate the similarity between the base regressors.

[0042] In this step, by continuously obtaining the test set data and calculating the collinearity values between the individual base regressors based on these data in real - time, the similarity between the base regressors is dynamically evaluated. The collinearity value is used to measure the degree of correlation between the output results of two regressors, and its calculation formula is where c i represents the predicted value obtained after the i - th group of features in the test set is input into a regression model, represents the mean of all predicted values obtained after all groups of features in the test set are input into this regression model; d i represents the predicted value obtained after the i - th group of features in the test set is input into another regression model, represents the mean of all predicted values obtained after all groups of features in the test set are input into this regression model.

[0043] In the implementation process, as production continues, new test data will be continuously obtained. These data are input into each base regression model in real - time, so as to obtain the predicted output values corresponding to each model. Taking two base regressors as an example, such as a random forest model and a BP neural network model, assuming that for the latest obtained test set, the predicted value obtained after the i - th group of features is input into the random forest model is denoted as c i , and the mean of all prediction results of this test set input into the random forest model is Correspondingly, the predicted value obtained after the i - th group of features is input into the BP neural network model is d i , and the mean of all prediction results of the test set input into the BP neural network model is Then, according to the given collinearity formula, the collinearity value between the prediction results of the two models is calculated.

[0044] In this way, for each input of new data, the collinearity values between the base regressors will be recalculated. This real - time and dynamic evaluation method can effectively reflect the correlation fluctuations of different models with the change of data, so as to help the system identify the model combinations with high similarity of prediction results for subsequent targeted optimization of model weights.

[0045] Step C. Based on the collinearity evaluation results, dynamically determine several base regressors with the highest collinearity, and dynamically assign weights based on the real-time performance of each base regressor, and combine the determined base regressors into a weighted sum integration model as the grain size prediction model.

[0046] Specifically, in one embodiment, Step C includes the following steps C1 - C3.

[0047] C1. Take the two base regressors with the highest collinearity values as the optimal base regressors.

[0048] First, select the two regressors with the highest collinearity values from multiple heterogeneous base regressors, which are called "optimal base regressors". The so-called collinearity in this solution refers to the degree of correlation between the predicted values output by different regression models. After calculating the collinearity through the formula, it is possible to accurately determine those model combinations that show convergent prediction results for specific test data. In an actual production scenario, for example, after adjusting the forging process parameters of a certain batch of forgings, the predicted grain sizes output by the SVR model and the RF model are ASTM 8.3 and 8.5 levels respectively, while the predicted grain size of the Ridge model is 7.0 levels. This means that the SVR and RF models are more similar, so they will be selected as the optimal combination by S31 to improve the stability and reliability of the prediction.

[0049] C2. Calculate the root mean square error (RMSE) of the predicted values output by the optimal base regressors, and calculate the weights of the two optimal base regressors based on the calculated root mean square error, where w i is the weight of the i-th model, and RMSE i is the root mean square error of the i-th model.

[0050] For the above two selected optimal base regressors, perform real-time calculation of the root mean square error (RMSE). The definition of the root mean square error is a measure of the difference between the model predicted values and the true values on a given test set. This index can intuitively reflect the actual prediction accuracy of the model. The lower the RMSE value of the model, the closer the prediction effect is to the real situation. For example, taking the grain size detection results of a specific batch of forgings as the standard, if the predicted RMSE of SVR is 0.15 and the predicted RMSE of RF is 0.20, it means that the current prediction accuracy of SVR is higher. Based on the magnitude of the RMSE value, use the weight calculation formula to assign different weights to the two regressors, where the weight size is inversely proportional to the model performance, that is, the smaller the RMSE of the model, the greater the weight assigned. Taking the above example, the weight of the SVR model will be higher than that of the RF model. This dynamic weight allocation mechanism can ensure that the integration model can timely utilize the prediction advantages of the optimal model in the current state and effectively reduce the prediction error of the overall model.

[0051] C3. Use the sum of the products of the calculated weights and the output results of the base regressors as the output of the weighted summation integration model.

