Statistical analysis device, statistical analysis method, equipment control method, prediction model generation device, prediction model generation method, and numerical analysis result prediction method

Abstract

To provide a statistical analysis device and a statistical analysis method capable of improving the accuracy of statistical analysis processing.SOLUTION: A statistical analysis device 1 for refining includes a prediction section which uses a physical quantity distribution image in a refining facility 20 obtained by numerical analysis as teacher data and predicts the physical quantity distribution image in the refining facility 20 in operation according to operating conditions of the refining facility 20 by statistical analysis using a prediction model which predicts the physical quantity distribution image in the refining facility 20 in operation by statistical analysis, which is generated by sorting the teacher data into a plurality of condition ranges set to a condition range in which the operating conditions of the refining facility 20 can take place, among a plurality of teacher data sets which are generated by machine learning based on the teacher data set selected according to operating results of the refining facility 20.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a statistical analysis device, a statistical analysis method, a facility control method, a prediction model generation device, a prediction model generation method, and a numerical analysis result prediction method.

Background Art

[0002] In the steel refining process, the composition of steel is adjusted by blowing oxygen gas onto the high-temperature molten steel in a converter. Since the inside of the converter is a harsh environment with high temperature where molten steel scatters, it is difficult to install measuring instruments inside the converter to directly measure the data inside the converter. Therefore, the data inside the converter is indirectly measured using measuring instruments installed outside the converter. And when acquiring the physical quantities inside the converter, the physical quantities inside the converter are predicted by using the indirectly measured data in empirical formulas, theoretical formulas, etc.

[0003] In the refining process, there are many factors that cause variations in operating results, and it is difficult to grasp the trend of operation with only data of a small number of charges. Therefore, statistical analysis using data of a large number of charges has been performed to improve operating results. Specifically, with the improvement in the performance of information processing devices in recent years, data science (DS) technology using big data has been used to improve operating results. For example, as a DS technology for converters, Patent Document 1 describes a method of determining parameters in a model by learning past operating information of secondary refining facilities and estimating the molten steel temperature using the model. Further, Patent Document 2 describes a technique of determining parameters in an in-furnace estimation model formula by real-time statistical analysis using data collected outside the furnace and calculating the time variation of the molten steel composition using the in-furnace estimation model formula. Also, Patent Document 3 describes a technique of estimating the molten steel composition by a reaction formula using data of converter exhaust gas and intermediate substances.

[0004] On the other hand, as a tool different from statistical analysis, numerical analysis is also widely used for improving operation results. Numerical analysis performs calculations using physical models to reproduce physical phenomena or calculate physical quantities. According to numerical analysis, analysis results can be obtained relatively quickly at low cost without the need for experimental equipment. Also, in numerical analysis, since the operating conditions of the equipment to be numerically analyzed can be easily changed for analysis, the degree of freedom in analysis is high. Furthermore, in numerical analysis, it is possible to reproduce (speculate) the state of a place where it is difficult to measure or visualize with an actual machine, or to obtain physical quantities that are difficult to measure using measuring instruments such as sensors as data. As an example, it is known that numerical analysis is applicable in the iron and steel manufacturing process as well.

[0005] For example, Patent Document 4 describes a technique for calculating and visualizing the temperature distribution, stress, etc. of a steel material during rolling by performing numerical analysis using operation data in the rolling process of the steel material. Also, Patent Document 5 describes a method of constructing a database of numerical analysis results by performing numerical calculations in advance using numerical analysis methods such as the finite element method, and extracting appropriate numerical analysis results from the database for each operating condition to determine the optimal load. Also, Patent Document 6 describes a method of performing numerical analysis using operation results in advance and extracting from the database the numerical analysis results calculated under conditions closest to the operating conditions when operating a heating furnace. Also, Patent Document 7 describes a method of performing calculations in advance regarding the operation of blowing diluent gas fuel into a sintering machine and statistically processing the calculation results in advance.

Prior Art Documents

Patent Documents

[0006]

Patent Document 1

Patent Document 2

Patent Document 3

Patent Document 4

[0007] However, with the above-described statistical analysis techniques, only local values within the refining equipment can be predicted, and distribution information is unknown. Also, since the scale effects between the actual equipment and the laboratory cannot be reflected in the prediction formula, a large deviation between the predicted value and the measured value is expected. Furthermore, since the influence of the jet flow by the top lance is not considered, physical quantities that can only be obtained by directly measuring inside the furnace, such as dynamic pressure and ignition area, are not used. On the other hand, with the above-described numerical analysis techniques, since the calculation time is long, it is impossible to complete one calculation during one cycle of operation of the refining equipment, and numerical analysis results cannot be provided for each cycle. Also, although the numerical analysis results are rich in information because they can calculate the physical quantity distribution, due to the excessive amount of information, it becomes an operation such as using some data at a specific location, and the numerical analysis results cannot be effectively utilized.

