JP7456592B2 - Graphite nodularization treatment device, learning device, and method for determining magnesium input amount - Google Patents

Description

The present invention relates to a graphite nodularization treatment device, a learning device, and a method for determining the amount of magnesium input.

Spheroidal graphite cast iron, or ductile cast iron, has excellent strength and ductility, so it is used in products in various fields such as iron pipes for water and sewage systems. Ductile cast iron is manufactured through a graphite spheroidization process in which magnesium is added to molten metal to spheroidize crystallized graphite.

In JP-A No. 2005-281804 (Patent Document 1), in order to deal with the fading of magnesium that occurs during graphite nodularization, the magnesium content in the molten metal is calculated in real time, and based on the results, Techniques have been proposed to control the timing and amount of supplemental addition of magnesium. Specifically, the magnesium content in the molten metal is predicted by substituting the measured values obtained by analyzing the components in the molten metal in the ladle, the molten metal temperature, the amount of molten metal, and the elapsed time into a predetermined integral formula. is calculating.

In Japanese Unexamined Patent Application Publication No. 2005-298843 (Patent Document 2), a calculation unit of a control device calculates the mass of molten iron, the S (sulfur) concentration in molten iron before treatment, and the molten iron after treatment input into an operation panel in advance. A technique has been proposed in which the Mg input amount is calculated as the length of an iron-coated Mg wire based on a target value of Mg (magnesium) concentration and a set value of Mg addition yield.

In addition, Toshitake Kanno, “Challenge to IoT in the casting field”, Sokeizai Vol. 59 (Non-Patent Document 1) describes an example in which a neural network is used to predict the Mg yield of a ladle in a graphite nodularization process. It is stated that if the data input to AI is only simple raw data, the correlation between the measured Mg yield (teacher data) and predicted Mg yield will be low, but if theoretical data is added to the raw data, the correlation will be high. There is. The raw data includes the specific surface area of the ladle, the amount of Mg added, the spheroidization temperature, the wire feeding speed, and the component values of the source water. The theoretical data is the amount of Mg that can be dissolved, the molecular weight ratio and reaction amount of various elements (O, S, N) that combine with Mg.

Japanese Patent Application Publication No. 2005-281804 JP2005-298843A (Patent No. 3939307)

Toshitake Kanno, “Challenge to IoT in the casting field”, Materials and Materials Vol. 59, No. 6, pp. 13-19, published June 2018

In the graphite spheroidization process, some of the magnesium added vaporizes or combines with other elements and cannot remain in the molten metal, so the magnesium content after treatment is determined by the amount of magnesium input, the main melting raw material, It varies in a complicated manner depending on various detailed conditions such as auxiliary raw materials, melting method, spheroidization method, etc. Therefore, when calculating the amount of magnesium to be added using techniques such as those disclosed in Patent Documents 1 and 2, it is necessary to set the yield in advance by a method such as human prediction.

In addition, in Non-Patent Document 1, the magnesium yield after graphite spheroidization is predicted by AI, but even if many condition data included in raw data and theoretical data are input as explanatory variables, the yield will be affected. Since it is difficult to know all the conditions in advance, it is not always possible to predict the magnesium yield with high accuracy. Therefore, even if the amount of magnesium to be added to the ladle is determined based on a large amount of condition data, there is a possibility that the target magnesium content cannot be achieved.

The present invention has been made to solve the above-mentioned problems, and its purpose is to provide a graphite nodularization treatment device that can efficiently achieve a target magnesium content, and to reduce the amount of magnesium input. The purpose is to provide a method for making decisions.

Another object of the present invention is to provide a learning device suitable for efficiently achieving a target magnesium content.

A graphite spheroidizing treatment device according to one aspect of the present invention is a graphite spheroidizing treatment device equipped with a magnesium supply means for supplying magnesium to the molten metal in a ladle, and is equipped with a data storage means and a determination means. The data storage means stores performance data including a pair of the amount of magnesium input in a past cup and the magnesium content of the molten metal after the graphite spheroidizing treatment, or result data indicating the yield. The determination means predicts the magnesium content or yield of the molten metal after treatment for each treatment cup based on the past performance data stored in the data storage means, and determines the amount of magnesium input from the magnesium supply means.

