Method for predicting amount of residual molten steel during slag tapping of steel ladle by using machine learning

Abstract

The invention discloses a method and system for predicting the amount of residual molten steel during slag tapping of a steel ladle through machine learning, and the method comprises the steps: S1, sampling the water simulation environment data of a laboratory through SMOGN, and carrying out the preprocessing of the sampled water simulation data, and obtaining the preprocessing data; s2, constructing a plurality of machine learning models based on the preprocessed data, and evaluating the plurality of machine learning models based on a decision coefficient, a mean square error and a mean absolute error; and S3, based on the evaluation result, a machine learning model is selected to predict the residual molten steel amount when the steel ladle with or without the wall side circular ring is subjected to slag tapping, and a residual molten steel amount prediction result is obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of iron and steel metallurgy and steelmaking, and in particular to a method for predicting the amount of remaining molten steel when slag is discharged from a ladle by using machine learning. Background Art

[0002] In the production process of clean steel, some enterprises have failed to fully realize the harm of slag discharge at the end of ladle pouring and when the liquid level of the tundish is too low. Due to the excessive emphasis on improving the recovery rate of molten steel, the amount of slag discharged from the ladle is too large. When the tundish capacity is small and the liquid level is low, the ladle slag is easily drawn into the tundish and eventually enters the crystallizer, forming slag inclusions in the ingot, which partially or completely offsets the purification effect of metallurgical technology. In order to prevent slag from entering the tundish, the ladle slag detection technology is currently mainly used. Although it significantly reduces the amount of slag entering the tundish and purifies the molten steel, it also leads to a decrease in the recovery rate of molten steel. The slag detected at the water inlet is often brought in by the vortex effect, which means that there is still a certain amount of molten steel in the ladle that has not been discharged, resulting in a waste of resources. Therefore, the harm of ladle slag discharge to the production of clean steel cannot be ignored. It is necessary to conduct in-depth research on the ladle slag discharge process and take measures to control the amount of slag discharge, so as to improve the cleanliness of molten steel and the recovery rate of molten steel. Through these measures, we can improve production efficiency and resource utilization while ensuring the quality of steel and achieve sustainable development of the steel industry.

[0003] Different simulation methods play an important role in studying the ladle slag process, which can help researchers and engineers better understand complex metallurgical phenomena and optimize production processes. Physical simulation is to simulate the actual ladle slag process by building a scaled-down model. Water model experiment is the most commonly used physical simulation method. It uses water and slag-like media to simulate the behavior of molten steel and slag based on the principle of similarity. This simulation can study the flow characteristics of the liquid in the ladle, the interaction between the slag-metal interface, the floating and separation effects of inclusions, etc. Numerical simulation is to analyze the fluid flow, heat transfer and mass transfer in the ladle slag process through computer simulation. Through numerical simulation, the slag effect under different operating conditions can be predicted without actual experiments, thereby guiding process optimization. Different simulation methods have their own advantages in studying the ladle slag process. Physical simulation can help understand macroscopic phenomena and processes, while numerical simulation can provide detailed fluid mechanics and thermodynamics information. The comprehensive use of these simulation methods can provide a more comprehensive understanding and optimization of the ladle slag process.

[0004] However, traditional simulation methods have some limitations when studying the ladle slag discharge process. Physical simulation experiments usually take a long time to design and execute, and may involve high experimental costs. In addition, physical simulations usually do not have the ability to adjust dynamically in real time, and their results are often limited to specific experimental settings and conditions. In contrast, machine learning, which has emerged in recent years, can make up for these shortcomings. Once a machine learning model is trained, it can perform large-scale predictions and analyses at a relatively low cost, greatly saving manpower, material resources and time costs. Machine learning can also provide rapid responses and optimization suggestions in combination with real-time data. At the same time, machine learning has better generalization capabilities and can be applied to a wider range of situations and parameter ranges. In recent years, the application of machine learning technology in the steel production process has covered multiple links such as ironmaking, steelmaking, casting and rolling. Taking the continuous casting process as an example, in the continuous casting of steel, machine learning technology plays a vital role in monitoring and controlling complex processes. Since the continuous casting process involves physical, mechanical and chemical components, it is a very complex process, and machine learning technology can provide effective monitoring and control means. Although machine learning has advantages in many aspects, it also has limitations, such as relying on high-quality data and possibly lacking intuitive explanations of physical processes. Therefore, in practical applications, machine learning and physical simulation can complement each other and jointly improve the understanding and optimization of the ladle slag discharge process. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a method for predicting the amount of molten steel remaining when the ladle is unloading slag by using machine learning, and the method comprises:

[0006] Step S1, using SMOGN to sample the water simulation environment data of the laboratory, and preprocessing the sampled water simulation data to obtain preprocessed data;

[0007] Step S2: construct several machine learning models based on the preprocessed data, and evaluate the several machine learning models based on decision coefficient, mean square error and mean absolute error;

[0008] Step S3: Based on the evaluation results, a machine learning model is selected to predict the amount of molten steel remaining when the ladle is unloading slag for ladles with or without wall side rings, and a prediction result of the amount of molten steel remaining is obtained.