[0052] By multiplying the calculated dynamic weights by the output results of the two best base regressors respectively and performing a summation operation, the output of the final weighted summation integration model is obtained. This process actually represents the weighted average of the prediction results of each best base regressor, integrating the differential advantages between models and compensating for the possible biases in the prediction of a single model, thereby improving the overall robustness and accuracy of the prediction results. Still taking the prediction of the grain size of forgings as an example, assume that the SVR model predicts the grain size of the forging to be ASTM 8.3 with a weight of 0.57, and the RF model predicts it to be ASTM 8.5 with a weight of 0.43. Then the output result of the integrated model after weighted summation is 8.386, which is more accurate and stable than the prediction of any single model and can more accurately guide the forging production process.

[0053] The embodiment of the present application also discloses a method for predicting the grain size of forgings. Refer to Figures 2-4 , and this method includes the following S1 - S6.

[0054] S1. Continuously obtain several process data measured during the forging process of the forging; among them, the process data is divided into a pre - obtained part and a continuously obtained part.

[0055] Process data refers to various data indicators directly related to the production process recorded during forging production. These data are usually automatically obtained by the factory digital process acquisition system to ensure the timeliness, continuity, and accuracy of data acquisition. Specifically, it includes the initial forging temperature, final forging temperature, blanking temperature, transfer time, forging time, etc. In different embodiments, it may also include important parameters such as deformation rate, forging pressure, die temperature, cooling rate, holding time, and annealing temperature. The changes in these process data have a direct impact on the grain size and grain structure of the forging. For example, the initial forging temperature and the final forging temperature determine the recrystallization degree of the material during deformation, the blanking temperature and the holding time affect the kinetic process of grain growth, the deformation rate and the forging pressure directly determine the deformation uniformity of the material and the degree of grain refinement, and the die temperature and the cooling rate further determine the final grain size and the microstructure uniformity of the forging.

[0056] In an actual production scenario, the original process dataset obtained by this solution can be divided into two parts according to the time characteristics of data collection, namely the pre-acquisition part and the continuous acquisition part. Different from the training of conventional machine learning models, conventional models usually use a one-time fixed dataset for model training. After obtaining the internal model parameters, these parameters are in a fixed state and no longer change. If the subsequent production environment changes and new data needs to be added, the general approach is to merge the new data with the old data and then perform a complete retraining. This method obviously has the defects of large computational volume, high training cost, and low efficiency. Another method is to freeze the parameters of the fully connected layer that has been trained previously in the model and only add a new fully connected layer for training behind it. Although this method reduces the computational volume, it may lead to insufficient model training, and the parameters cannot fully adapt to the changes in the new production environment, thereby reducing the prediction accuracy and generalization ability.

[0057] To avoid the above problems, this solution realizes a hierarchical model training strategy by partitioning the dataset. Specifically, the data in the pre-acquisition part is used to initially train base regressors with high computational complexity and parameter tuning difficulty. Once the training parameters of these regressors are determined, they are highly matched with the basic process environment of the specific production line. The continuously acquired part, on the other hand, serves as dynamic input data to continuously update and fine-tune the weight parameters of the ensemble model to adapt to the possible changes in the process environment during production. In this way, both the overall computational overhead and training cost are reduced, and the dynamic fluctuations of process parameters in the actual production environment can be captured and adapted in real time, enabling the prediction model to always maintain a high prediction accuracy and generalization ability, thus achieving efficient and accurate prediction of the grain size of forgings.

[0058] As an example, in this embodiment, the process dataset of the engine nine-stage disk forging is adopted. To collect this data, data is collected at the actual production site and stored in the database. The dataset records the entire process of the forgings from blanking, billet making, die forging to grain size detection from 2021 to 2022. This dataset contains a total of 795 samples, 27 features, and the grain size is used as the label.

[0059] S2. Perform data preprocessing.