Preferably, the determining means determines the magnesium content of the molten metal after treatment or It includes a prediction means for predicting yield.

It is desirable that the performance data stored in the data storage means further include molten metal condition data in addition to result data.

More preferably, the graphite nodularization processing apparatus further includes model storage means for storing a learning model generated by machine learning. In this case, the determining means (predicting means) inputs at least past performance data into the learning model to predict the magnesium content or yield of the molten metal after treatment.

A learning device according to another aspect of the invention includes a generation means and a model storage means. The generation means is a combination of the amount of magnesium input in the past cup and the magnesium content of the molten metal after graphite spheroidization treatment, or past performance data including result data indicating yield, and the molten metal after processing in the treated cup. Generate a learning model by machine learning the correlation with magnesium content or yield. The model storage means stores the learning model generated by the generation means.

Preferably, the learning device further includes evaluation means for evaluating the learning model. In this case, the evaluation means determines the error ( It is desirable to greatly evaluate the loss).

A method for determining the amount of magnesium input according to still another aspect of the present invention is a method for determining the amount of magnesium input in a graphite spheroidization treatment apparatus, which comprises comparing the amount of magnesium input in past cups and the magnesium content of the molten metal after the graphite spheroidization treatment. a step of storing performance data in a storage unit, including result data indicating a combination with a quantity or a yield; and a step of storing magnesium in the molten metal after treatment for each processing cup based on past performance data stored in the storage unit. The method includes a step of predicting the content or yield, and a step of determining the amount of magnesium to be added to the processing cup depending on the prediction result.

According to the present invention, the target magnesium content can be efficiently achieved.

1 is a diagram schematically showing a manufacturing process of ductile cast iron in an embodiment of the present invention. (A) is a block diagram showing a functional configuration of a graphite nodularization processing apparatus according to an embodiment of the present invention, and (B) is a block diagram showing a functional configuration of a learning device according to an embodiment of the present invention. It is. FIG. 3 is a diagram conceptually illustrating a method for predicting Mg yield using a prediction model in an embodiment of the present invention in comparison with a comparative example. It is a figure which shows the simulation result when Mg yield was predicted by the prediction model in the embodiment of this invention, compared with a comparative example. FIG. 3 is a diagram showing types of error functions (loss functions) used for evaluating a prediction model. It is a flowchart which shows graphite nodularization processing in an embodiment of the present invention.

The embodiment of the present invention will be described in detail with reference to the drawings. Note that the same or corresponding parts in the drawings are given the same reference numerals and their description will not be repeated.

<Overview of the manufacturing process of ductile cast iron>
First, with reference to FIG. 1, an overview of the manufacturing process of ductile cast iron will be explained. FIG. 1 is a diagram schematically showing the manufacturing process of ductile cast iron in this embodiment.

In this embodiment, the ductile cast iron is processed in the order of melting process P1, desulfurization process P2, molten metal holding process P3, molten metal component analysis process P4, component adjustment process P5, spheroidization process P6, slag removal process P7, and casting process P8. After that, it is manufactured.

In the melting process P1, molten metal such as pig iron, scrap iron, steel, etc. is melted in a melting furnace 11 such as a cupola. As the melting furnace 11, for example, an electric furnace may be employed.

In the desulfurization step P2, the molten metal produced in the melting furnace 11 is poured into the desulfurization ladle 12, and the sulfur content in the molten metal in the ladle 12 is removed. Note that the desulfurization step P2 may be omitted.

In the molten metal holding process P3, the desulfurized molten metal is poured into a low-frequency furnace 13 for holding the molten metal and is held there.

In the molten metal component analysis step P4, the components of the molten metal in the low frequency furnace 13 are analyzed by a QV (cantback) analyzer 14. After component analysis, the molten metal in the low frequency furnace 13 is poured into a ladle 15 having a predetermined capacity (for example, 6 tons).

In the component adjustment step P5, the components in the molten metal are adjusted based on the analysis results by the QV analyzer 14. Specifically, the analysis results by the QV analyzer 14 are input to the component adjustment device 16, and the component adjustment device 16 adjusts elemental components such as carbon (C) and silicon (Si) to be added to the molten metal in the ladle 15. The amount is adjusted.