[0009] Optionally, in step S1, the construction of the water simulation environment specifically includes:

[0010] A physical model of a ladle with a similarity ratio of 1:5 was established using acrylic glass, in which the height of step 1 was always twice that of step 2. Water was injected into the ladle to a height of 400 mm. After the water injection, a layer of oil with a thickness of 20 mm was covered and allowed to stand for 30 minutes to allow the flow field in the ladle to stabilize, thus completing the construction of the water simulation environment.

[0011] Optionally, the water simulation data after sampling includes: ladle height, ladle upper diameter, ladle lower diameter, side wall step height, side wall step thickness, bottom step height, suppression dam length, suppression dam thickness and water inlet diameter.

[0012] Optionally, in step S1, the sampled water simulation data is preprocessed to obtain the preprocessed data, which specifically includes:

[0013] Two partitions are constructed according to the target variable value: BinsR partition and BinsN partition; the data points of the BinsN partition are sparsely distributed, and the data points of the BinsN partition are densely distributed; random undersampling is used to process the data points contained in the BinsN partition, and oversampling is used to process the partition in the Bins R;

[0014] Based on each data point in the BinsR partition, the Euclidean distance between the current data point and the remaining water simulation data points in the BinsR partition is calculated, and the safety distance is calculated based on the calculation result of the Euclidean distance; the water simulation data is classified by the K nearest neighbor algorithm, and the number of nearest neighbors is set to 2;

[0015] If the distance between the current data point and the nearest neighbor is within the "safe distance", SmoteR is used to generate a new synthetic sample. If the distance between the current data point and the nearest neighbor exceeds the "safe distance", Gaussian noise is introduced to generate a new sample.

[0016] The classified data is normalized to obtain preprocessed data.

[0017] Optionally, in step S2, the several machine learning models constructed based on the preprocessed data specifically include: LASSO model, SVR model, ElasticNet model, MLP model and XGBoost model.

[0018] Optionally, in step S2, based on the decision coefficient R 2 , mean square error MSE and mean absolute error MAE to evaluate several machine learning models, including:

[0019]

[0020] Among them, y i and The measured and predicted values ​​of the remaining water volume respectively; is the average value of the measured value of the remaining water volume; n is the total number of sample data;

[0021] R 2 The value range is [0,1], based on the decision coefficient R2 , MSE values ​​and MAE values ​​are used to evaluate several machine learning models.

[0022] The present invention also discloses a system for predicting the amount of molten steel remaining when the ladle is unloading slag by using machine learning, the system comprising: a preprocessing module, a model selection module and a molten steel amount prediction module;

[0023] The preprocessing module is used to sample the water simulation environment data of the laboratory using SMOGN, and preprocess the sampled water simulation data to obtain preprocessed data;

[0024] The model selection module is used to construct several machine learning models based on the preprocessed data, and evaluate the several machine learning models based on the decision coefficient, mean square error and mean absolute error;

[0025] The molten steel quantity prediction module is used to select a machine learning model based on the evaluation result to predict the remaining molten steel quantity when the ladle is unloading slag for the ladle with or without the wall side ring, and obtain the remaining molten steel quantity prediction result.

[0026] Optionally, in the pre-processing module, the construction of the water simulation environment specifically includes:

[0027] A physical model of a ladle with a similarity ratio of 1:5 was established using acrylic glass, in which the height of step 1 was always twice that of step 2. Water was injected into the ladle to a height of 400 mm. After the water injection, a layer of oil with a thickness of 20 mm was covered and allowed to stand for 30 minutes to allow the flow field in the ladle to stabilize, thus completing the construction of the water simulation environment.