[0060] This step streamlines the original dataset through a feature selection method, screening out target features that have a significant impact on grain size prediction, thereby forming a high-quality dataset. The so-called target features are those that contribute the most to grain size prediction after the feature selection process. They have a strong internal correlation with the predicted grain size and can more effectively characterize the grain size change law of the material under specific process conditions. Relatively speaking, some parameters with less influence or high redundancy with other feature information, such as the slight temperature fluctuations generated during the simple process transfer, will be discarded after the feature selection process, thereby reducing the computational amount and avoiding the interference of invalid information on the model prediction accuracy.

[0061] The data preprocessing link is an important step before feature selection and model construction. It improves data quality, ensures data consistency and reliability by performing necessary cleaning, correction, and transformation on the original process data. Through good data preprocessing, the model can be prevented from being negatively affected by data anomalies, missing values, or redundant values.

[0062] In the present invention, the Recursive Feature Elimination (RFE) method is used to implement the feature screening process. This method first builds a model using all process parameters and evaluates the importance ranking of each feature in the model. For example, in the initial state, features such as forging temperature, forging pressure, and holding time that may have a direct impact on grain size rank among the top in terms of importance, while the temperature fluctuations during the transfer process are relatively unimportant. Subsequently, the method gradually removes features starting from the one with the lowest importance ranking. After removing one or more features each time, the model is rebuilt and evaluated until the model reaches the target state of stable performance. In this way, a reasonable and streamlined set of target features is finally formed.

[0063] Specifically, in one embodiment, S2 includes the following steps S21 - S24.

[0064] S21. Delete redundant values.

[0065] First, the redundant values existing in the dataset need to be deleted, that is, exactly the same duplicate data. This kind of redundant data usually stems from systematic problems in the data collection process, such as duplicate records caused by equipment jamming or network latency. For example, during the forging process, all parameter records such as the initial forging temperature, final forging temperature, and blanking temperature at a certain moment are exactly the same and appear multiple times. Such data should be identified and deleted to avoid the model overfitting to the repeated information and resulting in a decrease in subsequent prediction accuracy.

[0066] S22. Fill in missing values with the median.

[0067] The missing values are supplemented by the method of median filling. Specifically, if some parameters (such as forging temperature or forging pressure) in a batch of process data are not recorded or lost, the missing values can be effectively filled by statistically calculating the median of all the recorded data of this parameter. Compared with mean filling, median filling has better robustness, which can reduce the interference of extreme data on the filling result and ensure the stability of the data and the reliability of the subsequent model training process.

[0068] S23. Identify outliers using the box plot method.

[0069] To more accurately improve the data quality, it is also necessary to identify and process the outliers existing in the dataset. This step is implemented using the box plot method. The box plot method can quickly and effectively identify the outliers in the data through the upper and lower quartiles and the 1.5 times interquartile range standard. For example, the normal data of forging pressure usually concentrates in a specific range. If the forging pressure is abnormally high or low due to instrument failure during a certain measurement, it can be quickly identified by the box plot.

[0070] S24. For the values exceeding the upper and lower quartiles ± 1.5 times the interquartile range, truncate them to a preset range and fill the samples.

[0071] After identifying the outliers, it is further necessary to perform truncation operations on these data beyond the normal range. Specifically, for the data points exceeding the upper and lower quartiles ± 1.5 times the interquartile range, for example, if the billet making temperature in a certain abnormal record is significantly higher than the normal process range, it is directly truncated to the pre-set reasonable data range to prevent the extreme values of the outliers from affecting the training effect of the subsequent prediction model.

[0072] Specifically, in the embodiment of the present application, 27 features of the input dataset are preprocessed to find redundant values, outliers and missing values. The redundant values are deleted from the dataset, the outliers are adjusted and the missing values are filled to improve the quality of the input data. Then, feature selection technology is used to reduce the number of attributes, thereby upgrading the prediction model.

[0073] S3. Perform data splitting on the generated dataset, using stratified random sampling to generate a training set, a validation set and a test set; wherein, the continuously acquired part of the process data corresponds to the test set.

