Casting powder crystallization starting temperature prediction method based on interpretable machine learning

Abstract

The invention discloses a casting powder crystallization starting temperature prediction method based on interpretable machine learning, and relates to the technical field of steel continuous casting. Comprising the following steps: acquiring crystallization starting temperature data of casting powder with known components at different cooling speeds; selecting reliable data to construct a casting powder crystallization starting temperature database; dividing the data set into a training set and a test set; using the training set to construct an XGBoost integrated model; the trained XGBoost is used to predict a test set; and carrying out feature importance analysis on the trained XGBoost model by adopting the SHAP. And according to the SHAP value of each feature, the influence of different components on the crystallization starting temperature of the casting powder is judged, and importance ranking is carried out. The established model can accurately predict the crystallization starting temperature of the continuous casting covering slag with different components at different cooling speeds, and has certain robustness and generalization ability. The prediction method is scientific and reasonable, the prediction structure is reliable, and industrial application is facilitated.

Description

Technical Field

[0001] The present invention relates to the technical field of metallurgical continuous casting, and in particular to a method for predicting the crystallization start temperature of continuous casting protection slag based on machine learning. Background Art

[0002] Continuous casting is an important process in steelmaking, which can continuously convert liquid steel into solid steel, improving production efficiency and metal yield. In the process of molten steel solidification, continuous casting protection slag improves the solidification environment of the ingot in the crystallizer through insulation, lubrication, and heat transfer control, and is an indispensable auxiliary material for improving the quality of the ingot.

[0003] Considering the differences in solidification characteristics of different steel grades and process parameters such as casting speed and continuous casting temperature, in the mold slag design stage, it is often necessary to add a certain amount of CaF2, Na2O and other ingredients to the base material to change the initial crystallization temperature of the mold slag to achieve a reasonable heat control ability of the mold slag. Therefore, the initial crystallization temperature of the mold slag has become one of the important physical and chemical performance indicators that need to be considered in the mold slag design stage.

[0004] At present, researchers use thermogravimetric-insert thermal analysis, laser confocal scanning and other methods to observe the crystallization temperature of protective slag. However, due to the expensive experimental equipment and complicated experimental operation, the measurement of the crystallization temperature of protective slag requires a lot of time and money. Some researchers use single-wire / double-wire to observe the dynamic crystallization process of protective slag, but due to the presence of some special components, the protective slag is not transparent at high temperature, which is not conducive to observing the crystallization phenomenon. In addition, no quantitative formula has been found for the effect of different components on the crystallization temperature of protective slag, which makes the design of protective slag uncertain and challenging, especially when facing complex components, the design is somewhat blind. Summary of the invention

[0005] In order to address the deficiencies of the prior art, the main purpose of the present invention is to provide a method for predicting the crystallization start temperature of protective slag using machine learning, which is used to accurately predict the crystallization start temperature of protective slag under different compositions.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A method for predicting the crystallization start temperature of mold slag based on explainable machine learning, comprising the following steps:

[0008] S1. Obtaining the crystallization start temperature data of the mold slag with known composition at different cooling rates;

[0009] S2, deleting duplicate data and eliminating outliers, and using the screened data to build a mold slag crystallization start temperature database;

[0010] S3, divide the data set into training set and test set;

[0011] S4. Use the training set to build an XGBoost ensemble model, optimize the hyperparameters of the trained model, and establish a nonlinear relationship between the mold slag composition and the start crystallization temperature;

[0012] S5. Use the trained XGBoost to predict the test set, obtain the predicted values ​​of the crystallization start temperature of different mold slag components, and compare them with the true values; use the determination coefficient R 2 As a criterion, if R 2 If the XGBoost model is within the preset range, the model is judged to be qualified and used to predict the crystallization temperature of the mold slag; otherwise, it is unqualified and steps S2 to S5 are repeated;

[0013] S6. Use SHAP to perform feature importance analysis on the XGBoost model trained in S5. According to the SHAP value of each feature, determine the influence of different components on the crystallization temperature of the protective slag and rank them according to their importance.

[0014] In the present invention, the means of obtaining in S1 include consulting materials, experiments, etc.

[0015] In step S1, the protective slag contains CaO, SiO2, Al2O3; and contains Na2O, F - , Li2O, B2O3, MgO and at least one of the following:

[0016] Preferably, in step S2, the 3σ criterion is used to eliminate abnormal data. Assume that the data set is X={x1, x2, … ,x n}, then the mean and variance of the crystallization start temperature of the collected data are shown in formulas (1)-(2):

[0017]

[0018]

[0019] According to the 3σ criterion, when a certain crystallization temperature point or , they are considered as outliers and removed.