[0028] Compared with the prior art, the present invention has the following beneficial effects: (1) The method of the present invention oversamples the collected laboratory data through SMOGN, establishes five machine learning models to predict the residual water content of the water model data, selects the optimal calculation model and predicts and guides the actual production; (2) The method of the present invention can accurately and conveniently predict the residual molten steel volume when the ladle is unloading slag, thereby improving industrial production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0030] Figure 1 This is a schematic diagram of the water simulation environment structure of an embodiment of the present invention;

[0031] Figure 2 The statistical results and distribution diagram of the input variables of the embodiment of the present invention;

[0032] Figure 3 This is a comparison diagram of data distribution before and after oversampling in an embodiment of the present invention;

[0033] Figure 4 This is a comparison chart of prediction errors of five models in the embodiment of the present invention;

[0034] Figure 5 A comparison chart of the predicted remaining water volume and the actual remaining water volume by XGBoost according to an embodiment of the present invention;

[0035] Figure 6 This is a graph showing the results of a single variable impact analysis in an embodiment of the present invention;

[0036] Figure 7 This is a schematic diagram of comparing the influence of variables based on the SHAP average value according to an embodiment of the present invention;

[0037] Figure 8 This is a comparison chart of the dual variable impact analysis of an embodiment of the present invention;

[0038] Fig. 9 This is a method step diagram of a method for predicting the amount of molten steel remaining when slag is discharged from a ladle using machine learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0039] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] Embodiment 1

[0041] A method for predicting the amount of molten steel remaining when the ladle is unloading slag using machine learning, such as Fig. 9 As shown, the method includes:

[0042] Step S1, using SMOGN to sample the water simulation environment data of the laboratory, and preprocessing the sampled water simulation data to obtain preprocessed data.

[0043] The laboratory water simulation data was sampled and preprocessed using SMOGN. In this embodiment, a ladle of a domestic steel plant was used as a prototype, and a ladle physical model with a similarity ratio of 1:5 was established using acrylic plexiglass (such as Figure 1 As shown, Figure 1 (a) is a water model structure with a sidewall ring. Figure 1(b) is a water model structure without a sidewall ring), where the height of step 1 is always twice that of step 2. To simulate the process of slag entering the tundish due to the vortex in the ladle: we injected water into the ladle to 400 mm, then covered its surface with a layer of oil with a thickness of 20 mm, and let it stand for more than 30 minutes to stabilize the flow field in the ladle; then opened the water inlet, and used a high-speed camera to observe the process from the appearance of the oil layer vortex at the water inlet to the vortex penetrating the water inlet (oil discharged from the water inlet), and used the video frame by frame to find the photos of edible oil entering the water inlet, read the height corresponding to the remaining water in the ladle at this time, and further calculated the remaining water in the ladle.

[0044] Table 1

[0045]

[0046] In order to obtain the optimal bottom structure of the ladle, various parameters including the nozzle diameter, nozzle height, step height, suppression dam height, distance between step and nozzle, and the presence or absence of side wall rings were changed to simulate various experimental conditions. Three repeated experiments were carried out under each condition, and the average value was taken as the experimental result. The data required for the prediction model of the remaining molten steel volume were collected in the laboratory. Among them, 48 data with side wall rings, the input variables include nozzle diameter, nozzle height, step height difference, and suppression dam height. There are 69 data without side wall rings, and the input variables include nozzle diameter, nozzle height, step height difference, and the distance between the nozzle and the step. The statistical results and distribution of each input variable are shown in the figure. Figure 2 As shown. Since the data were collected in laboratory research experiments, and for the cases with and without sidewall rings, Figure 2 (a) is the data distribution diagram of the sidewall ring. Figure 2 (b) is the data distribution of the ring without sidewalls. There are only four input variables in each, so no further variable selection is performed in this paper. However, further analysis of the data shows that there are fewer data when the residual water volume is high in the collected samples. This imbalance may cause the model to perform poorly in predicting this situation. Therefore, the SMOGN package is used to pre-process the data by oversampling. If the distance between the current data point (given observation value) and the nearest neighbor point is within the "safe distance", SmoteR is used to generate a new synthetic sample. If the distance between the current data point and the nearest neighbor point exceeds the "safe distance", Gaussian noise is introduced to generate a new sample. For each point in the partition, the safe distance is a fixed value. If there are n points in BinsR, then after selecting the observation point, the Euclidean distance between it and the other n-1 points is calculated, and these distance values ​​are arranged in order of size. The median of this set of distance values ​​is selected and multiplied by one-half, which is the "safe distance" for the selected observation point.

[0047] Figure 3 Comparison of the distribution of output variables before and after oversampling. Figure 3 (a) is the distribution change diagram of the residual water volume in the side wall ring. Figure 3 (b) is the distribution change diagram of the residual water content of the ring without side wall. After oversampling, the data with higher residual water content in the sample is increased, and the number of data points in each interval is more balanced.

