Method for controlling steel plate tempering parameters based on machine learning and computer program product

Abstract

A method for controlling tempering parameters of steel plates based on machine learning includes: reading historical tempering datasets; training and establishing a tempering property prediction model corresponding to a property type based on the historical tempering datasets; using the tempering property prediction model to predict the difference in property before and after the first tempering under preset tempering parameters, so as to calculate the predicted property value after the first tempering; determining whether the predicted property value after the first tempering exceeds the specified property of the property type; if so, using the tempering property prediction model to predict and calculate the predicted property value after the second tempering under multiple candidate tempering parameters respectively; then calculating the target score corresponding to multiple candidate tempering parameters based on the property target of the property type using a property target function; and determining the optimal value of a tempering parameter from multiple candidate tempering parameters based on the target score.

Description

Technical Field

[0001] This disclosure relates to a method and computer program product for controlling the tempering parameters of steel plates, and in particular to a method and computer program product for predicting the properties of steel plates after tempering based on machine learning and optimizing the control of tempering parameters of steel plates. Prior Technology

[0002] During the production of steel plates, tempering heat treatment can be used to adjust properties to meet customer requirements. Tempering heat treatment utilizes the energy of temperature rise to enhance the activity of atoms in the steel, thereby rearranging and recombining them to adjust the steel plate's properties. This can include increasing toughness, adjusting hardness, improving mechanical properties, and stabilizing the microstructure. Specifically, it gradually transforms the steel plate from an unstable, unbalanced structure to a stable, balanced structure, reducing strength and increasing ductility to achieve the target strength and tensile strength. However, during the tempering process, due to multiple process variables, it is often difficult to quickly and accurately obtain the tempering parameters that allow the steel plate to meet the required mechanical properties. Summary of the Invention

[0003] This disclosure provides a machine learning-based method and computer program for controlling tempering parameters of steel plates, which helps to systematize the control of tempering parameters of steel plates and improves the speed and accuracy of tempering parameter adjustment, thereby stabilizing the quality of steel plate products.

[0004] In one embodiment, a method for adjusting tempering parameters of steel plates based on machine learning is provided, comprising: reading a historical tempering dataset, the historical tempering dataset including historical tempering parameters and historical differences in properties before and after tempering; training and establishing a tempering property prediction model corresponding to a property type based on the historical tempering dataset; predicting the difference in properties before and after tempering under preset tempering parameters using the tempering property prediction model to calculate the predicted value of the properties after the first tempering; determining whether the predicted value of the properties after the first tempering exceeds the standard properties of the property type; if so, predicting the difference in properties before and after tempering under multiple candidate tempering parameters using the tempering property prediction model to calculate the predicted value of the properties after the second tempering; calculating the target score corresponding to multiple candidate tempering parameters using a property target function based on the predicted value of the properties after the second tempering and the property target of the property type; and determining the optimal value of a tempering parameter from multiple candidate tempering parameters based on the target score.

[0005] In one embodiment, the mechanical property type includes at least one of yield strength and tensile strength.

[0006] In one embodiment, the tempering performance prediction model is trained using ensemble learning.

[0007] In one embodiment, the model used for ensemble learning is determined based on the results of K-fold cross-validation evaluation.

[0008] In one embodiment, the model used in ensemble learning includes at least one of a limit stochastic tree regression model and a linear regression model.

[0009] In one embodiment, each of the historical tempering parameters, preset tempering parameters, and candidate tempering parameters is at least one of the following: steel plate size, steel plate composition, tempering conditions, and rolling conditions.

[0010] In one embodiment, the target of the mechanism is the median of the normalized mechanism or the average value of the training mechanism data of the tempering mechanism prediction model.