[0020] Preferably, in step S3, 50-85% of the data is selected as the training set data, and the remaining data is used as the test set data. In practical applications, 75-80% of the data is generally used as the training set data, and the remaining data is used as the test set data.

[0021] Preferably, in step S4, six hyperparameters of learning_rate, max_depth, n_estimators, subsample, reg_lambda, and gamma in the XGBoost model are selected for the hyperparameter optimization process of the particle swarm optimization algorithm.

[0022] When applied in industry, the selection range of the hyperparameter learning_rate is 0.1-0.9, preferably 0.2-0.8; the selection range of max_depth is: greater than 0 and less than or equal to 20, preferably greater than 0 and less than or equal to 10; the selection range of n_estimators is 45-210, preferably 50-200; the selection range of subsample is 0.2-1; the selection range of reg_lambda is: greater than 0 and less than or equal to 10, preferably greater than 0 and less than or equal to 5; the selection range of gamma is: greater than 0 and less than or equal to 1. The present invention controls the selection range of hyperparameters, which can bring the benefits of improving the learning ability and generalization ability of the model.

[0023] Preferably, the particle swarm optimization algorithm uses the pyswarm library, wherein: the maximum number of iterations is set to m, m is an integer between 500 and 2000, preferably an integer between 900 and 1100, which of course includes 1000; the number of particles is set to n, n is an integer between 25 and 35, preferably 28 to 32, which of course includes 30; minfunc is set to k1, k1 is a number between 0 and 1, preferably 10 -2 ~10 -5 This includes 10 -4 ; minstep is set to k2, k2 is a number greater than 0, preferably 10 -2 ~10 -5 This includes 10 -4 .

[0024] At the same time, during the hyperparameter optimization process, 5-fold cross-validation is used to improve model performance. Specifically: the initial position of the particle will be evenly distributed in the entire hyperparameter search space. After a round of iterations, each particle will update its speed and position based on its individual historical optimal position and the global historical optimal position. This process will be repeated until the termination condition is met. The speed update of the particle can be regarded as being jointly determined by inertia, self-cognition, and group cognition. The inertia part represents the tendency of the particle to maintain its previous speed, self-cognition represents the tendency of the particle to find the optimal position it has found, and group cognition represents the tendency of the particle to find the optimal position found by the group. The speed of particle i at the iteration is shown in formula (3). The update of particle position determines the step size and moving direction of the particle, and the update rule is shown in formula (4).

[0025]

[0026]

[0027] in, is the historical optimal position of the individual in the i-th generation, is the global historical optimal position of the i-th generation, t represents the number of iterations, represents the velocity vector of particle i at the next time t+1, represents the position vector of particle i at the next time t+1, represents the velocity inertia term, is an individual learning item, is the global learning item, is the inertia factor, , is the learning factor, , is a random number between (0, 1), Represents the current position vector of particle i at iteration t. In the pyswarm library All are built-in values.

[0028] Preferably, in step S5, the trained XGBoost model can not only predict the crystallization start temperature of protective slag with different cooling rates and different compositions, but also sort the features by importance based on gain value or weight value according to its own interpretability feature_importances, thereby providing additional reference for protective slag design.

[0029] Preferably, in step S5, R 2 is an evaluation indicator used to evaluate model performance, as shown in formula (5):

[0030]

[0031] in, represents the sample size, represents the measured value of the crystallization start temperature of the i-th sample, represents the predicted value of the crystallization start temperature of the i-th sample, Represents the average of the sample measured values.

[0032] Preferably, in step S6, the influence of different components on the crystallization temperature of mold slag is judged and ranked according to the average SHAP value of each feature. SHAP can more intuitively measure the correlation between features and prediction results by calculating the marginal contribution of each feature to the prediction result in the model and judging the contribution of each feature to the result. In the prediction model f(x), the SHAP value of a feature in each sample is As shown in formula (6).

[0033]

[0034] in, represents the feature subset except feature i, is the model's prediction on the feature subset, is the predicted value of the model after adding feature i, indicating the marginal contribution of feature i. is the weight of the subset S. Finally, the SHAP values ​​of each feature in all samples are compared to obtain the feature importance ranking. i Represents the value of the sample at the i-th feature. Represents the total number of features, represents the factorial of the total number of features M, represents the number of features in the subset S, Represents the number of features in subset S The factorial of It represents the factorial of the number of features remaining after removing the subset S and the current feature from the total number of features M.