[0048] Finally, the oversampled data is normalized to [0, 1], and the two sets of data are divided into training set and test set according to 0.7:0.3. The data changes before and after oversampling and the number of data sets are shown in Table 2. In order to avoid the random influence of data partitioning, this embodiment uses the same training set and test set for different models.

[0049] Table 2

[0050] Side wall ring The amount of raw data After oversampling Number of training sets Number of test sets have 48 86 60 26 none 69 128 90 38

[0051] Step S2: construct several machine learning models based on the preprocessed data, and evaluate the several machine learning models based on the decision coefficient, mean square error and mean absolute error.

[0052] Different machine learning algorithms were established for the presence or absence of sidewall rings to predict the amount of remaining water when the water model was unloaded. In order to obtain a more accurate model, this embodiment established five machine learning models based on the presence or absence of sidewall rings, including LASSO, SVR, ElasticNet, MLP and XGBoost. The model was then further analyzed using the decision coefficient R 2 , mean square error MSE and mean absolute error MAE are three evaluation indicators to measure the prediction accuracy of the model. The formulas of the three indicators are defined as follows:

[0053]

[0054]

[0055] Among them, y i and The measured and predicted values ​​of the remaining water volume respectively; is the average value of the measured value of the remaining water volume; n is the total number of sample data; R 2 The value range is [0,1], based on the decision coefficient R 2 , MSE and MAE values ​​are used to evaluate several machine learning models. 2 The higher the , the lower the MSE and MAE values ​​are, and the better the model fit is.

[0056] The LASSO, SVR, ElasticNet, MLP, and XGBoost models used in the study were implemented based on Python's scikit-learn library and were run on a 64-bit, 12th Gen Intel(R) Core(TM) i9-12900H@2.50GHz Win11 system with 16Gb of RAM.

[0057] Compare with the actual results and select the optimal machine learning model. Figure 4 The comparison of prediction errors of the five models shows the R 2 , MSE and MAE, where Figure 4 (a) is R 2 , Comparison of prediction errors of MSE, MAELASSO and SVR for circular rings with side walls, Figure 4 (b) is R 2 , MSE, MAELASSO and SVR prediction error comparison chart for the ring without sidewall. 2 , the prediction effects of the models are not significantly different. According to Tables 3 and 4, except for LASSO and SVR with sidewall rings and LASSO without sidewall rings, the R 2 All reached above 0.95. For MSE and MAE, when predicting the two cases, the prediction errors of LASSO and SVR were large, especially the SVR model with sidewall rings, whose MSE and MAE values ​​were much larger than those of the other four models. MLP, ElasticNet and XGBoost all achieved good results, with MSE and MAE values ​​less than 1. Among them, the error value of the XGBoost model in predicting the two cases was closer to zero than that of other models, and R 2 The closer it is to 1, the better the prediction performance of the XGBoost model.

[0058] Table 3

[0059] algorithm <![CDATA[R 2 ]]> MSE MAE LASSO 0.88133 1.84874 1.2222 SVR 0.89518 1.63289 0.93448 MLP 0.98548 0.22621 0.37852 ElasticNet 0.99042 0.14925 0.29225 XGBoost 0.99716 0.0442 0.16745

[0060] Table 4

[0061] algorithm <![CDATA[R 2 ]]> MSE MAE SVR 0.84653 7.24873 2.07575 LASSO 0.96513 1.6468 1.06304 MLP 0.98509 0.7042 0.65255 ElasticNet 0.98599 0.66178 0.63391 XGBoost 0.99826 0.08226 0.21774

[0062] Figure 5 The figure shows the comparison between the predicted values ​​and the measured values ​​of the XGBoost model for the test set. The closer the points are to the diagonal line, the higher the prediction accuracy. Figure 5 (a) is the comparison diagram of the prediction with sidewall ring. Figure 5(b) is a comparison chart of the prediction of the ring without sidewalls. As can be seen from the figure, XGBoost has achieved high accuracy in the prediction of both cases, and the data points are basically distributed on the diagonal, indicating that the predicted value of the test set is almost the same as the actual value. Within the relative error range of 10%, the hit rate of both cases is 100%; when the relative error range is 5%, the hit rate of the ring with sidewalls is 88.46%, and the hit rate of the ring without sidewalls is 92.31%. The difference in the hit rate of the two cases is analyzed as the difference in the number of data. The number of data without the ring with sidewalls is larger. If the amount of data can be appropriately increased, the prediction performance of the model can be further improved.

[0063] Step S3: Based on the evaluation results, a machine learning model is selected to predict the amount of molten steel remaining when the ladle is unloading slag for ladles with or without wall side rings, and a prediction result of the amount of molten steel remaining is obtained.

