High manganese steel longitudinal crack process control method based on isolated forest and extreme random tree

Abstract

The invention discloses a high manganese steel longitudinal crack process control method based on an isolated forest and an extreme random tree. According to the method, process parameters and crack defect data in the high manganese steel continuous casting process are collected, a longitudinal crack database is constructed, and abnormal working condition data are removed through an isolated forest algorithm. Then, an extreme random tree algorithm is used for establishing a nonlinear relation between process parameters and longitudinal crack generation, and model hyper-parameters are optimized through a Bayesian optimization algorithm; after training is completed, a feature module of the model is adopted to carry out importance ranking on process parameters, an SHAP method is utilized to carry out feature dependency analysis on the process parameters which are high in ranking and controllable, and the influence of key process parameters on crack formation is quantified. According to the method, data-driven process optimization can be realized, the quality of the high-manganese steel casting blank is improved, the occurrence rate of longitudinal cracks is reduced, and an efficient and interpretable intelligent process control method is provided for the metallurgical industry.

Description

Technical Field

[0001] The present invention relates to the technical field of metallurgical continuous casting, and particularly to a method for controlling the longitudinal crack process of high manganese steel based on isolation forest and extremely randomized trees. Background Art

[0002] High manganese steel is widely used in high-strength structural materials due to its excellent mechanical properties. However, during continuous casting, longitudinal cracks are likely to occur in the billet, affecting product quality and production stability. Currently, crack control mainly relies on empirical adjustment of process parameters such as drawing speed, cooling intensity, superheat degree, etc., but this method is difficult to accurately and quickly reduce the risk of longitudinal cracks in the target steel grade.

[0003] In recent years, most research scholars have used numerical simulation and experimental research methods to analyze the crack formation mechanism during continuous casting and explore the influence of process parameters on crack defects. For example, patent application CN202411634313.2 reduces the occurrence of billet cracks by adjusting the narrow face taper cooling water flow rate and casting superheat degree of the chamfered mold. However, such methods often have high experimental costs and are difficult to update in real time. Currently, with the wide application of machine learning technology in the metallurgical field, data-driven methods provide new solutions for crack prediction and optimization. Using machine learning methods, the key factors affecting crack formation can be mined based on historical production data, and process parameters can be optimized through data-driven methods. Summary of the Invention

[0004] To solve the deficiencies of the prior art, an embodiment of the present invention provides a method for controlling the longitudinal crack process of high manganese steel based on isolation forest and extremely randomized trees, which can be used for visual analysis by combining the isolation forest and extremely randomized tree algorithms, accurately identify the key process parameters affecting crack formation, and optimize the process parameters in combination with on-site production experience to reduce the crack occurrence rate.

[0005] The technical solution of the embodiment of the present invention is realized as follows: An embodiment of the present invention provides a method for controlling the longitudinal crack process of high manganese steel based on isolation forest and extremely randomized trees, and the method includes: S1. Collect high manganese steel process parameters and crack defect data from the production line; S2. After deleting the missing data and eliminating the abnormal working condition data from the high manganese steel process parameters and crack defect data, construct a high manganese steel longitudinal crack database; S3. Divide the data in the database into a training set and a test set; S4. Use the training set to construct an extremely randomized tree ensemble model and perform hyperparameter optimization on the trained model to establish a non-linear relationship between process parameters and the occurrence of longitudinal cracks; S5, the trained model is used for test set testing to obtain the predicted value of whether longitudinal cracks occur under different process parameters, and the predicted value is compared with the true value; and the area under the curve is used as the judgment standard. If the area under the curve is within the preset range, the trained model is considered qualified, and the model is used for the prediction of longitudinal cracks in high manganese steel; otherwise, the trained model is considered unqualified, and steps S2 to S5 are repeated; S6. Perform feature importance analysis on the trained model and rank the importance; S7, selecting a preset number of continuous casting process parameters that are ranked high in importance and are easy to adjust on site, and using a visualization algorithm to perform feature dependency analysis on the selected parameters to obtain a parameter value range that is beneficial to reducing the occurrence of cracks in the ingot; S8. Adjust the process parameters on site according to the parameter value range obtained in step S7 in combination with prior knowledge to control the occurrence of longitudinal cracks in the ingot.

[0006] In one embodiment, in step S1, the process parameters include casting speed, nozzle insertion depth, superheat and type of protective slag.