[0008] The present invention has been made to solve the above problems, and an object thereof is to provide a statistical analysis device and a statistical analysis method capable of improving the accuracy of statistical analysis processing. Another object of the present invention is to provide a control method for equipment capable of improving the operation performance of the equipment. Another object of the present invention is to provide a prediction model generation device and a prediction model generation method capable of simultaneously realizing an improvement in prediction accuracy and a shortening of prediction time. Another object of the present invention is to provide a numerical analysis result prediction method capable of improving prediction accuracy. [Means for Solving the Problems]

[0009] The statistical analysis device according to the present invention uses, as teacher data, a physical quantity distribution image in a facility obtained by numerical analysis, and distributes the teacher data to a plurality of condition ranges set within a range of conditions that the operating conditions of the facility can take. Among the plurality of teacher data sets thus generated, a prediction model generated by machine learning based on the teacher data set selected according to the operating results of the facility, which predicts a physical quantity distribution image in the operating facility by statistical analysis, is used to predict, by statistical analysis, a physical quantity distribution image in the operating facility according to the operating conditions of the facility. A prediction unit, a physical quantity extraction unit that extracts values of physical quantities used when identifying the operating influence factors of the facility based on the physical quantity distribution image predicted by the prediction unit, and the values of the physical quantities extracted by the physical quantity extraction unit and the operating results of the facility corresponding to the operating conditions of the facility used by the prediction unit when predicting the physical quantity distribution image are used to obtain a correlation relationship, and a operating influence factor identification unit that identifies a physical quantity related to the operating results of the facility as an operating influence factor. It is characterized by comprising.

[0010] The statistical analysis method according to the present invention uses, as teacher data, a physical quantity distribution image in a facility obtained by numerical analysis, and distributes the teacher data to a plurality of condition ranges set within a range of conditions that the operating conditions of the facility can take. Among the plurality of teacher data sets thus generated, a prediction model generated by machine learning based on the teacher data set selected according to the operating results of the facility, which predicts a physical quantity distribution image in the operating facility by statistical analysis, is used to predict, by statistical analysis, a physical quantity distribution image in the operating facility according to the operating conditions of the facility. A prediction step, a physical quantity extraction step of extracting values of physical quantities used when identifying the operating influence factors of the facility based on the physical quantity distribution image predicted by the prediction step, and the values of the physical quantities extracted by the physical quantity extraction step and the operating results of the facility corresponding to the operating conditions of the facility used by the prediction step when predicting the physical quantity distribution image are used to obtain a correlation relationship, and an operating influence factor identification step of identifying a physical quantity related to the operating results of the facility as an operating influence factor. It is characterized by including.

[0011] The control method of the equipment according to the present invention is characterized by including a step of controlling the operation of the equipment by controlling the operation influencing factors specified by the above statistical analysis method.

[0012] The prediction model generation device according to the present invention includes a numerical analysis unit that generates a physical quantity distribution image in the equipment under a plurality of operating conditions under which the equipment can operate by numerical analysis, and a set of the operating conditions of the equipment and the physical quantity distribution image generated by the numerical analysis unit as teacher data, and sets a plurality of condition ranges according to the range of conditions that the operating conditions of the equipment can take, and divides the teacher data according to the plurality of condition ranges to generate a plurality of teacher data sets. A teacher data set creation unit, and a prediction model generation unit that selects the teacher data set according to the operation results of the equipment, performs machine learning, and generates a prediction model for predicting the physical quantity distribution image in the equipment during operation by statistical analysis.

[0013] The prediction model generation method according to the present invention includes a numerical analysis step of generating a physical quantity distribution image in the equipment under a plurality of operating conditions under which the equipment can operate by numerical analysis, and a set of the operating conditions of the equipment and the physical quantity distribution image generated by the numerical analysis step as teacher data, and sets a plurality of condition ranges according to the range of conditions that the operating conditions of the equipment can take, and divides the teacher data according to the plurality of condition ranges to generate a plurality of teacher data sets. A teacher data set creation step, and a prediction model generation step of selecting the teacher data set according to the operation results of the equipment, performing machine learning, and generating a prediction model for predicting the physical quantity distribution image in the equipment during operation by statistical analysis.

[0014] The numerical analysis result prediction method according to the present invention is characterized by including a step of predicting a physical quantity distribution image in the equipment during operation by statistical analysis using the prediction model generated by the above prediction model generation method.

Effects of the Invention

[0015] According to the statistical analysis apparatus and statistical analysis method according to the present invention, the accuracy of statistical analysis processing can be improved. Further, according to the facility control method according to the present invention, the operation performance of the facility can be improved. Further, according to the prediction model generation apparatus and prediction model generation method according to the present invention, it is possible to simultaneously improve the prediction accuracy and shorten the prediction time. Further, according to the numerical analysis result prediction method according to the present invention, the prediction accuracy can be improved.

Brief Description of the Drawings

[0016]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Embodiments for Carrying Out the Invention

[0017] Hereinafter, with reference to the drawings, a statistical analysis apparatus for refining, a statistical analysis method for refining, a control method for refining equipment, a prediction model generation apparatus for refining, a prediction model generation method for refining, and a numerical analysis result prediction method according to an embodiment of the present invention will be described.