[0064] The machine learning model is used to predict and guide the amount of molten steel remaining when the ladle is unloading slag in the actual production process. The influence of each input variable on the amount of residual water in the ladle is studied based on the XGBoost model. The single variable principle is adopted to change the value range of the variable under study, and the other input variables take the mode. The values ​​of different variables are shown in Tables 5 and 6.

[0065] Table 5

[0066]

[0067] Table 6

[0068]

[0069] Figure 6 The single variable effects of the experiment and prediction are compared. The red dotted line in the figure is the predicted value, and the black scattered points are the experimental values. Figure 6 (a) is the effect diagram of the four single variables of the water inlet height, step height, suppression dam height and water inlet diameter on the variables with side wall ring. Figure 6 (b) is a diagram showing the effects of the four single variables, namely, the sprue height, step height, suppression dam height and sprue diameter, on the variables without sidewall rings. As can be seen from the figure, for the case with sidewall rings, except for the predicted value corresponding to the suppression dam height which remains high, the predicted values ​​of the other three variables and the experimental values ​​are basically consistent in their effects on the remaining water volume; for the case without sidewall rings, except for the predicted value of the distance between the step and the sprue which is always higher than the experimental value, the two values ​​of the other three variables are also basically the same, and the change trends of the predicted value corresponding to the distance between the step and the sprue and the experimental value are roughly the same, both of which decrease first and then tend to be flat. In order to compare the contribution of each characteristic to the model, this paper adopts the Lundberg method.

[16] The SHAP (SHapley Additive explained) method introduced by et al. determines the impact of features on the ML model by changing the model input and calculates the SHAP average of each variable as its contribution value. The calculation results are shown in Figure 7 As shown in the figure, Figure 6 The two variables whose predicted values ​​differ greatly from the experimental values ​​contribute less to the model, that is, the model cannot thoroughly capture the substantial relationship between the variable and the predicted results, resulting in a large gap between the predicted values ​​and the experimental values.

[0070] Figure 6 (a) shows the effect of four input variables on the residual water volume in the ladle slag water simulation with sidewall rings. It can be seen that the lower the nozzle height and the higher the step height difference within a certain range, the less residual water volume, that is, the less molten steel is wasted in the ladle. The experimental results show that the influence of the suppression dam on the residual water volume is positively correlated, and the nozzle diameter is significantly reduced between 30 and 40 mm. Figure 6 (b) shows the influence of the four input variables when there is no sidewall ring. It can be seen from the figure that when there is no sidewall ring, the influence of the nozzle height, step height difference and nozzle diameter on the residual water volume is basically the same as that when there is a sidewall ring. As the distance between the step and the nozzle decreases, the residual water volume decreases first and then tends to be stable after 100mm.

[0071] for Figure 6 There are two main reasons for the formation of the ladder shape. The first is the structure of the model. As the most basic sub-model of XGBoost, the decision tree needs to divide the input data layer by layer according to the indicators to build the model. However, due to the limitations of the experimental data, the distribution of the training data is not dense and compact enough, that is, there are not enough data points. This is the second reason for the ladder shape. When predicting the data in the gaps, the model failed to divide the data during training, so when predicting the data, they were placed on the same "leaf". Each ladder represents a "leaf" formed by the integration of multiple decision trees.

[0072] Select Figure 7 The features with the top two contribution values ​​are used for bivariate prediction analysis. Figure 7 (a) is a comparison chart of the effect of SHAP average value on the variables with sidewall rings. Figure 7 (b) is a comparison of the effect of the SHAP average value on the variables of the ring without sidewalls. Figure 8 shown. Figure 8 (a) is a comparison diagram of the influence of double variables on the variables with sidewall rings. Figure 8(b) is a comparison chart of the impact of the double variables on the variable without sidewall rings. In the case of sidewall rings, when the step height difference is greater than 11mm and the sprue height is less than 12mm, the remaining water volume can be controlled below 7 liters. When the sprue height exceeds 24mm, the remaining water volume can reach more than 14 liters. For the case without sidewall rings, when the step height difference is greater than 11mm and the sprue height is less than 11mm, the remaining water volume is less than 10 liters. When the sprue height exceeds 26mm, the remaining water volume will reach more than 20 liters. Comparing the two situations, it can be seen that when the sprue height and the step height difference remain unchanged, the remaining water volume with sidewall rings is less than that without sidewall rings, which means that more molten steel can be saved during the production process.