[0007] In one embodiment, in step S2, an isolation forest machine learning algorithm is used to remove abnormal data: Assume the data set is ,sample The average path length in a forest is: The normalized anomaly score is: in, Representative samples In the The path length in the tree, is the total number of trees in the forest, is the sample size The expected path length under is used to normalize so that The value is between 0 and 1; when When it is greater than the set threshold, the sample is considered an outlier and removed.

[0008] In one embodiment, in step S3, 75% to 85% of the data are selected as training set data, and the remaining data are selected as test set data.

[0009] In one embodiment, in step S4, four hyperparameters in the model, namely max_leaf_nodes, max_features, max_depth, and n_estimators, are selected for the hyperparameter optimization process of the Bayesian optimization algorithm. Among them, the selection range of the hyperparameter max_leaf_nodes is 200 to 2000; the selection range of the hyperparameter max_features is 0.2 to 1; the selection range of the hyperparameter max_depth is 1 to 500; the selection range of the hyperparameter n_estimators is 50 to 2000; and the Bayesian optimization is used to optimize the hyperparameters of the model.

[0010] In one embodiment, in step S5, according to the self-explainability of the trained model, the features are sorted according to the gain value or the weight value.

[0011] In one embodiment, in step S6, according to the preset number of casting process parameters with relatively high importance ranking and easy to adjust on-site selected in step S5, SHAP is used to perform feature dependence analysis on them. Among them, using SHAP to perform feature dependence analysis on them includes: by calculating the marginal contribution of each process parameter to whether longitudinal cracks occur in the model, judging the contribution of different feature values of this feature to the result, and quantitatively analyzing the correlation between the process parameter and the longitudinal crack from the perspective of machine learning; when the SHAP value corresponding to the range of a certain process parameter is greater than 0, it is considered that the range of this process parameter is conducive to the occurrence of longitudinal cracks; when the SHAP value corresponding to the range of a certain process parameter is less than 0, it is considered that the range of this process parameter is not conducive to the occurrence of longitudinal cracks.

[0012] The method of this embodiment has the following beneficial effects: The method of this embodiment can, on the basis of accurately predicting whether longitudinal cracks occur in the slab, quantitatively analyze the influence of key process parameters on the occurrence of longitudinal cracks in the slab, and obtain the optimal process parameter range for on-site engineers to quickly and accurately adjust the process to reduce the occurrence of longitudinal cracks in high manganese steel. The method of this embodiment can significantly reduce the blindness of smelting engineers in optimizing process parameters, and at the same time reduce the time and economic losses caused by trial and error, providing a data-driven solution for the optimization of the high manganese steel continuous casting process. At the same time, the method of this embodiment has universality and can also provide a reference for reducing longitudinal cracks in other steel grades, promoting the application of intelligent manufacturing in the metallurgical field. Description of the Drawings

[0013] Figure 1 It is a schematic flowchart of the method for controlling the longitudinal crack process of high manganese steel based on the isolation forest and the extremely randomized tree according to the embodiment of the present invention; Figure 2 It is a schematic diagram of the implementation process of the embodiment of the present invention; Figure 3 It is a schematic diagram of the importance ranking of process parameters affecting the longitudinal cracking of high manganese steel obtained based on the interpretability of the extreme random tree model itself in the embodiments of the present invention; Figure 4 It is a schematic diagram of the analysis of the characteristic dependence diagram of the mold powder parameters based on the SHAP value in the embodiments of the present invention; Figure 5 It is a schematic diagram of the analysis of the characteristic dependence diagram of the superheat degree parameters based on the SHAP value in the embodiments of the present invention; Figure 6 It is a schematic diagram of the analysis of the characteristic dependence diagram of the casting speed parameters based on the SHAP value in the embodiments of the present invention. Specific embodiments

[0014] The present invention will be further described in detail below in conjunction with the drawings and embodiments.