[0011] In one embodiment, a computer program product is provided, which stores a program that, when loaded into a computer, executes the following: It reads a historical tempering dataset from a historical database, the historical tempering dataset containing historical tempering parameters and historical differences in mechanical properties before and after tempering; it trains and establishes a tempering mechanical property prediction model corresponding to a mechanical property type based on the historical tempering dataset; it uses the tempering mechanical property prediction model to predict the difference in mechanical properties before and after a first tempering under preset tempering parameters, thereby calculating the predicted mechanical property value after the first tempering; it then determines whether the predicted mechanical property value after the first tempering exceeds the specified mechanical property of the mechanical property type; if so, it uses the tempering mechanical property prediction model to predict the difference in mechanical properties before and after a second tempering under multiple candidate tempering parameters, thereby calculating the predicted mechanical property value after the second tempering; and based on the predicted mechanical property value after the second tempering and the mechanical property target of the mechanical property type, it calculates the target score corresponding to multiple candidate tempering parameters using a mechanical property target function, and determines the optimal value of a tempering parameter from multiple candidate tempering parameters based on the target score.

[0012] In one embodiment, the tempering performance prediction model is trained using ensemble learning.

[0013] In one embodiment, each of the historical tempering parameters, preset tempering parameters, and candidate tempering parameters is at least one of the following: steel plate size, steel plate composition, tempering conditions, and rolling conditions. Simple Explanation of the Diagram

[0014] To gain a more complete understanding of the embodiments and their advantages, the following description, taken in conjunction with the accompanying drawings, is provided, wherein: Figure 1 is a flowchart of the steel plate tempering parameter control method based on machine learning according to an embodiment of this disclosure; Figure 2 shows one example of the Delta value prediction results using a single extreme random tree regression model; and Figure 3 shows one example of the Delta value prediction results using an ensemble learning model; Figure 4 is a flowchart for finding the optimal value of tempering parameters using an objective function. Implementation

[0015] The embodiments disclosed herein are discussed in detail below. However, it will be understood that the embodiments provide many applicable concepts that can be implemented in a wide variety of specific contexts. The discussed and disclosed embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. The terms "first," "second," etc., used herein are not descriptions of sequence or order, but are used only to distinguish elements or operations described by the same technical terms.

[0016] In the tempering heat treatment process of steel plates, tempering parameters are affected by multiple process variables, including the steel plate's dimensions, composition, tempering conditions, and rolling conditions. Ultimately, the post-tempering properties of the steel plate must meet the customer's specifications, requiring precise control of the tempering parameters. This disclosure provides a machine learning-based method for controlling steel plate tempering parameters. Using ensemble learning in machine learning, a predictive model corresponding to various tempering parameters and post-tempering properties is established, and the optimal values ​​for tempering parameter control are found by combining this model with a tempering objective function.

[0017] Figure 1 is a flowchart of the machine learning-based steel plate tempering parameter control method 100 of this embodiment. The machine learning-based steel plate tempering parameter control method 100 is described as follows. First, step S110 is performed to read the historical tempering dataset. The historical tempering dataset is composed of various historical tempering parameters corresponding to various steel plates, as well as the corresponding historical steel plate properties before and after tempering, the difference in historical steel plate properties before and after tempering Delta, etc., after data cleaning to remove abnormal and unreasonable values. These data are set as X and Y values ​​as needed, so as to establish a tempering property prediction model based on the correlation between X and Y. Among them, the tempering parameters may include multiple serial process variables, such as: steel plate size such as steel plate thickness; steel plate composition such as carbon, silicon, manganese, etc.; tempering conditions such as tempering temperature, tempering time; rolling conditions such as furnace temperature, furnace time, initial cooling temperature, final cooling temperature, cooling rate, etc. The properties before and after tempering may be, for example, the yield strength and tensile strength of the steel plate before and after tempering.

[0018] Table 1 is an example of one of the historical tempering datasets: (Table 1) Where Y represents the difference in mechanical properties before and after tempering, Delta can be the difference in tensile strength before and after tempering, while X can be any tempering parameter that may affect the tempering process, such as tensile strength before tempering, steel plate thickness, tempering temperature, tempering time, furnace temperature, furnace time, initial cooling temperature, final cooling temperature, cooling rate, and the content of carbon, silicon, and manganese in the steel plate composition.