[0073] Embodiment 2

[0074] A system for predicting the amount of molten steel remaining when the ladle is unloading slag using machine learning, the system comprising: a preprocessing module, a model selection module and a molten steel amount prediction module;

[0075] The preprocessing module is used to sample the water simulation environment data of the laboratory using SMOGN, and preprocess the sampled water simulation data to obtain preprocessed data.

[0076] The laboratory water simulation data was sampled and preprocessed using SMOGN. In this embodiment, a steel ladle of a domestic steel plant was used as a prototype, and a steel ladle physical model with a similarity ratio of 1:5 was established using acrylic plexiglass (such as Figure 1 As shown), the height of step 1 is always twice that of step 2. To simulate the process of slag entering the tundish due to the vortex in the ladle: water was injected into the ladle to 400 mm, and then a layer of oil with a thickness of 20 mm was covered on its surface. The ladle was left to stand for more than 30 minutes to stabilize the flow field in the ladle. Then the water inlet was opened, and a high-speed camera was used to observe the process from the appearance of the oil layer vortex at the water inlet to the vortex penetrating the water inlet (oil discharged from the water inlet). The video was used to find the photos of edible oil entering the water inlet frame by frame, and the height corresponding to the remaining water in the ladle was read, and the remaining water in the ladle was further calculated.

[0077] In order to obtain the optimal bottom structure of the ladle, various parameters including the nozzle diameter, nozzle height, step height, suppression dam height, distance between step and nozzle, and the presence or absence of side wall rings were changed to simulate various experimental conditions. Three repeated experiments were carried out under each condition, and the average value was taken as the experimental result. The data required for the prediction model of the remaining molten steel volume were collected in the laboratory. Among them, 48 data with side wall rings, the input variables include nozzle diameter, nozzle height, step height difference, and suppression dam height. There are 69 data without side wall rings, and the input variables include nozzle diameter, nozzle height, step height difference, and the distance between the nozzle and the step. The statistical results and distribution of each input variable are shown in the figure. Figure 2 shown.

[0078] Since the data were collected in laboratory research experiments and there were only four input variables for both the case with and without sidewall rings, no further variable selection was performed in this paper. However, further analysis of the data showed that there were fewer samples collected when the residual water content was high, and this imbalance may cause the model to perform poorly in predicting this situation. Therefore, the SMOGN package was used to preprocess the data for oversampling. When the KNN distance of a given observation was close, SmoteR was used to generate new synthetic samples, and Gaussian noise was introduced when the KNN distance was far. Figure 3 The distribution of the output variables before and after oversampling is compared. After oversampling, the data with higher residual water content in the sample is increased, and the number of data points in each interval is more balanced.

[0079] Finally, the oversampled data is normalized to [0, 1], and the two sets of data are divided into a training set and a test set according to 0.7:0.3. In order to avoid the randomness of data division, the same training set and test set are used for different models in this embodiment.

[0080] The model selection module is used to build several machine learning models based on preprocessed data and evaluate several machine learning models based on decision coefficient, mean square error and mean absolute error.

[0081] Different machine learning algorithms were established for the cases with and without sluice gate cofferdams to predict the amount of remaining water when the water model was unloaded. In order to obtain a more accurate model, this paper established five machine learning models, including LASSO, SVR, ElasticNet, MLP, and XGBoost. The model was then further analyzed using the decision coefficient R 2 , mean square error MSE and mean absolute error MAE are used to measure the prediction accuracy of the model. The formulas of the three indicators are defined as follows:

[0082]

[0083] Among them, y i and The measured and predicted values ​​of the remaining water volume respectively; is the average value of the measured value of the remaining water volume; n is the total number of sample data; R 2 The value range is [0,1], based on the decision coefficient R 2 , MSE and MAE values ​​are used to evaluate several machine learning models. 2 The higher the , the lower the MSE and MAE values ​​are, and the better the model fit is.

[0084] The LASSO, SVR, ElasticNet, MLP, and XGBoost models used in the study were implemented based on Python's scikit-learn library and were run on a 64-bit, 12th Gen Intel(R) Core(TM) i9-12900H@2.50GHz Win11 system with 16Gb of RAM.

[0085] Compare with the actual results and select the optimal machine learning model. Figure 4 The comparison of prediction errors of the five models shows the R 2 , MSE and MAE. For R with and without sidewall rings 2 The prediction effects of the models are not significantly different. Except for LASSO and SVR with sidewall rings and LASSO without sidewall rings, the R 2 All reached above 0.95. For MSE and MAE, when predicting the two cases, the prediction errors of LASSO and SVR were large, especially the SVR model with sidewall rings, whose MSE and MAE values ​​were much larger than those of the other four models. MLP, ElasticNet and XGBoost all achieved good results, with MSE and MAE values ​​less than 1. Among them, the error value of the XGBoost model in predicting the two cases was closer to zero than that of other models, and R 2 The closer it is to 1, the better the prediction performance of the XGBoost model.