[0018] 〔Configuration〕 First, with reference to FIG. 1, the configuration of a statistical analysis apparatus for refining according to an embodiment of the present invention will be described.

[0019] FIG. 1 is a block diagram showing the configuration of a statistical analysis device for refining according to an embodiment of the present invention. As shown in FIG. 1, a statistical analysis device 1 for refining according to an embodiment of the present invention is a device that improves the operation performance of a refining facility 20 by optimizing the operation conditions of the refining facility 20 through statistical analysis processing, and is configured by an information processing device capable of information communication between the refining facility 20 and its control device 21. Examples of the refining facility 20 to which the present invention can be applied include a converter, an RH (Ruhrstahl-Heraeus) furnace, an electric furnace, a VOD (Vacuum Oxygen Decarburization) furnace, and the like.

[0020] A statistical analysis device 1 for refining according to an embodiment of the present invention includes a numerical analysis unit 2, a teacher data set creation unit 3, a prediction model generation unit 4, an operation data acquisition unit 5, a prediction unit 6, a physical quantity extraction unit 7, an operation influence factor identification unit 8, a teacher data database (teacher data DB) 9, and an operation data database (operation data DB) 10. The numerical analysis unit 2, the teacher data set creation unit 3, the prediction model generation unit 4, the operation data acquisition unit 5, the prediction unit 6, the physical quantity extraction unit 7, and the operation influence factor identification unit 8 are functional blocks realized by an arithmetic processing device such as a CPU (Central Processing Unit) inside the information processing device executing a computer program.

[0021] The statistical analysis device 1 for refining having such a configuration improves the accuracy of the statistical analysis processing of the refining process and contributes to the improvement of the operation performance of the refining facility 20 by executing the statistical analysis processing for refining. The statistical analysis device 1 for refining is configured to be able to quickly predict and provide the numerical analysis result by using the past data of the numerical analysis as teacher data and predicting the specified numerical analysis result by a statistical analysis method.

[0022] 〔Statistical analysis processing for refining〕 FIG. 2 is a flowchart showing the flow of a statistical analysis process for refining according to an embodiment of the present invention. The process shown in FIG. 2 is executed by the statistical analysis device 1 for refining. In this description, the statistical analysis process for refining is simply described as the analysis process.

[0023] In the process of step S1, the numerical analysis unit 2 performs numerical analysis by changing the operating parameters (explanatory variables) of the refining facility 20 within a predetermined range. The numerical analysis unit 2 obtains, by numerical analysis, the physical quantity distribution image within the refining facility 20 under a plurality of operating conditions under which the refining facility 20 can operate. That is, the numerical analysis unit 2 generates, as a result of the numerical analysis, a physical quantity distribution image (an image including information on the target variable) within the refining facility 20. Then, the numerical analysis unit 2 stores the physical quantity distribution image of the numerical analysis result in the teacher data DB9 as teacher data.

[0024] For example, the numerical analysis unit 2 stores, in the teacher data DB9 as teacher data, a set of the operating parameters of the refining facility 20 and the physical quantity distribution image obtained by numerical analysis. This teacher data uses the operating parameters of the refining facility 20 as input data and the image of the numerical analysis result as output data.

[0025] The range for changing the value of the operating parameter in the numerical analysis process by the numerical analysis unit 2 is determined according to the range of the actual values of the operating conditions of the refining facility 20 stored in the operating data DB10. Examples of the operating parameters of the refining facility 20 include the molten steel charge amount, molten steel temperature, molten steel composition, blown oxygen flow rate, lance shape, lance height, bottom blowing flow rate, and auxiliary raw material input amount.

[0026] In addition, examples of the physical quantities obtained by numerical analysis include molten steel flow velocity, pressure, molten steel temperature, molten steel density, molten steel composition, and mass fraction of the molten steel phase - slag phase - gas phase (where the phase exists within the calculation mesh). Note that the higher the degree of the numerical analysis model formula, the more physical quantity distribution images can be obtained. On the other hand, the calculation cost becomes higher, so it takes a lot of time to accumulate a sufficient number of teacher data. The number of teacher data to be accumulated depends on how much the operating parameters of the refining facility 20 fluctuate, so it varies for each refining facility 20. Also, the number of teacher data to be stored in the teacher data DB9 also changes depending on the type of model used for the image analysis described later.

[0027] In this way, in the process of step S1, the numerical analysis unit 2 generates past data of numerical analysis as teacher data. Thereby, the process of step S1 is completed, and the analysis process proceeds to the process of step S2.

[0028] In the process of step S2, the teacher data set creation unit 3 generates a teacher data set in which the teacher data is distributed to predetermined groups. The teacher data set is data created based on the numerical analysis results (teacher data) obtained by the numerical analysis unit 2, and is data obtained by grouping the teacher data within a preset condition range.

[0029] For the operating conditions of the refining facility 20 that are input data among the teacher data, the teacher data set creation unit 3 sets a plurality of groups according to a predetermined range for specific operating parameters. In the data set creation process, not all of the operating parameters of the refining facility 20 are targeted, and grouping is performed for important variables that have a great influence on the operating results of the refining facility 20 among the operating parameters of the refining facility 20. The variables to be grouped are preset.