[0019] Next, in step S120, a tempering probability prediction model is established based on the X and Y values ​​of the aforementioned historical tempering dataset, corresponding to different types of tempering parameters and the difference in probability before and after tempering (Delta). In step S120, the machine learning used in model building is an ensemble learning heterogeneous model machine learning algorithm. By integrating different machine learning models, the overall prediction accuracy of the ensemble learning model is improved, and the overfitting of the model is reduced, making the prediction trend smoother and more reasonable.

[0020] In this embodiment, an optimization method is introduced to search for the optimal combination of model parameters that yields the best model metrics for K-fold cross-validation. For example, Bayesian optimization is used to find the model parameters with the lowest mean absolute percentage error (MAPE) in 5-fold cross-validation. Furthermore, 5-fold cross-validation is used to evaluate various machine learning algorithms. Training and prediction are performed on training and testing datasets taken from historical tempering data. For example, the Delta difference in performance before and after tempering under different tempering parameters is predicted, and the model with the smallest mean absolute percentage error (MAPE) is selected as the ensemble learning model. The model used for ensemble learning can be a decision tree-based model, such as Extra Trees Regressor, CatBoost Regressor, Random Forest Regressor, or any combination thereof, but is not limited to these.

[0021] Furthermore, the ensemble learning method used in this embodiment employs stacking. The base learner uses superior models selected through the aforementioned evaluation, such as the Limiting Random Tree Regression model and the Lasso model (Minimum Absolute Value Convergence and Selection Formula). To avoid overfitting, ensure reasonable prediction trends, and broadly apply to the prediction of tempering parameter data points not present in the training dataset, the meta learner uses a linear regression model. The ensemble models used in ensemble learning are not limited to these, as long as they improve prediction accuracy, reduce overfitting, and ensure reasonable prediction trends.

[0022] Figures 2 and 3 show examples of prediction results for the Delta value of the probability difference before and after tempering using a single extreme random tree regression model and the ensemble learning model of this embodiment, respectively, with training datasets. X is the tempering parameter, and Y is the Delta value of the probability difference before and after tempering. Comparing the trends of prediction points in ranges F2 and F3, in the case where there is no X=55 and its corresponding Y value in the training dataset, as shown in Figure 2, when the extreme random tree regression model is used alone for prediction, the prediction results near X=55 show a trendless horizontal phenomenon. That is, when this model is used for prediction, it cannot correctly consider the correlation between X and Y values ​​for prediction when the input tempering parameter does not exist in the training dataset. As shown in Figure 3, if the tempering probability prediction model established by the ensemble learning model of this embodiment is used for prediction, the linear regression model of the meta-learner can be used to further generate the final prediction based on the prediction results generated by the base learner, so as to improve the unreasonable phenomenon of the predicted value being horizontal when the value changes.

[0023] Next, in step S130, the trained ensemble learning model is used as a tempering performance prediction model. Before the steel plate enters the tempering heat treatment, known tempering parameters such as the preset steel plate size, steel plate composition, tempering conditions, rolling conditions, and pre-tempering performance are input to predict the difference in performance before and after tempering, Delta. Then, the pre-tempering performance is added to the predicted value of the difference in performance before and after tempering, Delta, to obtain the predicted value of the post-tempering performance. Then, in step S140, it is determined whether the predicted value of the post-tempering performance of the steel plate under the current preset tempering parameters exceeds the customer's required performance specifications. If it does not exceed the requirements, the process proceeds to step S150, maintaining the original preset tempering parameter settings, and proceeding to step S170 for tempering heat treatment. If the requirements exceed the requirements, the process proceeds to step S160, where an optimal value for the tempering parameters is found using a performance objective function, and this optimal value is used as the new tempering parameter for the tempering heat treatment in step S170.