[0015] The embodiments of the present invention provide a method for controlling the longitudinal crack process of high manganese steel based on the isolated forest and the extreme random tree, as Figure 1 shown, the method includes: S1. Collect the process parameters of high manganese steel and crack defect data from the production line; S2. After deleting the missing data and eliminating the abnormal working condition data from the process parameters of high manganese steel and crack defect data, construct a longitudinal crack database of high manganese steel; S3. Divide the data in the database into a training set and a test set; S4. Use the training set to construct an extreme random tree ensemble model, and optimize the hyperparameters of the trained model to establish a non-linear relationship between the process parameters and the occurrence of longitudinal cracks; S5. Use the trained model to test the test set, obtain the predicted values of whether longitudinal cracks occur under different process parameters by the model, and compare them with the true values; and use the area under the curve as the evaluation criterion. If the area under the curve is within the preset range, it is considered that the trained model is qualified, and the model is used for the prediction of longitudinal cracks in high manganese steel; otherwise, it is considered that the trained model is unqualified, and steps S2 to S5 are performed again; S6. Perform feature importance analysis on the trained model and perform importance ranking; S7. Select a preset number of continuous casting process parameters with a high importance ranking and easy to adjust on site, and use a visualization algorithm to perform feature dependence analysis on the selected parameters to obtain the parameter value range conducive to reducing the occurrence of billet cracks; S8. Adjust the on-site process parameters according to the parameter value range obtained in step S7 in combination with prior knowledge to control the occurrence of longitudinal cracks in the billet.

[0016] The method of this embodiment can, on the basis of accurately predicting whether longitudinal cracks occur in the continuous casting billet, quantitatively analyze the influence of key process parameters on the occurrence of longitudinal cracks in the continuous casting billet, and obtain the optimal process parameter range for on-site engineers to quickly and accurately adjust the process to reduce the occurrence of longitudinal cracks in high manganese steel. The method of this embodiment can significantly reduce the blindness of metallurgical engineers in optimizing process parameters, and at the same time reduce the time and economic losses caused by trial and error, providing a data-driven solution for the optimization of the continuous casting process of high manganese steel. At the same time, the method of this embodiment has universality, and can also provide reference for reducing longitudinal cracks in other steel grades, promoting the application of intelligent manufacturing in the metallurgical field.

[0017] Specifically, referring to Figure 2 , in actual application, the implementation process of this embodiment can be completed as follows: S1: Collect process parameters and crack defect data of high manganese steel from the production line. It can include: specific composition of the steel grade, casting speed, submerged nozzle insertion depth, superheat, type of mold powder, etc.

[0018] S2: Take parameters such as the specific composition of the steel grade, casting speed, submerged nozzle insertion depth, superheat, type of mold powder, etc. as input features, and use binary coding of whether longitudinal cracks occur as labels to construct a database. Delete the repeatedly collected data, and calculate the average path length and standardized anomaly score of the samples using formulas (1)-(2) in the forest, and eliminate the abnormal data.

[0019] Here, the isolation forest machine learning algorithm can be used to eliminate the abnormal data. That is, a forest is constructed by randomly selecting features and cutting points, and the path length of the sample point in the tree is calculated. The shorter the path of the sample, the more likely it is to be an outlier.

[0020] Assume the data set is , then the average path length of the sample in the forest is shown in formula (1), and the standardized anomaly score is shown in formula (2): (1) (2) Where, represents the path length of the sample in the th tree, is the total number of trees in the forest, is the number of samples under the expected path length, which is used for normalization so that takes values between 0 and 1; when is greater than the set threshold, the sample is regarded as an outlier and eliminated.

[0021] Through the above-mentioned Isolation Forest algorithm, abnormal process parameter data can be effectively eliminated, the quality of the data set can be improved, and thus the accuracy and stability of the extremely randomized trees prediction model can be enhanced.

[0022] S3: Divide the database of longitudinal cracks in high manganese steel into a training set of 85% and a test set of 15%.

[0023] S4: The training set of longitudinal cracks in high manganese steel divided in S3 is used to train the extremely randomized trees model.

[0024] S41: Select the hyperparameters with greater influence according to the construction principle of the extremely randomized trees for adjusting their parameter values. The selection and range of the hyperparameters of the extremely randomized trees are shown in Table 1: S42: Use Bayesian optimization to optimize the hyperparameters of the extremely randomized trees with AUC as the evaluation index. 10-fold cross-validation is adopted during each optimization iteration to improve the generalization ability of the model. Bayesian optimization constructs a probabilistic surrogate model and uses the expected improvement criterion to search for the optimal hyperparameters. When the number of iterations meets the preset stopping condition, stop the optimization process and output the optimal hyperparameter combination and the corresponding optimal objective function value; S43: Use the optimal hyperparameter combination for training the extremely randomized trees. The optimal hyperparameter combination of the extremely randomized trees is shown in Table 2: S5: Compare the prediction results of the optimized extremely randomized trees on the test set data with the measured results. Also use AUC as the evaluation index to evaluate the performance of the model. If AUC≥0.8, it is considered that the constructed extremely randomized trees model is qualified and is used for subsequent feature importance ranking and feature dependence analysis.