In the spheroidizing process P6, a graphite spheroidizing apparatus (hereinafter abbreviated as "spheroidizing apparatus") 20 performs a process of spheroidizing graphite in the molten metal. In this embodiment, magnesium is introduced into the ladle 15 by the cored wire device 21.

The cored wire device 21 includes a supply device 23 that draws out the wire 22 from the wire coil 26 and supplies it to the ladle 15, and a guide member 24 that guides the movement of the wire 22 from the wire coil 26 to the supply device 23. The wire 22 is a cored wire in which granular magnesium alloy is enclosed within an iron shell. By supplying the wire 22 to the ladle 15, magnesium is added (injected) to the molten metal in the ladle 15. The supply device 23 functions as a magnesium supply means and includes a drive unit for drawing out the wire 22 (not shown).

The supply device 23 is controlled by a control device 25. The control device 25 determines the amount of magnesium to be added to the ladle 15 each time the spheroidization process is performed (that is, for each processing cup), and instructs the supply device 23 to do so. The supply device 23 supplies the wire 22 to the ladle 15 in a length corresponding to the specified amount of magnesium to be fed. The control device 25 is connected to the above-mentioned QV analyzer 14 by wire or wirelessly, and can use the analysis results by the QV analyzer 14 to determine the amount of magnesium to be fed. The spheroidization processing apparatus 20 according to the present embodiment includes a cored wire device 21, a control device 25, and a QV analysis device 14. Note that the control device 25 may be incorporated into the cored wire device 21.

In the slag removal step P7, the ladle 15A after the spheroidization process is taken out from the cored wire device 21, and slag is removed from the molten metal in the ladle 15A. The ladle 15B after the sludge removal process is transported to the casting process P8 by a transport means such as a crane. The above-mentioned QV analyzer 14 analyzes the components of the molten metal in the ladle 15A after the spheroidization process, and outputs the analysis results to the control device 25 of the spheroidization process apparatus 20. Note that a QV analyzer that analyzes the components of the molten metal before the spheroidization process and a QV analyzer that analyzes the components of the molten metal after the spheroidization process may be provided separately.

In the casting process P8, the molten metal in the ladle 15B is poured into a mold, and ductile cast iron is cast.

If the magnesium content of the molten metal in the ladle 15A after the spheroidization treatment (hereinafter also referred to as "residual Mg amount") is less than the specified amount, the quality of the ductile cast iron will deteriorate. Therefore, in the spheroidization process P6, it is desirable to charge magnesium into the ladle 15 as much as possible, so that the amount of residual Mg becomes the target residual Mg of the target value (a specified amount or more).

The spheroidization processing apparatus 20 according to the present embodiment mainly determines the amount of magnesium input (hereinafter also referred to as "the amount of Mg input") based on past performance data including result data indicating Mg yield. It has the following characteristics. The Mg yield is expressed by the following formula.
Mg yield = {processed hot water amount (output hot water amount + remaining hot water amount) x residual Mg amount}/input Mg amount

Note that, as expressed in the above formula, the Mg yield is determined by the input Mg amount and the residual Mg amount, so instead of the Mg yield, a set of the input Mg amount and the residual Mg amount may be used as the result data. .

<Functional configuration of spheroidization processing device>
The functional configuration of the spheroidization processing device 20 will be described with reference to FIG. 2(A). FIG. 2(A) is a block diagram showing the functional configuration of the spheroidization processing apparatus 20 according to this embodiment.

The spheroidization processing device 20 includes a data storage section 31, a model storage section 32, a determination section 33, an instruction output section 34, and a result acquisition section 35. These units 31 to 35 are included in the control device 25 shown in FIG.

The data storage unit 31 stores past performance data for each spheroidization process (for each cup). The data storage unit 31 may store only the most recent predetermined number (for example, two) of performance data (past cup information). The performance data includes molten metal condition data indicating various conditions of the spheroidization process and result data of the spheroidization process.

The molten metal condition data includes all or at least one of the amount of treated molten metal, the temperature before treatment, and the source molten metal components. The source hot water component contains both or at least one of C content and S content.