[0074] When performing data splitting on the generated dataset, using stratified random sampling can ensure that the data of each category are distributed in the training set, the validation set and the test set according to the same proportion. This method is particularly suitable for the situation of unbalanced category distribution, which helps to improve the generalization ability of the model and avoid training bias caused by fewer samples in some categories.

[0075] The basic implementation process of stratified random sampling includes the following key steps. First, the dataset generated after collecting the entire process data is grouped according to the categorical variable, so that each group of data only contains samples of a certain category. Then, random sampling is carried out within each category, and it is divided into a training set, a validation set, and a test set respectively according to the set ratio to ensure that the class distribution of the original dataset can still be maintained in the divided subsets.

[0076] The model is trained using the training set. If the model is evaluated on the same data that it is trained on, it may perform well on a specific dataset, but it cannot be generalized to new data and overfitting may occur; validation helps to detect and prevent this problem. In addition, the validation set allows hyperparameters to be adjusted without introducing test set bias. The test set provides an unbiased assessment of the final performance of the model, indicating how well it performs on new actual data. Generally, dividing the data into a training set, a validation set, and a test set can ensure that the machine learning model is robust, can generalize well to new data, and can be reliably executed in actual scenarios. More specifically, in the embodiments of the present application, 60% of the 795 samples are assigned to the training set, while 20% are assigned to the validation set and the test set.

[0077] S4. The recursive feature elimination method screens a number of target features and standardizes the features by removing the mean and scaling to unit variance.

[0078] Recursive feature elimination (RFE) is a wrapper feature selection technique that recursively finds the most useful subset of features for model prediction by building a model, evaluating the importance of features, and gradually removing the least important features.

[0079] Refer to Figure 5 , in the implementation process, RFE first constructs an initial prediction model using all the preprocessed features in the dataset and calculates the Gini importance scores of each feature to rank the importance of the features. As shown in Table 5, in this embodiment, the recursive feature elimination method is used to select 13 important features as input features and the grain size as the label. For the features such as the billet-making temperature, blanking length, transfer time, and forging time shown in the table, after training the model with the RFE method, if it is found that the transfer time contributes the least to the prediction performance of the model, it will be removed first, and then the remaining features are used to reconstruct the model and the importance of each feature is evaluated again. This is recursively repeated until a preset target state is reached, that is, a core feature subset that is most conducive to grain size prediction is found.

[0080] The advantage of this method lies in its ability to accurately identify the features that have the most significant impact on model prediction, thereby significantly reducing the data dimension, minimizing the interference of redundant attributes, and enhancing the prediction efficiency and accuracy of the model. Additionally, through a recursive step-by-step elimination approach, it ensures that the retained features truly have a crucial impact on grain size prediction, effectively avoiding the waste of computational resources and the decline in prediction performance caused by redundant features during model training.

[0081] Descriptive Statistics of Part of the Training Dataset Furthermore, in this step, the features are also standardized by removing the mean and scaling to unit variance. This mainly targets the important features selected from the dataset for standardization to ensure that all features are within the same scale range, thereby avoiding the unbalanced impact of features with different scales on the prediction model. Since there are often significant differences in the dimension and numerical range of feature data, for example, the billet temperature value is as high as several hundred degrees Celsius, while the transfer time and forging time are measured in seconds, this difference can easily lead to the prediction model overly focusing on features with larger scales during training, thus reducing the overall prediction accuracy.