[0025] S6: Conduct feature importance analysis and importance ranking on the qualified extremely randomized trees model after being trained in S5, as Figure 3 shown.

[0026] When ranking, based on the interpretability of the trained model itself, the features can be ranked according to the gain value or the weight value.

[0027] S7: Based on the importance ranking results obtained in S6, it is found that the mold powder composition, superheat degree, and drawing speed are the continuous casting process parameters with relatively high importance rankings and are easy to adjust on site. Use the visualization algorithm to conduct feature dependence analysis on the selected parameters to obtain the parameter value range that is beneficial to reducing the occurrence of slab cracks. As Figures 4 - 6 shown.

[0028] Here, by calculating the marginal contribution of each process parameter to the occurrence of longitudinal cracks in the extremely randomized tree, the contribution of different eigenvalue of this feature to the result can be evaluated, and the correlation between process parameters and longitudinal cracks can be quantitatively analyzed from the perspective of machine learning. When the SHAP value corresponding to the range of a certain process parameter is greater than 0, it is considered that the range of this process parameter is conducive to the occurrence of longitudinal cracks. When the SHAP value corresponding to the range of a certain process parameter is less than 0, it is considered that the range of this process parameter is not conducive to the occurrence of longitudinal cracks.

[0029] S8: Adjust the on-site process parameters through the results obtained in S7 and combined with prior knowledge to control the occurrence of longitudinal cracks in the continuous casting billet.

[0030] Comparative Example 1 During a period of time, a total of 3 different types of mold fluxes, namely No. 1, No. 2, and No. 3, were successively used in the production of high manganese steel by a certain continuous casting machine in this steel plant. Since the process parameters were continuously adjusted during the replacement of the mold flux, it was impossible to directly determine the optimal mold flux selection from the produced continuous casting billets. Among the 30 continuous casting billets investigated along the casting direction, the average number of longitudinal cracks per billet was 29.

[0031] Example 1 According to the importance ranking results of the extremely randomized tree, it is found that the type of mold flux is the most important parameter affecting the occurrence of longitudinal cracks in high manganese steel. Using it for the feature dependence analysis of SHAP, it is found that even though the process parameters are constantly changing, when using No. 3 mold flux for the production of high manganese steel, the corresponding SHAP values are mostly less than 0 and are overall lower than the SHAP values corresponding to the use of No. 1 and No. 2 mold fluxes. This indicates that this parameter range has an inhibitory effect on the occurrence of longitudinal cracks in high manganese steel and is conducive to the production of continuous casting billets that meet the quality requirements. It is determined to select No. 3 mold flux for continuous casting production. Among the 30 continuous casting billets investigated along the casting direction, the average number of longitudinal cracks per billet was 18. The number of longitudinal cracks decreased, and most of them were micro-cracks, and the quality of the continuous casting billets was improved.

[0032] Comparative Example 2 During a period of time, during the production of high manganese steel by a certain continuous casting machine in this steel plant, the superheat range was between 12°C and 35°C, and the casting speed was between 0.75 m / min and 0.95 m / min. Due to the large fluctuations in process parameters, the quality stability of the continuous casting billets was poor, with large fluctuations. When using No. 3 mold flux for the production of high manganese steel, among the 30 continuous casting billets investigated along the casting direction, the average number of longitudinal cracks per billet was 18.

[0033] Example 2 According to the importance ranking results of extremely randomized trees, it is found that the superheat and casting speed are the main parameters affecting longitudinal cracks in high manganese steel except for the mold powder. Using them for the feature dependence analysis of SHAP, when the superheat is around 15°C to 19°C and the casting speed is 0.8 m / min to 0.86 m / min, the corresponding SHAP values are mostly less than 0, which indicates that this parameter range has an inhibitory effect on the generation of longitudinal cracks in high manganese steel and is beneficial to producing billets that meet the quality requirements. It is determined to select No. 3 mold powder, control the superheat at about 17°C, and control the casting speed at about 0.83 m / min. Among the total 30 billets, the average number of longitudinal cracks per billet is 13. The number of longitudinal cracks decreases, and the quality of the billets is improved.