[0030] For example, in a converter, the oxygen flow rate has a great influence on the operating results, and the results of numerical analysis also vary greatly depending on the oxygen flow rate. Therefore, the teacher data set creation unit 3 targets the oxygen flow rate among the operating parameters, and divides the oxygen flow rate into groups in a predetermined range in advance. For example, the teacher data set creation unit 3 divides the oxygen flow rate into 6 groups at intervals of 10,000 Nm 3 / Hr in the range of 20,000 to 80,000 Nm 3 / Hr. Then, the teacher data set creation unit 3 distributes the teacher data generated by the numerical analysis unit 2 to any one of the 6 groups set for the oxygen flow rate. In this way, the teacher data set creation unit 3 executes a setting process of setting a plurality of condition ranges according to the condition ranges that the operating conditions of the refining facility 20 can take for the preset variables, and a classification process of sorting and storing the teacher data generated by the numerical analysis unit 2 according to the plurality of condition ranges (groups) set by the setting process.

[0031] In addition, the group can be set for a plurality of variables. The important variables to be grouped may be a plurality of variables. Specifically, as an important variable different from the oxygen flow rate, the lance height can be mentioned. For example, the lance height is divided into 5 groups at intervals of 0.5 m in the range of 1.5 to 4.0 m. In this case, a teacher data set of 6 groups of oxygen flow rate × 5 groups of lance height = 30 groups is generated. By applying such grouping, it is possible to create a data set that summarizes only the teacher data under conditions close to the conditions to be predicted in the prediction process described later. Therefore, it is possible to simultaneously improve the prediction accuracy and shorten the prediction time. Thereby, the process of step S2 is completed, and the analysis process proceeds to the process of step S3.

[0032] In the process of step S3, the prediction model generation unit 4 performs machine learning using the teacher data set to generate a prediction model. The prediction model generation unit 4 selects a teacher data set corresponding to the value of the operating condition according to the operating results of the refining facility 20, and performs machine learning based on the selected teacher data set.

[0033] When using a prediction model generated using teacher data far from the conditions to be predicted during image prediction described later, it will cause deterioration of the prediction accuracy. Therefore, when generating a prediction model, it is not preferable to use all the prepared teacher data together. In many prediction models, teacher data far from the conditions to be predicted is made to have a small weight and is less likely to be reflected in the prediction result, but the influence is not completely zero. When high prediction accuracy is required, data far from the specified conditions is harmful in prediction. Therefore, in the generation process of the prediction model executed by the prediction model generation unit 4, machine learning is performed using the teacher data set for each group divided into a predetermined range, and a prediction model corresponding to each group is generated.

[0034] The prediction model generation unit 4 performs machine learning using the teacher data set (a set of the operation parameters of the refining facility 20 corresponding to a predetermined condition range and the physical quantity distribution image) stored in the teacher data DB 9. That is, the prediction model generation unit 4 performs machine learning using the operation conditions of the refining facility 20 stored in the operation data DB 10. At this time, the prediction model generation unit 4 generates a prediction model for each group constituting the teacher data set. The generated prediction model is stored in the storage unit provided in the statistical analysis device 1 for refining. That is, the statistical analysis device 1 for refining includes the numerical analysis unit 2, the teacher data set creation unit 3, and the prediction model generation unit 4, and functions as a prediction model generation device that generates a prediction model. Thereby, the process of step S3 is completed, and the analysis process proceeds to the process of step S4.

[0035] In the process of step S4, the operation data acquisition unit 5 acquires, as operation data, information regarding the actual values of the operation conditions and operation results of the refining facility 20 from the refining facility 20 during the operation of the refining facility 20, and stores the acquired operation data in the operation data DB 10. The process of step S4 is executed when the refining facility 20 is in operation. That is, in step S4, the refining facility 20 performs an operation, and operation data such as the conditions and results during the operation are acquired by the operation data acquisition unit 5 and stored in the operation data DB 10.

[0036] Examples of the information regarding the operation conditions of the refining facility 20 include information necessary for performing numerical analysis in the process of step S1, such as the molten steel charging amount, molten steel temperature, molten steel composition, blown oxygen flow rate, lance shape, lance height, bottom blowing flow rate, and auxiliary raw material input amount. Examples of the information regarding the operation results of the refining facility 20 include the decarburization rate, which is the reduction rate of the molten steel composition, the dephosphorization rate, the usage amount of the auxiliary raw material, and the refining time. Thereby, the process of step S4 is completed, and the analysis process proceeds to the process of step S5.

[0037] In the process of step S5, the prediction unit 6 uses the operating conditions obtained in the process of step S4 as the required points (conditions for which a physical quantity distribution image is desired to be predicted), and predicts the physical quantity distribution image (numerical analysis image) for the required points by statistical analysis using the prediction model generated in the process of step S3. The prediction unit 6 predicts the numerical analysis image for the operating conditions of the refining facility 20 during operation by statistical analysis based on the prediction model.