In this embodiment, the result data represents the Mg yield. As described above, the result data may be the amount of Mg added and the amount of residual Mg. In other words, the data storage unit 31 only needs to store performance data including a pair of the amount of magnesium added in a past cup and the magnesium content of the molten metal after the graphite spheroidizing process, or result data indicating the yield.

The model storage unit 32 stores a prediction model 32A. Prediction model 32A is a learning model generated by learning device 100, which will be described later.

The determining unit 33 predicts the Mg yield of the molten metal after processing for each processing cup based on past performance data stored in the data storage unit 31, and determines the amount of Mg to be input. In other words, the determining unit 33 inputs at least past performance data into the prediction model 32A, and the prediction unit 36 predicts the Mg yield, and determines the input Mg amount according to the prediction result (predicted yield) in the prediction unit 36. It includes an arithmetic unit 37 that performs calculations. The method for predicting the Mg yield by the prediction unit 36 will be described in detail later.

The calculation unit 37 calculates the amount of Mg to be added using the following formula.
Amount of Mg added = (target residual Mg x amount of molten metal to be processed) / predicted yield

The instruction output unit 34 instructs the supply device 23 to supply the ladle 15 with the wire 22 having a length corresponding to the input Mg amount determined by the determination unit 33, thereby spheroidizing the graphite.

The result acquisition unit 35 acquires the analysis result of the molten metal component in the ladle 15A after the graphite nodularization process from the QV analyzer 14, and detects the amount of residual Mg. Furthermore, the Mg yield is calculated based on the detected residual Mg amount. The calculated Mg yield is stored in the data storage section 31 as result data.

Note that the functions of the determining section 33, the instruction output section 34, and the result acquisition section 35 can be realized by a processor such as a CPU executing software. The data storage unit 31 and the model storage unit 32 are realized by a nonvolatile memory or a recording medium that is removable from the computer.

<Method for predicting Mg yield>
Referring to FIG. 3, a method for predicting the Mg yield using the prediction model 32A in this embodiment will be described. FIG. 3A conceptually shows a prediction method according to a comparative example, and FIG. 3B conceptually shows a prediction method according to the present embodiment (example).

Referring to FIG. 3(A), in the case of a normal prediction model, machine learning is performed to predict the Mg yield (or residual Mg amount), which is the objective variable, using the process cup information to be processed as an explanatory variable. . The processing cup information corresponds to the molten metal condition data Da, and the Mg yield (actually measured value) corresponds to the result data Db. The molten metal condition data Da indicates, for example, the basic information of each treatment (date, time, product type) and the ratio of each elemental component (C, Si, Mn, P, S, Cu, etc.) of the molten metal before treatment. It includes the source hot water component, the amount of treated hot water specified by the amount of remaining hot water and the amount of tapped hot water, and the source hot water temperature that indicates the temperature of the molten metal before treatment, and further includes the amount of additives.

On the other hand, referring to FIG. 3(B), prediction model 32A according to the present embodiment uses not only processed cup information but also past cup information, that is, performance data, as an explanatory variable to predict Mg yield. has been machine learned. Here, the performance data includes the Mg yield, which is the answer to the objective variable (actual measurement result).

When the performance data for the most recent two cups are included in the explanatory variables, the performance data for the previous two cups includes molten metal condition data Da1 and result data Db1 regarding the spheroidization process for the previous cup, and the performance data for the two cups previous includes the molten metal condition data Da1 and result data Db1 regarding the spheroidization process for the previous cup. It includes molten metal condition data Da2 and result data Db2 regarding the previous spheroidization process.

<How to generate a prediction model>
A method for generating the prediction model 32A will be described with reference to FIG. 2(B). FIG. 2(B) is a block diagram showing the functional configuration of the learning device 100.

The learning device 100 includes a historical data storage section 101, a model generation section 102, a model storage section 103, and an evaluation section 104. The learning device 100 is realized by a computer including a processor such as a CPU (Central Processing Unit) and a memory.

The model generation unit 102 generates processing history data for each spheroidization process in the spheroidization processing device 20, and stores the generated processing history data 1, 2, . . . , n in the history data storage unit 101.

Each processing history data (k) corresponds to the above-mentioned performance data and includes molten metal condition data and result data stored in association with unique identification data. The identification data includes, for example, the date and product type included in the basic information shown in FIG. 3(A). The molten metal condition data includes all or at least one of the amount of treated molten metal, the pre-treatment temperature, and the source molten metal components. The molten metal condition data may also include processing time.