[0034] It should be noted that the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent in such process, method, commodity or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, commodity or device including the element.

[0035] The above are only the embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A method for process control of longitudinal cracks in high manganese steel based on isolation forest and extreme random tree, characterized in that: The method comprises: S1. Collect high manganese steel process parameters and crack defect data from the production line; S2. After deleting missing data and eliminating abnormal working condition data from the high manganese steel process parameters and crack defect data, a high manganese steel longitudinal crack database is constructed; S3, dividing the data in the database into a training set and a test set; S4. Using the training set to construct an extreme random tree ensemble model, and performing hyperparameter optimization on the trained model to establish a nonlinear relationship between process parameters and the occurrence of longitudinal cracks; S5, the trained model is used for test set testing to obtain the predicted value of whether longitudinal cracks occur under different process parameters, and the predicted value is compared with the true value; and the area under the curve is used as the judgment standard. If the area under the curve is within the preset range, the trained model is considered qualified, and the model is used for the prediction of longitudinal cracks in high manganese steel; otherwise, the trained model is considered unqualified, and steps S2 to S5 are repeated; S6. Perform feature importance analysis on the trained model and rank the importance; S7, selecting a preset number of continuous casting process parameters that are ranked high in importance and are easy to adjust on site, and using a visualization algorithm to perform feature dependency analysis on the selected parameters to obtain a parameter value range that is beneficial to reducing the occurrence of cracks in the ingot; S8. Adjust the process parameters on site according to the parameter value range obtained in step S7 in combination with prior knowledge to control the occurrence of longitudinal cracks in the ingot.

2. The method for process control of high manganese steel longitudinal cracks based on isolation forest and extreme random tree according to claim 1 is characterized in that: In the step S1, the process parameters include casting speed, nozzle insertion depth, superheat and type of protective slag.

3. The method for process control of high manganese steel longitudinal cracks based on isolation forest and extreme random tree according to claim 1 is characterized in that: In step S2, an isolation forest machine learning algorithm is used to remove abnormal data: Assume the data set is ,sample The average path length in a forest is: The normalized anomaly score is: in, Representative samples In the The path length in the tree, is the total number of trees in the forest, is the sample size The expected path length under is used to normalize so that The value is between 0 and 1; when When it is greater than the set threshold, the sample is considered an outlier and removed.

4. The method for process control of high manganese steel longitudinal cracks based on isolation forest and extreme random tree according to claim 1 is characterized in that: In step S3, 75% to 85% of the data are selected as training set data, and the remaining data are selected as test set data.

5. The method for process control of high manganese steel longitudinal cracks based on isolation forest and extreme random tree according to claim 1 is characterized in that: In the step S4, four hyperparameters, namely max_leaf_nodes, max_features, max_depth and n_estimators, are selected in the model for the hyperparameter optimization process of the Bayesian optimization algorithm; wherein the selection range of the hyperparameter max_leaf_nodes is 200-2000; the selection range of the hyperparameter max_features is 0.2-1; the selection range of the hyperparameter max_depth is: 1-500; the selection range of the hyperparameter n_estimators is 50-2000; and Bayesian optimization is used to optimize the hyperparameters of the model.

6. The method for process control of high manganese steel longitudinal cracks based on isolation forest and extreme random tree according to claim 1 is characterized in that: In step S5, the features are ranked in importance based on gain values ​​or weight values ​​according to the interpretability of the trained model itself.

7. The method for process control of high manganese steel longitudinal cracks based on isolation forest and extreme random tree according to claim 1 is characterized in that: In the step S6, SHAP is used to perform feature dependency analysis on the preset number of continuous casting process parameters selected in step S5 that are ranked high in importance and are easy to adjust on site; wherein, the feature dependency analysis using SHAP includes: by calculating the marginal contribution of each process parameter to whether longitudinal cracks occur in the model, judging the contribution of different eigenvalues ​​of the feature to the result, and quantitatively analyzing the correlation between the process parameters and the longitudinal cracks from the perspective of machine learning; when the SHAP value corresponding to the range of a certain process parameter is greater than 0, it is considered that the range of the process parameter is conducive to the occurrence of longitudinal cracks; when the SHAP value corresponding to the range of a certain process parameter is less than 0, it is considered that the range of the process parameter is not conducive to the occurrence of longitudinal cracks.

Images