[0038] In this case, the prediction unit 6 selects a prediction model generated using conditions close to the operating conditions based on the operating conditions of the refining facility 20 during operation. In the prediction process executed by the prediction unit 6, a prediction model generated using only teacher data (image of numerical analysis results) calculated under conditions close to the conditions to be predicted is selected.

[0039] For example, the prediction unit 6 identifies the group that matches the conditions to be predicted (operating conditions during operation) among the grouped condition ranges in the teacher data set, and selects a prediction model generated by machine learning using only the teacher data of the identified group. In this way, the prediction unit 6 identifies the group corresponding to the operating conditions during operation, and selects a prediction model generated using the teacher data set belonging to that group. When the actual operation of the converter is performed, a prediction model generated by the teacher data set that matches the value of the operating conditions is selected and used for prediction. The prediction unit 6 configured in this way uses a prediction model that predicts the physical quantity distribution image in the refining facility 20 during operation by statistical analysis, which is generated by machine learning based on the teacher data set selected according to the operation results of the refining facility 20, to predict the physical quantity distribution image (numerical analysis result) in the refining facility 20 during operation by statistical analysis according to the operating conditions of the refining facility 20.

[0040] Note that at this time, since the image file only has color information of the image, it is necessary to convert what physical quantity value the color of the image represents. Usually, the conversion is performed using the relational expression between the color and the physical quantity value required when creating the contour diagram of the numerical analysis image. Since an image is a collection of a large number of dots, it is necessary to predict the values of all dots according to the number of pixels. Also, naturally, the analytical load increases as the number of pixels increases. Examples of statistical analysis methods include regression models (linear regression, multiple regression, local regression, etc.), principal component analysis, neural networks, and the like. Thereby, the process of step S5 is completed, and the analysis process proceeds to the process of step S6.

[0041] In the process of step S6, the physical quantity extraction unit 7 extracts, by image analysis technology, the value of the physical quantity used when identifying the operation influence factor described later from the physical quantity distribution image predicted in the process of step S5.

[0042] For example, when the physical quantity distribution image is a distribution image of the jet flow velocity, the physical quantity extraction unit 7 extracts the maximum value of the molten steel flow velocity, the average value of the molten steel flow velocity, the molten steel flow velocity immediately after the oxygen discharge from the lance, the angle at which the jet deviates from the straight-ahead direction, and the jet shape (width, spread), etc. at the molten steel surface height coordinate. Note that image analysis technology has been established in various technical fields, and by using these technologies, image analysis becomes easier. For example, in the field of steel materials, specific crystal grains are recognized from a microscopic crystal structure image to extract the grain size and number of crystal grains, etc. However, since it takes time to do this manually, it is being done automatically by a computer program. Therefore, by applying this technology, processes such as extracting the ignition area of the region above a predetermined pressure from the molten steel surface position pressure distribution can be performed.

[0043] In addition, when performing numerical analysis of multiphase flow, information such as the shape and depth of the cavity (the depression of the molten steel generated by top-blowing oxygen) can be extracted from the molten steel mass distribution in the longitudinal section of the converter through image analysis. Also, it is possible to recognize and count the splashes in the image and take statistics on the sizes of the splashes. The advantages of performing image analysis include the ability to extract necessary physical quantities in various forms even when the information volume of the physical quantity distribution image is large. One of the main purposes of statistical analysis is to identify operating influence factors (physical quantities that greatly affect the operating results). However, since it is not known before performing statistical analysis what physical quantities should be extracted as candidates for operating influence factors and how to extract them, it is desirable to increase the candidates for physical quantities as much as possible. By programming the candidate physical quantities and their extraction methods in image analysis, the number of extracted physical quantities can be increased. As a result, the process of step S6 is completed, and the analysis process proceeds to the process of step S7.

[0044] In the process of step S7, the operating influence factor identification unit 8 analyzes the correlation between the physical quantities extracted from the physical distribution image in the process of step S6 and the operating results corresponding to the operating conditions of the refining facility 20 used when predicting the physical quantity distribution image, thereby identifying the physical quantities related to the operating results of the refining facility 20 as operating influence factors. Note that the correlation between the physical quantity and the operating results may be evaluated by simple correlation analysis, but if a method such as a neural network is applied, the influence of the correlation between multiple physical quantities on the operating results can also be analyzed. Also, a regression equation may be created using the physical quantities correlated with the operating results. For example, if it is identified by correlation analysis that the operating influence factors on the decarburization rate are the molten steel surface flow rate, oxygen flow rate, and lance height, a regression equation such as decarburization rate = a × molten steel surface flow rate + b × oxygen flow rate + c × lance height can be created to estimate the improvement cost of the decarburization rate when the operating conditions are changed. Also, the optimal operating conditions can be identified from the regression equation and used for controlling the operating conditions of the subsequent refining facility 20. As a result, the process of step S7 is completed, and the series of analysis processes ends.