[0082] In the specific implementation process, this step realizes the feature standardization process by removing the mean of each feature and scaling it to unit variance. Suppose the j-th sample in the dataset contains multiple feature values, such as billet temperature, transfer time, forging time, etc., a total of 13 important features, then x j = [f 1，j , f 2，j , f 3，j , f 4，j , f 5，j , f 6，j , f 7，j , f 8，j , f 9，j , f 10，j , f 11，j , f 12，j , f 13，j is the j-th sample in the dataset, where f 1，j , f 2，j , f 3，j , f 4，j , f 5，j , f 6，j , f 7，j , f 8，j , f 9，j , f 10，j , f 11，j , f 12，j , f 13，j are the feature values corresponding to the j-th sample. During the calculation process, first, the mean and standard deviation σ of each feature in the training dataset are determinedi For example, if the average value of the training data set for the blank-making temperature is 940.2 °C and the standard deviation is 90.8 °C, the standardized value of the blank-making temperature feature for any sample is calculated through the formula where, is the mean of the training samples, and σ i is the standard deviation of the training samples. Each feature processed in this way has a unified scale centered at 0 with a standard deviation of 1.

[0083] S5. Train the training set and the validation set with the above-mentioned training method of the forging grain size prediction model to obtain a dynamically adjusted weighted summation integration model.

[0084] S6. Input the real-time updated test set into the weighted summation integration model to obtain the forging grain size prediction result.

[0085] In S5 and S6, a prediction mechanism is established that is initially trained based on process data samples of a specific production line and can realize dynamic model update and optimization according to the input of subsequent real-time process data. In the specific implementation process, the process data is first processed through the previous steps to finally obtain a usable training set, validation set, and test set. Among them, the training set and the validation set are specifically used for the training and optimization of the model in the initial stage to determine the internal parameters of several heterogeneous base regressors (such as SVR, RF, Ridge, BP neural network, etc.). Once these model parameters are trained, they remain fixed and are no longer adjusted, thus ensuring that the model has a stable prediction basis.

[0086] At the same time, the test set plays a crucial continuous dynamic role. The test set is not just a one-time static data, but process parameter data continuously and dynamically obtained from the forging production process. These process parameter data usually include key parameters such as the forging temperature, forging speed, deformation amount, and cooling speed of the forging. Each time a forging goes through the forging process, these process data are collected in real time and incorporated into the test set after corresponding preprocessing procedures, and then input into the previously trained weighted summation integration model to obtain the predicted value of the forging grain size. By continuously inputting these real-time data, the collinearity between each base regressor can be calculated to evaluate the similarity of the prediction trends between regression models in real time. According to the real-time calculated collinearity results, instead of being limited to two fixed base regressors, several base regressors with the highest collinearity can be dynamically selected as the optimal base regressors, thus being able to flexibly cope with the model performance fluctuations caused by changes in process parameters or equipment conditions during the actual production process.

[0087] After the optimal basis regressors are dynamically determined, the system further calculates the prediction error metrics of each optimal basis regressor in real time, such as the root mean square error (RMSE). The RMSE can quantify the difference between the real-time prediction value of the model and the actual grain size value. The better the real-time performance of the model (the lower the RMSE), the higher the weight the system assigns to it, so that the model with the highest current prediction accuracy has the greatest impact on the final prediction result. For example, in a certain production line, when process conditions such as production line temperature, forging speed, and deformation degree change significantly in several consecutive production batches, initially the SVR and RF basis regressors may be selected due to their high collinearity and prediction accuracy. However, in subsequent batches, as the process conditions change, the BP neural network and Ridge regression may exhibit higher collinearity and lower RMSE values. At this time, the system automatically adjusts the combination of the optimal basis regressors and their respective weights to quickly adapt to the new situation of the current production line.

[0088] Optionally, the forging grain size prediction method is used for predicting the grain size of GH4169 alloy turbine forgings.

[0089] It should be understood that the magnitudes of the sequence numbers of the steps in the above embodiments do not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0090] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as Figure 6 shown. The computer device includes a processor, a memory, a network interface, and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used for data related to the training method of the forging grain size prediction model and / or the forging grain size prediction method. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a training method of a forging grain size prediction model and / or a forging grain size prediction method.

[0091] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements a training method of a forging grain size prediction model and / or a forging grain size prediction method in the above embodiments, such as Figure 1 or Figure 2 the steps shown. To avoid repetition, it will not be elaborated here.