[0024] Figure 4 is a flowchart of step S160, in which the optimal value of the tempering parameter is found using the objective function of the mechanical property. First, step S410 is performed, where the tempering mechanical property prediction model for each type of mechanical property, established in step S120, predicts the difference in mechanical property before and after tempering (Delta) under different candidate tempering parameters for each type of mechanical property. Then, step S420 is performed, where the difference in mechanical property before and after tempering (Delta) is added to the corresponding mechanical property before tempering to obtain the predicted value of the mechanical property after tempering. Next, proceed to step S430, where the predicted values ​​of the above-mentioned mechanical properties after tempering are... Substitute the objective of each mechanism into the objective function to generate the objective score corresponding to different candidate tempering parameters.

[0025] The objective function of the mechanism is as shown in equation (1): (1) in, , For the number of machine types; The weights of each performance metric are assigned to adjust the importance of each performance metric in compliance. This is the predicted value of mechanical properties after tempering; For the objective of mechanism, it can be the median of the mechanism specification or the average value of the training mechanism data used during the training of the tempering mechanism prediction model; The standard deviation of the training data for the tempering property prediction model is given. In equation (1), the predicted tempering property values ​​for each property type under multiple candidate tempering parameters are given respectively. With machine target The absolute value of the difference divided by The values ​​are then summed to obtain the target score for each candidate tempering parameter corresponding to the mechanical properties.

[0026] Finally, in step S440, the candidate tempering parameter corresponding to the smallest applicable target score is selected as the optimal tempering parameter value.

[0027] Table 2 shows an example of the target score corresponding to each candidate tempering time calculated using the mechanical objective function: (Table 2) In this example, tempering time is used as a candidate tempering parameter, but practical applications are not limited to this; it can be any tempering parameter that has an influence on the tempering heat treatment process, such as other steel plate dimensions, steel plate composition, tempering conditions, rolling conditions, etc. In this example, The value is set to 2, and the mechanical properties are yield strength YS and tensile strength TS, respectively. , The yield strength YS and tensile strength TS are the median values ​​of their respective mechanical properties. , The weights of yield strength YS and tensile strength TS are both set to 1; and the mechanical target of yield strength YS is... The mechanical target for tensile strength TS is 517. The value is 645. The explanation of this example is as follows: First, using the tempering probability prediction model for yield strength YS and the tempering probability prediction model for tensile strength TS, the difference in yield strength YS before and after tempering (Delta 1) and the difference in tensile strength TS before and after tempering (Delta 2) are predicted for different candidate tempering times (corresponding to step S410). Then, these are added together with the corresponding pre-tempering probability to obtain the post-tempering probability prediction value of yield strength YS. And the predicted values ​​of tensile strength TS after tempering and mechanical properties. (corresponding to step S420), then substitute the above parameters into the objective function of equation (1) to calculate the objective score for each candidate tempering time (corresponding to step S430), and finally sort these objective scores from low to high, and determine the lowest and most applicable objective score from the corresponding candidate tempering times (corresponding to step S440). For example, in this embodiment, the tempering time is 70 minutes, which is the optimal value for tempering time.

[0028] Based on the above process, the optimal value of the tempering parameters can be determined. Then, as shown in Figure 1, the tempering heat treatment in step S170 is performed with the optimal value of the tempering parameters. The predicted value of the tempered properties, the corresponding target score, and the optimal value of the tempering parameters obtained under each candidate tempering parameter in the above process can all be stored in the database for future reference when needed.

[0029] This disclosure discloses a machine learning-based method for controlling tempering parameters in steel plates and a computer program product. It utilizes ensemble learning in machine learning to obtain a tempering property prediction model composed of multiple models, which is then used to predict the post-tempering properties of steel plates under specific tempering parameters. If the predicted post-tempering properties exceed the customer's required specifications, the tempering parameters are optimized to obtain the optimal values. The method provided in this disclosure helps to systematize the control of tempering parameters in steel plates, improving the speed and accuracy of parameter adjustments and stabilizing the quality of steel plate products.