[0045] Thereafter, the operator of the refining facility 20 changes the operating conditions based on the operation impact factors identified in the process of step S7 to improve the operation performance of the refining facility 20. For example, if it is found that the operation performance of the refining facility 20 is improved by increasing the top blowing oxygen flow rate from the current amount in the process of step S7, the operator of the refining facility 20 manually changes the operating conditions of the refining facility 20 to increase the top blowing oxygen flow rate. Note that the control device 21 may automatically change the operating conditions of the refining facility 20. For example, if it is found that the lance height should be set to 3 m when the molten steel amount is 300 tons in the process of step S7, the control device 21 may automatically change the lance height to 3 m during the operation when the molten steel amount is 300 tons. Further, the processes of steps S4 to S7 may be performed online every time after each charge, and the operation may be performed for each charge under the optimized operating conditions. For this analysis process, the processes of steps S1 to S3 can be used as a prediction model generation method, and the processes of steps S4 to S7 can be used as a statistical analysis method. Furthermore, the refining statistical analysis device 1 can perform a control method for the refining facility 20 including a step of controlling the operation of the refining facility 20 by controlling the operation impact factors identified by this statistical analysis method.

Example

[0046] In the example, the target refining facility 20 is the converter 30 shown in FIG. 3. As the numerical analysis model, the top blowing jet flow analysis model of the gas phase shown in FIG. 4 was used. Note that FIG. 3 shows the molten steel 31, the slag 32, the lance 33, and the jet flow 34. Further, FIG. 4 shows the lance 33, the furnace mouth 36, and the molten steel interface 37.

[0047] The parameters of the up-blow jet analysis model are the oxygen flow rate and lance height, each with 20 levels, and a total of 400 levels of analysis were performed. The analysis result image shown in Fig. 5 (the example shown in Fig. 5 is the flow velocity distribution image), which serves as the training data, was prepared. Next, the operation data for 1000 charges of the converter 30 were obtained, and the numerical analysis result images for each charge were predicted by image analysis using a regression model. At that time, in order to improve the prediction accuracy and shorten the prediction time, the oxygen flow rate and lance height, which are important variables with a large impact on the operation, were each divided into groups, and all the training data were distributed among the groups.

[0048] Specifically, the oxygen flow rate was set in six range groups such as 20000 - 30000 Nm 3 / Hr, 30000 - 40000 Nm 3 / Hr, 40000 - 50000 Nm 3 / Hr, 50000 - 60000 Nm 3 / Hr, 60000 - 70000 Nm 3 / Hr, 70000 - 80000 Nm 3 / Hr. For example, if the training data is calculated at an oxygen flow rate of 22000 Nm 3 / Hr, it will be assigned to the group of 20000 - 30000 Nm 3 / Hr. When the actual converter 30 is operated under the condition of 25000 Nm 3 / Hr, the prediction is made using the prediction model generated using all the training data belonging to 20000 - 30000 Nm 3 / Hr.

[0049] Next, for each of the six groups divided by the oxygen flow rate, the lance height was divided into groups of 1.5 - 2.0 m, 2.0 - 2.5 m, 2.5 - 3.0 m, 3.0 - 3.5 m, and 3.5 - 4.0 m. Since there are six groups for the oxygen flow rate and five groups for the lance height, a total of 6×5 = 30 groups are divided. When the converter 30 is operated with an oxygen flow rate of 35000 Nm 3 / Hr and a lance height of 2.7 m, 40000 - 50000 Nm 3Predict using the prediction model generated using only the teacher data belonging to the group with a lance height of 2.5 to 3.0 m and / Hr.

[0050] At the same time, the prediction accuracy was checked. As a method, out of a large number of data prepared as teacher data, only one was excluded and not used for prediction. Furthermore, as an input condition when performing prediction, the analysis conditions (explanatory variables) of the excluded teacher data were used. Since the result predicted by the excluded teacher data becomes the correct answer, the accuracy check can be performed from the degree of deviation between the prediction result and the correct answer. As a result, in the prediction when all teacher data were used, a deviation of ± 5 m / sec occurred in the predicted value of the maximum flow velocity of the molten metal surface. However, when using only the group of teacher data under conditions close to the operating conditions, the predicted value of the maximum flow velocity of the molten metal surface was reduced to ± 1 m / sec, and the prediction accuracy could be improved. In addition, the numerical analysis image prediction time per charge could be shortened to 1 / 30. This showed that the classification of teacher data is effective for improving the prediction accuracy and shortening the time.

[0051] The distribution images of physical quantities use the flow velocity distribution, pressure distribution, temperature distribution, density distribution, and composition distribution. From these images, values of a number of physical quantities such as the maximum flow velocity, discharge flow velocity, straightness (merging) of the jet, maximum dynamic pressure of the molten metal surface, ignition point area, average dynamic pressure of the ignition point, maximum temperature of the ignition point, furnace atmosphere temperature, exhaust gas concentration, exhaust gas flow rate, exhaust gas temperature, etc. were extracted.

[0052] When calculating the importance using the regression tree for 1000 charge prediction data, it was found that the importance of the average dynamic pressure of the molten metal surface and the decarburization rate was high. Therefore, a large number of numerical calculations with the lance shape changed were further carried out to search for and identify a lance that improves the average dynamic pressure of the molten metal surface. By using the identified lance shape, the decarburization rate was improved, which could lead to an improvement in operation.