The model generation unit 102 generates a learning model for predicting the Mg yield of the processed cup, that is, a prediction model 103A, based on the plurality of processing history data stored in the history data storage unit 101. Specifically, the prediction model 103A is generated by performing machine learning on the correlation between the molten metal condition data and past performance data (including Mg yield) of the processing cup and the Mg yield, which is the result data of the processing cup. That is, as shown in FIG. 3(B), a prediction model 103A is generated using the molten metal condition data and past performance data of the processing cup as explanatory variables and the Mg yield of the processing cup as an objective variable.

As an algorithm used for machine learning, for example, (multiple) regression analysis may be adopted, but other techniques such as neural network, support vector machine, decision tree, k-NN, etc. may also be adopted.

The evaluation unit 104 evaluates the accuracy of the predictive model 103A generated by the model generation unit 102, and outputs the evaluation result to the model generation unit 102. The model generation unit 102 appropriately modifies the prediction model 103A so as to eliminate the error between the predicted value and the measured value (result) of the Mg yield.

Here, the evaluation unit 104 determines that if the actual measured value (result) is lower than the Mg yield predicted using the prediction model 103A, the error (loss) will be higher than if it is higher by the same amount. It is desirable to highly evaluate the The same applies when the objective variable of the prediction model is the amount of residual Mg in the molten metal after treatment.

The error function (loss function) used to evaluate the prediction model 103A will be described with reference to FIG. 5. As error functions used for machine learning evaluation, the following examples 1 and 2 are common. As shown in the graph of FIG. 5B, the error function of Example 1 expresses the difference between the predicted value and the actual value as an absolute value, and calculates the average of the absolute values as the evaluation result. As shown in the graph of FIG. 5(C), the error function of Example 2 squares the difference between the predicted value and the actual value, and calculates the square root of the average as the evaluation result.

On the other hand, as shown in the graph of FIG. 5(D), the error function of the present embodiment multiplies the value by a constant ( For example, the average of the absolute values is calculated as the evaluation result. Therefore, as shown in FIG. 5A, if there are several cases in which the difference between the predicted value and the measured value is positive (on the + side), the evaluation result (error) by the evaluation unit 104 is It may be higher than the evaluation result (error) of No. 2.

By employing such an evaluation method, it is possible to prevent the quality of the finished product obtained through the subsequent casting process P8 from deteriorating.

A prediction model 103A generated by the model generation unit 102 is stored in advance in the model storage unit 32 of the spheroidization processing device 20 as a prediction model 32A. Note that the spheroidization processing device 20 may have the functions of the learning device 100, so that the prediction model 32A can be updated at any time in the spheroidization processing device 20.

<Simulation results>
FIG. 4(A) is a table showing simulation results when Mg yield is predicted using the prediction model of the comparative example, and FIG. 4(B) is a table showing simulation results when Mg yield is predicted using the prediction model 103A of the present embodiment. 2 is a table showing simulation results.

FIG. 4(A) is a table showing the predicted value and actual measured value (objective variable) of Mg yield according to the comparative example shown in FIG. 3(A) as a difference from the reference value (target value of Mg yield). It is. FIG. 4(B) shows the predicted value and actual measured value (target variable) of Mg yield according to the prediction method in this embodiment shown in FIG. 3(B), as well as the reference value (target value of Mg yield). This is a table showing the difference between

Referring to FIG. 4(A), in the comparative example, only the cup data (processed cup information) is included in the explanatory variable column CA1. The cup data corresponds to the molten metal condition data shown in FIG. 3(A). In the objective variable column CA2, the Mg yield actual value for each time is shown as a difference from the reference value (target value). In the predicted value column CA3, each predicted Mg yield when predicted using the explanatory variables in the column CA1 as input is shown as a difference from a reference value (target value).

Referring to FIG. 4(B), in the present embodiment, the explanatory variable column CB1 includes data from the first cup and data from the second cup in addition to the cup data. The data before one cup and the data before two cups correspond to performance data and include molten metal condition data and yield data (result data). In the objective variable column CB2 and the predicted value column CB3, each value is shown as a difference from a reference value (target value), similarly to the above columns CA2 and CA3.