[0035] In step S6, the feature is selected from all the features in the database constructed in step S2, such as the mold slag composition.

[0036] Based on S6, the contents of different components and their corresponding SHAP values ​​are obtained to determine the effects of different components on the crystallization temperature of the mold slag, which can be used as a reference for researchers to adjust the components during the mold slag design stage. At the same time, based on SHAP, the importance of different components in affecting the crystallization temperature of the mold slag is ranked, providing a basis for researchers to reasonably select the components of the mold slag during the mold slag design stage.

[0037] Compared with the prior art, the above technical solution has at least the following benefits:

[0038] The constructed model can accurately predict the crystallization start temperature of continuous casting mold slag with different compositions at different cooling rates, and has certain robustness and generalization ability. This application can significantly reduce the time loss and equipment loss caused by metallurgical workers when testing the crystallization start temperature of mold slag, and at the same time provide designers with more accurate and scientific decisions for designing mold slag with reasonable crystallization performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. The drawings are only used to illustrate the preferred embodiments and are not considered to be limitations of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work.

[0040] Figure 1 It is a principle flow chart of the present invention.

[0041] Figure 2 This is a flow chart of the particle swarm optimization algorithm used in the present invention for hyperparameter optimization.

[0042] Figure 3 Iteration diagram of the particle swarm optimization algorithm used in the present invention for the XGBoost hyperparameter optimization process.

[0043] Figure 4 This is a comparison chart of the prediction results and the measured results of the test set data of the present invention.

[0044] Figure 5 It is a ranking diagram of the SHAP values ​​of the components of the present invention. DETAILED DESCRIPTION

[0045] In order to make the technical solutions and advantages of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] Example 1

[0047] The present invention proposes a method for predicting the crystallization start temperature of mold slag based on XGBoost, which saves a lot of time, manpower and material costs compared with the existing experimental detection methods. The constructed XGBoost prediction model has good prediction accuracy and robustness, and is suitable for predicting the crystallization start temperature of a variety of different continuous casting mold slags. At present, no empirical formula or model related to the crystallization start temperature of mold slag has been found. The present invention can meet the demand for mold slag crystallization prediction under the current mold slag development.

[0048] like Figure 1 As shown, the method for predicting the starting crystallization temperature of mold slag includes the following steps:

[0049] S1: Collect the data of the crystallization start temperature of the mold slag under different cooling rates and different mold slag compositions. Among them, the key components that affect the crystallization start temperature of the mold slag may include: CaO, SiO2, Al2O3, Na2O, F - , Li2O, B2O3, MgO, and the collected components are expressed in mass percentage.

[0050] S2: The cooling rate and the mold slag composition are used as input features, and the start crystallization temperature is used as a label to build a database. The duplicated data are deleted, and the mean and variance of the start crystallization temperature are calculated using formulas (1)-(2) to eliminate abnormal data.

[0051] In step S2, the 3σ criterion is used to eliminate abnormal data. Assume that the data set is , then the mean and variance of the crystallization start temperature of the collected data are shown in formulas (1)-(2):

[0052]

[0053]

[0054] according to The criterion is that when a certain crystallization temperature point or , it is considered as an outlier and removed. i represents the i-th sample value, n represents the number of samples, S3: The mold slag crystallization start temperature database is divided into 80% training set and 20% test set.

[0055] S4: The training set divided in S3 is used to build the XGBoost start crystallization temperature prediction model, where the flow chart of the particle swarm optimization algorithm for hyperparameter optimization is as follows: Figure 2 As shown;

[0056] S41: According to the construction principle of XGBoost, the hyperparameters with greater influence are selected for optimization. The selection and range of XGBoost hyperparameters are shown in Table 1.

[0057]

[0058] S42: Construct an objective function to transform hyperparameter optimization into an optimization problem. The objective function accepts the hyperparameters of the particle as input and assigns them to the XGBoost model. The model is evaluated through five-fold cross validation to ensure that the model has a better generalization ability for data prediction. The evaluation indicator is the R² score, as shown in formula (5). Among them, since the default value of the pyswarm library pso is to find the minimum value, -R² is used as the output value of the objective function for comparing the fitness value of the particle.

[0059]

[0060] in, represents the sample size, represents the measured value of the crystallization start temperature of the i-th sample, represents the predicted value of the crystallization start temperature of the i-th sample, Represents the average of the sample measured values.

[0061] S43: The particle swarm optimization algorithm is used to search for the best XGBoost hyperparameter combination. According to the particle swarm optimization principle, the particle swarm is first initialized. Each particle represents a hyperparameter combination of the XGBoost model, and the score of the current particle is obtained by putting it into the objective function.