[0086] The molten steel volume prediction module is used to select a machine learning model based on the evaluation results to predict the remaining molten steel volume when the ladle is slag-free for ladles with or without wall side rings, and obtain the remaining molten steel volume prediction results. The machine learning model is used to predict and guide the remaining molten steel volume when the ladle is slag-free in the actual production process. The influence of each input variable on the remaining water volume in the ladle is studied based on the XGBoost model. The single variable principle is adopted to change the value range of the variable under study, and the other input variables take the mode. The values ​​of different variables are shown in Tables 7 and 8.

[0087] Table 7

[0088]

[0089] Table 8

[0090]

[0091] Figure 6The single variable influence of the experiment and prediction is compared. The red dotted line in the figure is the predicted value, and the black scattered points are the experimental values. It can be seen from the figure that for the case with side wall rings, except for the predicted value corresponding to the suppression dam height which has always been high, the predicted values ​​of the other three variables and the experimental values ​​are basically consistent in their influence on the remaining water volume; for the case without side wall rings, except for the predicted value of the distance between the step and the sprue which has always been higher than the experimental value, the two values ​​of the other three variables are also basically the same, and the change trend of the predicted value corresponding to the distance between the step and the sprue is roughly the same as the experimental value, both of which decrease first and then tend to be flat. In order to compare the contribution of each characteristic to the model, this paper adopts the Lundberg method.

[16] The SHAP (SHapley Additive explained) method introduced by et al. determines the impact of features on the ML model by changing the model input and calculates the SHAP average of each variable as its contribution value. The calculation results are shown in Figure 7 As shown in the figure, Figure 6 The two variables whose predicted values ​​differ greatly from the experimental values ​​contribute less to the model, that is, the model cannot thoroughly capture the substantial relationship between the variable and the predicted results, resulting in a large gap between the predicted values ​​and the experimental values.

[0092] Figure 6 (a) shows the effect of four input variables on the residual water volume in the ladle slag water simulation with sidewall rings. It can be seen that the lower the nozzle height and the higher the step height difference within a certain range, the less residual water volume, that is, the less molten steel is wasted in the ladle. The experimental results show that the influence of the suppression dam on the residual water volume is positively correlated, and the nozzle diameter is significantly reduced between 30 and 40 mm. Figure 6 (b) shows the influence of the four input variables when there is no sidewall ring. It can be seen from the figure that when there is no sidewall ring, the influence of the nozzle height, step height difference and nozzle diameter on the residual water volume is basically the same as that when there is a sidewall ring. As the distance between the step and the nozzle decreases, the residual water volume decreases first and then tends to be stable after 100mm.

[0093] for Figure 6 There are two main reasons for the formation of the ladder shape. The first is the structure of the model. As the most basic sub-model of XGBoost, the decision tree needs to divide the input data layer by layer according to the indicators to build the model. However, due to the limitations of the experimental data, the distribution of the training data is not dense and compact enough, that is, there are not enough data points. This is the second reason for the ladder shape. When predicting the data in the gaps, the model failed to divide the data during training, so when predicting the data, they were placed on the same "leaf". Each ladder represents a "leaf" formed by the integration of multiple decision trees.

[0094] Select Figure 7The two features with the highest contribution value were subjected to bivariate prediction analysis. The results are as follows: Figure 8 As shown. In the case of a side wall ring, when the step height difference is greater than 11mm and the sprue height is less than 12mm, the remaining water volume can be controlled below 7 liters. When the sprue height exceeds 24mm, the remaining water volume can reach more than 14 liters. For the case without a side wall ring, when the step height difference is greater than 11mm and the sprue height is less than 11mm, the remaining water volume is less than 10 liters. When the sprue height exceeds 26mm, the remaining water volume will reach more than 20 liters. Comparing the two situations, it can be seen that when the sprue height and the step height difference remain unchanged, the remaining water volume with the side wall ring is less than that without the side wall ring, that is, more molten steel can be saved during the production process.

[0095] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. It should be understood that the present invention is not limited to the forms disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the application concept described herein through the above teachings or the technology or knowledge in the relevant fields. The changes and modifications made by those skilled in the art shall not deviate from the spirit and scope of the present invention, and shall be within the scope of protection of the claims attached to the present invention.