Evaluation of the simulation results in FIGS. 4(A) and 4(B) is shown below.

As shown in Table 1, the usage ratio of magnesium is the same in both the present example and the comparative example, and the prediction error and target lower limit deviation rate are higher in the comparative example than in the example. The prediction error here is the root mean square error calculated according to Example 2 of FIG. 5, and the difference between the predicted value and the actual value is squared, and the square root of the average is expressed as the error. The target lower limit deviation rate represents the rate at which the predicted value deviates from the target lower limit and becomes lower. The usage ratio of magnesium represents the amount of Mg when the amount of Mg for spheroidization treatment (input amount) used when operating according to the actual measured value of Mg yield (answer) is 1.

The above evaluation results show that by including past performance data in the explanatory variables of the prediction model 103A, as in this embodiment, it is possible to predict the Mg yield with greater accuracy than in the comparative example. In particular, it was found that the prediction accuracy can be improved by including the objective variable, the Mg yield (actual measured value), in the performance data.

<Operation of spheroidization processing device>
The operation of the spheroidization processing apparatus 20 will be described with reference to FIG. 2(A) and FIG. 6. FIG. 6 is a flowchart showing the graphite nodularization process in this embodiment.

First, the determination unit 33 of the spheroidizing treatment device 20 acquires molten metal condition data for the treatment cup from the QV analysis device 14 or the like, and acquires performance data for the past two cups stored in the data storage unit 31 (step S1).

The prediction unit 36 of the determination unit 33 predicts the Mg yield of the processed glass based on the information acquired in step S1 (step S3). That is, the molten metal condition data for the treated cup and the performance data for the past two cups are input into the prediction model 32A, and the output value of the prediction model 32A is obtained as the predicted value.

Subsequently, the calculation unit 37 of the determination unit 33 determines the input Mg amount based on the Mg yield predicted in step S3, the predetermined target residual Mg amount, and the processing amount included in the molten metal condition data of the processing cup. Calculate (step S5).

The instruction output unit 34 instructs the supply device 23 of the cored wire device 21 to supply wire according to the input Mg amount calculated in step S5 (step S7). As a result, an appropriate amount of magnesium is automatically added to the molten metal in the ladle 15, and the graphite in the molten metal is spheroidized (step S9).

When the spheroidization of graphite is completed, the result acquisition unit 35 acquires the component analysis result of the molten metal in the ladle 15A, and detects the amount of residual Mg (step S11). Furthermore, the Mg yield (actual measurement value) is calculated based on the detected residual Mg amount, the processed hot water amount, and the input Mg amount calculated in step S5 (step S13).

The result acquisition unit 35 stores the calculated Mg yield (actual measurement value) in the data storage unit 31 (step S15). That is, performance data including the molten metal condition data of the processed cup acquired in step S1 and the result data obtained in step S13 is stored in the data storage unit 31 as the most recent (one cup previous) performance data.

When this process is finished, the process returns to step S1 and the above process is repeated.

<Modified example>
(1) Past performance data used for predicting Mg yield is not limited to the previous two cups. For example, performance data from one cup before and three cups ago may be used. Note that it is desirable that the track record data used for predicting the Mg yield be track record data for the same product type.

(2) The past performance data used for predicting the Mg yield is preferably a plurality of cups, but is not limited to two cups, and may be three or more cups. Alternatively, the Mg yield may be predicted based on past performance data for one cup.

(3) Although the past performance data used for predicting the Mg yield includes both the molten metal condition data and the result data of the processing cup, it may include only the result data.

(4) Although both the molten metal condition data of the processing cup and the past performance data are used to predict the Mg yield, it is sufficient that at least the past performance data is used. That is, the explanatory variables of the prediction model 103A may be only past performance data.

(5) In the present embodiment, the method in which the control device 25 of the spheroidization processing apparatus 20 determines the amount of Mg to be introduced using the prediction model 103A has been described. It can also be executed on a computer independent of 20.

(6) The method for determining the amount of Mg to be introduced, which is executed by the control device 25 of the spheroidization processing device 20, can also be provided as a program. Similarly, the learning method executed by the learning device 100 can also be provided as a program. Such a program can be provided by being recorded on an optical medium such as a CD-ROM (Compact Disc-ROM) or a computer-readable non-transitory recording medium such as a memory card. The program can also be provided by downloading over a network.