[0062] S44: According to the performance of each particle, the algorithm continuously iterates and updates the position and velocity of the particle swarm through formulas (3)-(4), so that the particles move closer to the global optimal position. The parameter selection of the particle swarm optimization algorithm is shown in Table 2.

[0063] S45: When the number of iterations reaches the upper limit or the change of the particle swarm meets the preset stop condition, the optimization process is stopped and the optimal hyperparameter combination and the corresponding optimal objective function value are output. The parameter selection of the particle swarm optimization algorithm is shown in Table 2. Figure 3 In this embodiment, when R 2 When the value is greater than or equal to 0.9, the optimization can be stopped.

[0064]

[0065] S45: According to the optimal hyperparameter combination for XGBoost training, the optimal XGBoost hyperparameter combination is shown in Table 3.

[0066]

[0067] S5: Compare the prediction results of the optimized XGBoost on the test set data with the measured results, also using R 2 is an evaluation indicator used to evaluate the model performance. The results are as follows Figure 4 shown.

[0068] S6. Use the SHAP algorithm to analyze the importance of features (such as mold slag composition) in the trained XGBoost model. The influence of different components on the crystallization temperature of the mold slag is judged based on the average SHAP value and the importance is ranked. Among them, SHAP calculates the marginal contribution of each feature to the prediction result in the model, judges the contribution of each feature to the result, and can more intuitively measure the correlation between the feature and the prediction result. For the prediction model , the SHAP value of a feature in each sample As shown in formula (6).

[0069]

[0070] in, represents the feature subset except feature i, is the model's prediction on the feature subset, is the predicted value of the model after adding feature i, indicating the marginal contribution of feature i. is the weight of subset S. Finally, the SHAP values ​​of each feature in all samples are compared to obtain the feature importance ranking. The results are as follows Figure 5 shown.

[0071] S7: The optimized XGBoost mold slag crystallization start temperature prediction model is used to predict the mold slag crystallization start temperature under different compositions.

[0072] For example:

[0073] In this embodiment, a continuous casting mold slag was selected and measured by a single-wire experiment. When the cooling rate was 0.9°C / s, the starting crystallization temperature was 1342°C. The XGBoost model based on the optimal hyperparameter combination predicted that the starting crystallization temperature of the mold slag was 1347°C. The influence of different components of the mold slag on the starting crystallization temperature was quantitatively analyzed by SHAP. The SHAP values ​​of different components are shown in Table 4. As can be seen from Table 4, the CaO and F in the mold slag - The SHAP values ​​corresponding to CaO, F and B2O3 are all greater than 0, indicating that compared with the established mold slag crystallization temperature database, - , B2O3 content, the crystallization start temperature of the mold slag can be effectively increased, which is beneficial to improving the crystallization performance of the mold slag. The SHAP values ​​corresponding to SiO2, Al2O3, and Na2O are all less than 0, indicating that compared with the established mold slag crystallization start temperature database, at this SiO2, Al2O3, and Na2O content, the crystallization start temperature of the mold slag can be effectively reduced, which is not conducive to improving the crystallization performance of the mold slag.

[0074]

[0075] In the embodiment of the present invention, the crystallization start temperature of the protective slag can be predicted based on the protective slag composition and cooling rate, which significantly reduces the time, manpower and material costs. The constructed XGBoost prediction model has good prediction accuracy and robustness. At present, no empirical formula or model related to the crystallization start temperature of the protective slag has been found. The present invention can meet the crystallization prediction needs of the protective slag under the current protective slag development. In addition, the visualization analysis of the constructed model through SHAP is conducive to improving the operator's understanding of the complex decision-making of the integrated model. It can guide researchers to design protective slag components with reasonable crystallization performance.

[0076] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for predicting the crystallization start temperature of mold slag based on explainable machine learning, characterized in that: The steps include: S1. Obtaining the crystallization start temperature data of the mold slag with known composition at different cooling rates; S2, deleting duplicate data and eliminating outliers, and using the screened data to build a mold slag crystallization start temperature database; S3, divide the data set into training set and test set; S4. Use the training set to build an XGBoost ensemble model, optimize the hyperparameters of the trained model, and establish a nonlinear relationship between the mold slag composition and the start crystallization temperature; S5. Use the trained XGBoost to predict the test set, obtain the predicted values ​​of the crystallization start temperature of different mold slag components, and compare them with the true values; use the determination coefficient R 2 As a criterion, if R 2 If the XGBoost model is within the preset range, the model is judged to be qualified and used to predict the crystallization temperature of the mold slag; otherwise, it is unqualified and steps S2 to S5 are repeated; S6. Use SHAP to perform feature importance analysis on the XGBoost model trained in S5. Determine the influence of different components on the crystallization temperature of the protective slag based on the average SHAP value of each feature and rank the importance.