Claims

1. A method for predicting the amount of molten steel remaining when the ladle is unloading slag using machine learning, characterized in that: Methods include: Step S1, using SMOGN to sample the water simulation environment data of the laboratory, and preprocessing the sampled water simulation data to obtain preprocessed data; Step S2: construct several machine learning models based on the preprocessed data, and evaluate the several machine learning models based on decision coefficient, mean square error and mean absolute error; Step S3: Based on the evaluation results, a machine learning model is selected to predict the amount of molten steel remaining when the ladle is unloading slag for ladles with or without wall side rings, and a prediction result of the amount of molten steel remaining is obtained.

2. The method for predicting the amount of molten steel remaining when the ladle is unloading slag by machine learning according to claim 1, characterized in that: In step S1, the construction of the water simulation environment specifically includes: A physical model of a ladle with a similarity ratio of 1:5 was established using acrylic glass, in which the height of step 1 was always twice that of step 2. Water was injected into the ladle to a height of 400 mm. After the water injection, a layer of oil with a thickness of 20 mm was covered and allowed to stand for 30 minutes to allow the flow field in the ladle to stabilize, thus completing the construction of the water simulation environment.

3. The method for predicting the amount of molten steel remaining when the ladle is unloading slag by machine learning according to claim 1, characterized in that: The sampled water simulation data include: ladle height, ladle upper diameter, ladle lower diameter, side wall step height, side wall step thickness, bottom step height, suppression dam length, suppression dam thickness and water inlet diameter.

4. The method for predicting the amount of molten steel remaining when the ladle is unloading slag by machine learning according to claim 1, characterized in that: In step S1, the sampled water simulation data is preprocessed to obtain the preprocessed data, which specifically includes: Two partitions are constructed according to the target variable value: BinsR partition and BinsN partition; the data points of the BinsN partition are sparsely distributed, and the data points of the BinsN partition are densely distributed; random undersampling is used to process the data points contained in the BinsN partition, and oversampling is used to process the partition in BinsR; Based on each data point in the BinsR partition, the Euclidean distance between the current data point and the remaining water simulation data points in the BinsR partition is calculated, and the safety distance is calculated based on the calculation result of the Euclidean distance; the water simulation data is classified by the K nearest neighbor algorithm, and the number of nearest neighbors is set to 2; If the distance between the current data point and the nearest neighbor point is within the "safe distance", SmoteR is used to generate a new synthetic sample. If the distance between the current data point and the nearest neighbor point exceeds the "safe distance", Gaussian noise is introduced to generate a new sample. The classified data is normalized to obtain preprocessed data.

5. The method for predicting the amount of molten steel remaining when the ladle is unloading slag by using machine learning according to claim 1, characterized in that: In step S2, the several machine learning models constructed based on the preprocessed data specifically include: LASSO model, SVR model, ElasticNet model, MLP model and XGBoost model.

6. The method for predicting the amount of molten steel remaining when the ladle is unloading slag by using machine learning according to claim 5, characterized in that: In step S2, based on the decision coefficient R 2 , mean square error MSE and mean absolute error MAE to evaluate several machine learning models, including: Among them, y i and The measured and predicted values ​​of the remaining water volume respectively; is the average value of the measured value of the remaining water volume; n is the total number of sample data; R 2 The value range is [0,1], based on the decision coefficient R 2 , MSE values ​​and MAE values ​​are used to evaluate several machine learning models.

7. A system for predicting the amount of molten steel remaining when the ladle is unloading slag using machine learning, the system being used to implement the method for predicting the amount of molten steel remaining according to any one of claims 1 to 6, characterized in that: The system includes: a pre-processing module, a model selection module and a molten steel quantity prediction module; The preprocessing module is used to sample the water simulation environment data of the laboratory using SMOGN, and preprocess the sampled water simulation data to obtain preprocessed data; The model selection module is used to construct several machine learning models based on the preprocessed data, and evaluate the several machine learning models based on the decision coefficient, mean square error and mean absolute error; The molten steel quantity prediction module is used to select a machine learning model based on the evaluation result to predict the remaining molten steel quantity when the ladle is unloading slag for the ladle with or without the wall side ring, and obtain the remaining molten steel quantity prediction result.

8. The system for predicting the amount of molten steel remaining when the ladle is unloading slag by using machine learning according to claim 7, characterized in that: In the pre-processing module, the construction of the water simulation environment specifically includes: A physical model of a ladle with a similarity ratio of 1:5 was established using acrylic glass, in which the height of step 1 was always twice that of step 2. Water was injected into the ladle to a height of 400 mm. After the water injection, a layer of oil with a thickness of 20 mm was covered and allowed to stand for 30 minutes to allow the flow field in the ladle to stabilize, thus completing the construction of the water simulation environment.

Images