The program according to the present invention may execute processing by calling necessary modules in a predetermined sequence at a predetermined timing among program modules provided as part of a computer operating system (OS). . In that case, the program itself does not include the above module and processes are executed in cooperation with the OS. A program that does not include such a module may also be included in the program according to the present invention.

Further, the program according to the present invention may be provided by being incorporated into a part of another program. Even in that case, the program itself does not include the modules included in the other programs, and the processing is executed in cooperation with the other programs. Programs incorporated into such other programs may also be included in the programs according to the present invention.

The embodiments disclosed this time should be considered to be illustrative in all respects and not restrictive. The scope of the present invention is indicated by the claims rather than the above description, and it is intended that all changes within the meaning and range equivalent to the claims are included.

20 spheroidization processing device, 21 cored wire device, 23 supply device, 25 control device, 31 data storage section, 32, 103 model storage section, 32A, 103A prediction model, 33 determination section, 34 instruction output section, 35 result acquisition section , 36 prediction unit, 37 calculation unit, 100 learning device, 101 history data storage unit, 102 model generation unit, 104 evaluation unit.

Claims (7)

A graphite spheroidization treatment device equipped with a magnesium supply means for introducing magnesium into molten metal in a ladle,
a data storage means for storing performance data including a combination of the amount of magnesium input in a past cup and the magnesium content of the molten metal after graphite spheroidization treatment, or result data indicating the yield;
Predicting the magnesium content or yield of the molten metal after treatment for each processing cup , using a predetermined number of recent performance data including the result data among past performance data stored in the data storage means , A graphite nodularization processing apparatus, comprising a determining means for determining the amount of magnesium input from the magnesium supply means based on the prediction result, a predetermined target residual magnesium amount, and the amount of hot water to be processed in the processing cup . The determining means determines the magnesium content of the molten metal after treatment based on molten metal condition data of the treatment cup including at least one of the amount of molten metal to be treated, the temperature before treatment, and the source hot water component, and the latest predetermined number of performance data. The graphite nodularization processing apparatus according to claim 1, further comprising a prediction means for predicting the amount or yield. 3. The graphite nodularization processing apparatus according to claim 2, wherein the performance data stored in the data storage means further includes the molten metal condition data in addition to the result data. Further comprising a model storage means for storing the learning model generated by machine learning,
4. The graphite spherical according to claim 1, wherein the determining means inputs at least the latest predetermined number of performance data into the learning model to predict the magnesium content or yield of the molten metal after treatment. processing equipment. A combination of the amount of magnesium input and the magnesium content of the molten metal after the graphite spheroidization treatment, or the yield rate, for each graphite spheroidization treatment in a graphite spheroidization treatment apparatus equipped with a magnesium supply means for introducing magnesium into the molten metal in a ladle. historical data storage means for storing past performance data including result data indicating;
Among the past performance data stored in the history data storage means, the correlation between a predetermined number of recent performance data including the result data and the magnesium content or yield of the molten metal after processing in the processing cup is machine learned. a generation means for generating a learning model;
A learning device comprising: model storage means for storing the learning model generated by the generation means. Further comprising an evaluation means for evaluating the learning model,
The evaluation means increases the error when the actual measured value is lower than the magnesium content or yield of the molten metal after treatment predicted using the learning model than when it is higher by the same amount. The learning device according to claim 5, which evaluates. A method for determining an amount of magnesium to be input in a graphite spheroidizing treatment device, comprising:
storing performance data in a storage unit, the performance data including a combination of the amount of magnesium charged in a past cup and the magnesium content of the molten metal after the graphite spheroidizing treatment, or result data indicating a yield;
predicting the magnesium content or yield of the molten metal after processing for each processing cup using a predetermined number of most recent performance data including the result data from among the past performance data stored in the memory unit;
A method for determining an amount of magnesium to be added , comprising a step of determining an amount of magnesium to be added to a treatment cup based on the prediction result in the prediction step, a predetermined target amount of residual magnesium, and the amount of hot water to be treated in the treatment cup.

Images