[0053] As is clear from the above description, according to the refining statistical analysis device 1, since prediction by statistical analysis is performed using a prediction model generated with numerical analysis results calculated under conditions close to the conditions to be predicted as teacher data, the accuracy of the statistical analysis process of the refining process can be improved. Further, by controlling the operation of the refining facility 20 by controlling the identified operation influencing factors, the operation performance of the refining facility 20 can be improved.

[0054] Specifically, regarding data acquisition in the refining facility 20, which is an issue in statistical analysis in the refining facility 20, it can be supplemented by using the physical quantity distribution data in the refining facility 20 calculated by numerical analysis. However, since numerical analysis takes a long time to calculate, it is not possible to obtain a sufficient number of data without shortening the calculation time. Therefore, in this embodiment, by using past data of numerical analysis as teacher data and predicting specified numerical analysis results by a statistical analysis method, the numerical analysis results can be provided quickly. Also, at that time, the input and output of numerical analysis are made into images with a large amount of information, and by combining image analysis, a large amount of numerical analysis data can be effectively utilized. Thereby, the accuracy of the statistical analysis process of the refining process can be improved.

[0055] Note that the variable to be grouped only needs to include at least one of the oxygen flow rate and the lance height. That is, a configuration may be adopted in which a teacher data set grouped for both the oxygen flow rate and the lance height is generated, a configuration may be adopted in which a teacher data set grouped only for the oxygen flow rate is generated, or a configuration may be adopted in which a teacher data set grouped only for the lance height is generated.

[0056] In addition, the number of groups in the teacher dataset can be set to any number. That is, the number of groups for the oxygen flow rate is not limited to six. The number of groups for the lance height is not limited to five. When grouping the oxygen flow rate and the lance height, the number of groups is not limited to thirty. That is, the range of conditions when creating the teacher dataset is not limited to the above-described range of conditions. It can be set within the range of operable conditions to a range of conditions that contribute to improving the prediction accuracy.

[0057] As described above, embodiments to which the invention made by the present inventors is applied have been described. However, the present invention is not limited by the description and drawings that form part of the disclosure of the present invention according to this embodiment. For example, although this embodiment applies the present invention to a refining facility, the application range of the present invention is not limited to a refining facility, and it can be applied to all facilities capable of performing numerical analysis. Thus, all other embodiments, examples, and operation techniques made by those skilled in the art based on this embodiment are included in the scope of the present invention.

Explanation of Reference Numerals

[0058] 1 Statistical analysis device for refining 2 Numerical analysis unit 3 Teacher dataset creation unit 4 Prediction model generation unit 5 Operation data acquisition unit 6 Prediction unit 7 Physical quantity extraction unit 8 Operation influence factor identification unit 9 Teacher data database (teacher data DB) 10 Operation data database (operation data DB) 20 Refining facility 21 Control device 30 Converter 31 Molten steel 32 Slag 33 Lance 34 Jet 36 Furnace mouth 37 Molten steel interface

Claims

1. Using a prediction model that predicts the physical quantity distribution image inside the facility during operation by statistical analysis, which is generated by machine learning based on the teacher data set selected according to the operation performance of the facility among a plurality of teacher data sets generated by distributing the physical quantity distribution image inside the facility obtained by numerical analysis as teacher data to a plurality of condition ranges set within the range of conditions that the operation conditions of the facility can take, a prediction unit that predicts the physical quantity distribution image inside the facility during operation by statistical analysis according to the operation conditions of the facility; A physical quantity extraction unit that extracts the value of the physical quantity used when identifying the operation influencing factors of the facility based on the physical quantity distribution image predicted by the prediction unit; An operation influencing factor identification unit that identifies the physical quantity related to the operation performance of the facility as an operation influencing factor by obtaining the correlation between the value of the physical quantity extracted by the physical quantity extraction unit and the operation performance of the facility corresponding to the operation conditions of the facility used by the prediction unit when predicting the physical quantity distribution image; Comprising: The facility is a refining facility; The physical quantities obtained by the numerical analysis are molten steel flow rate, pressure, molten steel temperature, molten steel density, molten steel composition, and mass fraction of molten steel phase - slag phase - gas phase; The physical quantity distribution image is an image using flow velocity distribution, pressure distribution, temperature distribution, density distribution, and composition distribution; The information regarding the operation conditions is molten steel charge amount, molten steel temperature, molten steel composition, blown oxygen flow rate, lance shape, lance height, bottom blowing flow rate, and auxiliary raw material input amount; The information regarding the operation performance is decarburization rate, dephosphorization rate, which are the reduction rates of the composition of molten steel, the usage amount of auxiliary raw materials, and refining time; The values of the physical quantities extracted by the physical quantity extraction unit are the maximum flow velocity per height, discharge flow velocity, straight - advance degree of the jet, maximum dynamic pressure on the molten steel surface, ignition point area, average dynamic pressure of the ignition point, maximum temperature of the ignition point, furnace atmosphere temperature, exhaust gas concentration, exhaust gas flow rate, and exhaust gas temperature A statistical analysis device characterized by the above.