2. The method for predicting the starting crystallization temperature of mold slag based on explainable machine learning according to claim 1, characterized in that: In step S1, the protective slag contains CaO, SiO2, Al2O3; and contains Na2O, F - , Li2O, B2O3, MgO and at least one of the following:

3. The method for predicting the starting crystallization temperature of mold slag based on explainable machine learning according to claim 1, characterized in that: In step S2, the 3σ criterion is used to eliminate abnormal data. Assume that the data set is X={x1, x2, … ,x n }, the mean and variance of the crystallization start temperature of the collected data are shown in formulas (1)-(2): ； ； According to the 3σ criterion, when a certain crystallization temperature point or , they are considered as outliers and removed.

4. The method for predicting the mold slag crystallization start temperature based on explainable machine learning according to claim 1, characterized in that: In step S3, 50-85% of the data are selected as training set data, and the remaining data are selected as test set data.

5. The method for predicting the mold slag crystallization start temperature based on explainable machine learning according to claim 1, characterized in that: In step S4, six hyperparameters, namely, learning_rate, max_depth, n_estimators, subsample, reg_lambda, and gamma, in the XGBoost model are selected for the hyperparameter optimization process of the particle swarm optimization algorithm; The selection range of the hyperparameter learning_rate is 0.1~0.9, preferably 0.2~0.8; the selection range of max_depth is: greater than 0 and less than or equal to 20, preferably greater than 0 and less than or equal to 10; the selection range of n_estimators is 45~210, preferably 50~200; the selection range of subsample is 0.2~1; the selection range of reg_lambda is: greater than 0 and less than or equal to 10, preferably greater than 0 and less than or equal to 5; the selection range of gamma is: greater than 0 and less than or equal to 1.

6. The method for predicting the mold slag crystallization start temperature based on explainable machine learning according to claim 5, characterized in that: The particle swarm optimization algorithm uses the pyswarm library, where: the maximum number of iterations is set to m, m is an integer between 500 and 2000, preferably an integer between 900 and 1100, which of course includes 1000; the number of particles is set to n, n is an integer between 25 and 35, preferably 28 to 32, which of course includes 30; minfunc is set to k1, k1 is a number between 0 and 1, preferably 10 -2 ~10 -5 This includes 10 -4 ; minstep is set to k2, k2 is a number greater than 0, preferably 10 -2 ~10 -5 This includes 10 -4 ; At the same time, during the hyperparameter optimization process, 5-fold cross validation was used to improve model performance.

7. The method for predicting the mold slag crystallization start temperature based on explainable machine learning according to claim 6, characterized in that: The initial positions of the particles are evenly distributed in the entire hyperparameter search space. After a round of iterations, each particle will update its speed and position based on its individual historical optimal position and the global historical optimal position. This process will be repeated until the termination condition is met. The particle movement process is shown in formulas (3)-(4): ； ； in, is the historical optimal position of the individual in the i-th generation, is the global historical optimal position of the i-th generation, is the inertia factor, , is the learning factor, , is a random number between (0, 1).

8. The method for predicting the mold slag crystallization start temperature based on explainable machine learning according to claim 1, characterized in that: In step S5, R 2 is an evaluation indicator used to evaluate model performance, as shown in formula (5): ； in, represents the sample size, represents the measured value of the crystallization start temperature of the i-th sample, represents the predicted value of the crystallization start temperature of the i-th sample, Represents the average of the sample measured values.

9. The method for predicting the mold slag crystallization start temperature based on explainable machine learning according to claim 1, characterized in that: In step S6, the influence of different components on the crystallization temperature of mold slag is judged and the importance is ranked according to the average SHAP value of each feature; SHAP calculates the marginal contribution of each feature to the prediction result in the model, and judges the contribution of each feature to the result, which can more intuitively measure the correlation between the feature and the prediction result; in the feature set In the prediction model f(x), the SHAP value of a feature in each sample is As shown in formula (6); ； in, represents the feature subset except feature i, is the model's prediction on the feature subset, is the predicted value of the model after adding feature i, indicating the marginal contribution of feature i. is the weight of subset S; finally, the SHAP values ​​of each feature in all samples are compared to obtain the feature importance ranking.

Images