[0092] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the training method of a forging grain size prediction model and / or the forging grain size prediction method in the above embodiments, such as Figure 1 or Figure 2 the steps shown. To avoid repetition, it will not be elaborated here.

[0093] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments of the present application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0094] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the above division of each functional unit and module is used as an example. In actual applications, the above functions can be allocated to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0095] The above embodiments are only used to illustrate the technical solutions of the present invention, not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A training method for a prediction model of the grain size of forgings, characterized in that It includes the following steps: Training several heterogeneous base regressors using the training set and validation set formed by preprocessed and standardized process data, and determining the internal parameters of each base regressor; Continuously obtaining the test set data formed by preprocessed and standardized process data, and using the test set data to calculate the collinearity value between each base regressor in real time to dynamically evaluate the similarity between the base regressors; Based on the collinearity evaluation result, dynamically determining several base regressors with the highest collinearity, and dynamically assigning weights based on the real-time performance of each base regressor, and combining the determined base regressors into a weighted sum integration model as the grain size prediction model.

2. The training method of the forging grain size prediction model according to claim 1, characterized in that The several heterogeneous base regressors include: SVR, RF, Ridge, BP neural network.

3. The training method of the forging grain size prediction model according to claim 1, characterized in that The calculation formula for the collinearity value between the base regressors is where c i represents the predicted value obtained after the i-th group of features in the test set is input into a regression model, represents the mean of all predicted values obtained after all groups of features in the test set are input into this regression model; d i represents the predicted value obtained after the i-th group of features in the test set is input into another regression model, represents the mean of all predicted values obtained after all groups of features in the test set are input into this regression model.

4. The training method of the forging grain size prediction model according to claim 1, characterized in that The step of dynamically determining several base regressors with the highest collinearity based on the collinearity evaluation result, dynamically assigning weights based on the real-time performance of each base regressor, and combining the determined base regressors into a weighted sum integration model as the grain size prediction model includes the following steps.

5. Taking the two base regressors with the highest collinearity value as the best base regressors; Calculate the root mean square error of the output prediction values of the optimal basis regressor, and calculate the weights of the two optimal basis regressors according to the calculated root mean square error, where, w i is the weight of the i-th model, and RMSE i is the root mean square error of the i-th model; Taking the sum of the products of the calculated weights and the output results of the base regressors as the output of the weighted sum integration model.

6. A method for predicting the grain size of forgings, characterized in that, It includes the following steps: S1. Continuously obtaining several items of process data measured during the forging process of the forging; wherein, the process data is divided into a pre-acquired part and a continuously acquired part; S2. Performing data preprocessing; S3. Splitting the generated data set, using stratified random sampling to generate a training set, a validation set and a test set; wherein, the continuously acquired part of the process data corresponds to the test set; S4. Screening several target features based on the recursive feature elimination method, and standardizing the features by removing the mean and scaling to unit variance; S5. Training the training set and the validation set with the training method of the forging grain size prediction model described in any one of claims 1-4 to obtain a weighted sum integration model; S6. Inputting the test set into the weighted sum integration model to obtain the forging grain size prediction result.

7. The training method of the forging grain size prediction model according to claim 5, characterized in that, The S2 includes the following steps: S21. Deleting redundant values; S22. Filling missing values with the median; S23. Identifying outliers using the box plot method; S24. For the values exceeding the upper and lower quartiles by ±1.5 times the interquartile range, truncating them to a preset range and filling the samples.

8. The forging grain size prediction method according to claim 6, characterized in that, Used for grain size prediction of GH4169 alloy turbine forgings.

9. A computer device, characterized in that, It includes: One or more processors; A memory; One or more applications, wherein the one or more applications are stored in the memory and are configured to be executed by the one or more processors, and the one or more programs are configured to: Execute the forging grain size prediction method described in any one of claims 1 to 4; And / or, execute the forging grain size prediction method described in any one of claims 5 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by a processor to implement: Executing the forging grain size prediction method according to any one of claims 1 to 4; And / or, executing the forging grain size prediction method according to any one of claims 5 to 7.

Images