2. Using a prediction model that predicts the physical quantity distribution image inside the facility during operation by statistical analysis, which is generated by machine learning based on the teacher data set selected according to the operation results of the facility among a plurality of teacher data sets generated by distributing the physical quantity distribution image inside the facility obtained by numerical analysis as teacher data to a plurality of condition ranges set within the range of conditions that the operation conditions of the facility can take, a prediction step of predicting the physical quantity distribution image inside the facility during operation by statistical analysis according to the operation conditions of the facility; A physical quantity extraction step of extracting the value of the physical quantity used when identifying the operation influence factor of the facility based on the physical quantity distribution image predicted in the prediction step; An operation influence factor identification step of identifying the physical quantity related to the operation result of the facility as an operation influence factor by obtaining the correlation between the value of the physical quantity extracted in the physical quantity extraction step and the operation result of the facility corresponding to the operation conditions of the facility used when the prediction step predicts the physical quantity distribution image; including; The facility is a refining facility; The physical quantities obtained by the numerical analysis are molten steel flow rate, pressure, molten steel temperature, molten steel density, molten steel composition, and mass fraction of molten steel phase - slag phase - gas phase; The physical quantity distribution image is an image using flow rate distribution, pressure distribution, temperature distribution, density distribution, and composition distribution; The information regarding the operation conditions is molten steel charge amount, molten steel temperature, molten steel composition, blown oxygen flow rate, lance shape, lance height, bottom blowing flow rate, and auxiliary raw material input amount; The information regarding the operation result is decarburization rate, dephosphorization rate, which are the reduction rates of the composition of molten steel, the usage amount of auxiliary raw materials, and refining time; The values of the physical quantities extracted in the physical quantity extraction step are the maximum flow rate per height, discharge flow rate, straight - advancing degree of the jet, maximum dynamic pressure on the molten steel surface, ignition point area, average dynamic pressure of the ignition point, maximum temperature of the ignition point, furnace atmosphere temperature, exhaust gas concentration, exhaust gas flow rate, and exhaust gas temperature A statistical analysis method characterized by this.

3. A facility control method characterized by including a step of controlling the operation of the facility by controlling the operation influence factor identified by the statistical analysis method according to Claim 2.

4. A numerical analysis unit that generates a physical quantity distribution image inside the facility under a plurality of operation conditions under which the facility can operate by numerical analysis; Using the pair of the operating conditions of the equipment and the physical quantity distribution image generated by the numerical analysis unit as teacher data, and setting a plurality of condition ranges according to the range of conditions that the operating conditions of the equipment can take, and classifying the teacher data according to the plurality of condition ranges to generate a plurality of teacher data sets, a teacher data set creation unit; Selecting the teacher data set according to the operating results of the equipment and performing machine learning to generate a prediction model for predicting the physical quantity distribution image in the operating equipment by statistical analysis, a prediction model generation unit; comprising; the equipment is a refining equipment; the physical quantities obtained by the numerical analysis are molten steel flow rate, pressure, molten steel temperature, molten steel density, molten steel composition, and mass fractions of the molten steel phase - slag phase - gas phase; the physical quantity distribution image is an image using flow rate distribution, pressure distribution, temperature distribution, density distribution, and composition distribution; the information on the operating conditions is molten steel charge amount, molten steel temperature, molten steel composition, blown oxygen flow rate, lance shape, lance height, bottom blowing flow rate, and auxiliary raw material input amount; A prediction model generation device characterized by the above.

5. A numerical analysis step of generating a physical quantity distribution image in the equipment under a plurality of operating conditions that the equipment can operate by numerical analysis; Using the pair of the operating conditions of the equipment and the physical quantity distribution image generated by the numerical analysis step as teacher data, and setting a plurality of condition ranges according to the range of conditions that the operating conditions of the equipment can take, and classifying the teacher data according to the plurality of condition ranges to generate a plurality of teacher data sets, a teacher data set creation step; Selecting the teacher data set according to the operating results of the equipment and performing machine learning to generate a prediction model for predicting the physical quantity distribution image in the operating equipment by statistical analysis, a prediction model generation step; including; the equipment is a refining equipment; the physical quantities obtained by the numerical analysis are molten steel flow rate, pressure, molten steel temperature, molten steel density, molten steel composition, and mass fractions of the molten steel phase - slag phase - gas phase; the physical quantity distribution image is an image using flow rate distribution, pressure distribution, temperature distribution, density distribution, and composition distribution; the information on the operating conditions is molten steel charge amount, molten steel temperature, molten steel composition, blown oxygen flow rate, lance shape, lance height, bottom blowing flow rate, and auxiliary raw material input amount; A prediction model generation method characterized by the above.

6. A numerical analysis result prediction method characterized by including a step of predicting a physical quantity distribution image in an operating facility by statistical analysis using a prediction model generated by the prediction model generation method according to claim 5.

Images