[0030] Although this disclosure has been disclosed above with reference to embodiments, it is not intended to limit this disclosure. Anyone skilled in the art may make some modifications and refinements without departing from the spirit and scope of this disclosure. Therefore, the scope of protection of this disclosure shall be determined by the appended claims.

[0031] 100: A Machine Learning-Based Method for Adjusting Tempering Parameters in Steel Plates S110, S120, S130, S140, S150, S160, S170: Steps S410, S420, S430, S440: Steps F2, F3: Range

Claims

1. A method for controlling tempering parameters of steel plates based on machine learning, executed by a computer and comprising: reading a historical tempering dataset, the historical tempering dataset containing historical tempering parameters and historical differences in tempering properties before and after tempering; training and establishing a tempering property prediction model corresponding to a property type based on the historical tempering dataset; The difference in mechanical properties before and after the first tempering under preset tempering parameters is predicted using the tempering mechanical property prediction model, so as to calculate the predicted value of the mechanical properties after the first tempering. And determine whether the predicted value of the first tempering property exceeds the standard property of the property type. If so, use the tempering property prediction model to predict the difference in property before and after the second tempering under multiple candidate tempering parameters, calculate the predicted value of the second tempering property, and calculate the target score corresponding to the multiple candidate tempering parameters using a property target function based on the predicted value of the second tempering property and the property target of the property type. Then, determine the optimal value of a tempering parameter from the multiple candidate tempering parameters based on the target score.

2. The method for controlling the tempering parameters of steel plates based on machine learning as described in claim 1, wherein the type of machine learning includes at least one of yield strength and tensile strength.

3. The method for controlling the tempering parameters of steel plates based on machine learning as described in Request 1, wherein the training of the tempering property prediction model is ensemble learning.

4. The machine learning-based method for controlling tempering parameters of steel plates as described in claim 3, wherein the model used in the ensemble learning is determined based on the results of K-fold cross-validation.

5. The method for controlling the tempering parameters of steel plates based on machine learning as described in claim 3, wherein the model used in the ensemble learning includes at least one of a limit stochastic tree regression model and a linear regression model.

6. The machine learning-based steel plate tempering parameter control method as described in claim 1, wherein each of the historical tempering parameter, the preset tempering parameter, and the candidate tempering parameter is at least one of steel plate size, steel plate composition, tempering conditions, and rolling conditions.

7. The machine learning-based method for controlling the tempering parameters of steel plates as described in claim 1, wherein the target property is the median of the specified property or the average value of the training property data of the tempering property prediction model.

8. A computer program product, which stores a program that, when loaded into a computer, executes the following: reading a historical tempering dataset from a historical database, the historical tempering dataset containing historical tempering parameters and the difference in tempering properties before and after historical tempering; and training and establishing a tempering property prediction model corresponding to the property type based on the historical tempering dataset. The difference in mechanical properties before and after the first tempering under preset tempering parameters is predicted using the tempering mechanical property prediction model, so as to calculate the predicted value of the mechanical properties after the first tempering. And determine whether the predicted value of the first tempering property exceeds the standard property of the property type. If so, use the tempering property prediction model to predict the difference in property before and after the second tempering under multiple candidate tempering parameters, calculate the predicted value of the second tempering property, and calculate the target score corresponding to the multiple candidate tempering parameters using a property target function based on the predicted value of the second tempering property and the property target of the property type. Then, determine the optimal value of a tempering parameter from the multiple candidate tempering parameters based on the target score.

9. The computer program product as described in claim 8, wherein the tempering performance prediction model is trained by ensemble learning.

10. The computer program product as described in claim 8, wherein each of the historical tempering parameter, the preset tempering parameter, and the candidate tempering parameter is at least one of the following: steel plate size, steel plate composition, tempering conditions, and rolling conditions.

Images