WO2022239157A1

WO2022239157A1 - Learning control device for rolling process

Info

Publication number: WO2022239157A1
Application number: PCT/JP2021/018091
Authority: WO
Inventors: 真康関本
Original assignee: 東芝三菱電機産業システム株式会社
Priority date: 2021-05-12
Filing date: 2021-05-12
Publication date: 2022-11-17
Also published as: CN115623864A; JPWO2022239157A1; JP7323051B2

Abstract

The learning control device according to an embodiment of the present invention calculates a predicted value on the basis of an actual value measured in a rolling process, calculates an instantaneous value of a learning rate on the basis of the difference between the calculated predicted value and the actual value of the rolling process, calculates an updated value of the learning rate on the basis of the instantaneous value of the learning rate and a previous value of a cell having corresponding rolling conditions in a learning rate table, updates the learning rate of the cell having the corresponding rolling conditions in the learning rate table, identifies an occurrence time of a sudden change in the learning rate on the basis of the instantaneous value of the learning rate stored in a rolling information database, detects a deviation from a standard of the instantaneous value of the learning rate before and after the occurrence time of the sudden change in the learning rate as a learning rate sudden change component, and re-updates the learning rate in the learning table by a correction based on the instantaneous value of the learning rate stored in the rolling information database at and after the identified occurrence time of the sudden change of the learning rate and on the learning rate sudden change component.

Description

Learning controller for rolling process

The present invention relates to a learning control device for rolling processes.

At the rolling mill, for example, steel materials, aluminum, copper and other non-ferrous materials are rolled to produce metal strips used in the manufacture of automobiles and electrical products. In addition, there are various types of rolling mills (rolling processes) such as hot thin plate rolling mills, thick plate rolling mills, cold rolling mills, and rolling mills for rolling wire rods.

In any rolling process, control is performed so that the product after the completion of manufacturing matches the target values such as the desired dimensions, shape, and temperature that affect the mechanical properties. Generally, the control of the rolling process includes setting control and dynamic control.

In the setting control, using a mathematical model that reproduces the phenomena during the rolling process, the speed of the rolling mill, the flow rate of the cooling water, the gap between the rolling rolls, and other equipment are adjusted so that the material to be rolled has the desired dimensions and temperature. A set value has been determined. Here, the mathematical model is often simplified from the viewpoint of reducing the computational load.

Therefore, there may be a deviation between the predicted value of the target rolling phenomenon calculated by the mathematical model and the actual value measured by the measuring instrument installed in the facility. The deviation between the predicted value and the actual value appears as an error in the product target size and temperature, which causes the product to become a defective product outside the allowable range of quality assurance.

Furthermore, in recent years, the demands for product specifications have become more sophisticated and diversified, and strict control is required for product quality assurance.

In addition, in rolling process control, a learning coefficient is provided in the mathematical model, and by adjusting the learning coefficient based on the deviation between the predicted value and the actual value, the accuracy of prediction by the mathematical model is improved and stabilized.

In general, for the variables in the mathematical model, the learning coefficient is determined by comparing the predicted value of the prediction target obtained from the actual value and the actual value of the prediction target. The learning coefficient obtained here is the learning coefficient for the material to be rolled, that is, the instantaneous value.

In addition, the instantaneous value of the learning coefficient varies greatly due to factors omitted due to the simplification of the mathematical model, measurement errors of measuring instruments, and various disturbances in the rolling process. Therefore, the instantaneous value of the learning coefficient is applied as an update value after being smoothed.

The updated value of the learning coefficient is generally determined based on the rolling conditions, which are processing conditions such as product target thickness and width, temperature, material composition, rolling reduction, and number of processing passes. , which are called “cells”).

In other words, the learning control device for the rolling process can acquire appropriate learning coefficients corresponding to the rolling conditions by using a table classified according to the rolling conditions. In this way, the rolling process learning control device uses an appropriate learning coefficient to improve the prediction accuracy of the rolling phenomenon by the mathematical model and to ensure rolling stability.

In addition, even when simply updating the learning coefficient of one cell in the table corresponding to the rolling conditions, in order to sufficiently update and converge the learning coefficients of all the cells in the table that can be handled in terms of operation. requires a large amount of rolling experience.

As a countermeasure, for example, as described in Patent Document 1, when updating the learning coefficient of one cell in the table corresponding to the rolling condition, the learning coefficient of the cell with similar rolling conditions is also updated at the same time. It has been known. Here, multiple cells adjacent to the relevant cell are updated.

According to this method, it is possible to sufficiently update the learning coefficient with a smaller number of rolling records. can be secured.

On the other hand, learning control using a table classified by rolling conditions has the problem that it is difficult to follow chronological changes in the rolling process. For example, even if the learning coefficient in the table that can be handled in terms of operation is sufficiently updated, if there is a cell that has not been rolled under the relevant rolling conditions for a while, if the rolling process changes, the learning coefficient will not be appropriate. Therefore, there is a possibility that the prediction accuracy may be remarkably lowered under the rolling conditions.

For this problem, for example, a method disclosed in Patent Document 2 has been proposed. For example, in Patent Document 2, a time-series learning coefficient that compensates for an error caused by a time-series change included in the deviation between a predicted value and an actual value by a mathematical model is separated from a learning coefficient corresponding to the rolling conditions. However, it is proposed to modify the predicted value based on these two types of learning coefficients.

The time-series change here means a behavior that changes linearly, such as the influence of reduction in roll diameter due to wear caused by rolling friction of rolling rolls. According to this method, it is possible to appropriately obtain the learning coefficient for each rolling condition excluding such chronological changes in the rolling process.

Japanese Patent Laid-Open No. 6-259107 Japanese Patent Laid-Open No. 4-367901

Prediction errors in setting calculations, which are one of the causes of hot rolling quality defects, include factors such as mechanical errors and measurement errors, in addition to prediction errors in mathematical models that represent rolling material deformation characteristics. Of these, mechanical errors and measurement errors may occur suddenly due to equipment failure, roll change, repair/replacement, poor calibration, operator's erroneous operation, or change in weather conditions.

However, it is extremely difficult to immediately determine the cause of these sudden errors from the information measured by the sensors on the rolling line, and several to several dozen coils are produced after the cause of the error occurs. In many cases, the error factors are identified after

In this case, the prediction errors in the setting calculations detected by sensors, etc. on the rolling line between the time these error factors occur and the time they are identified and appropriate countermeasures are taken are assumed to be due to the prediction errors in the mathematical model. be learned.

In the conventional learning method for each coil and the learning method based on past chronological changes, if the errors of these different factors are regarded as the prediction errors of the formula model and learned, the formula for the original target event There is a possibility that the model error becomes a disturbance to the learning, the learning coefficient is updated incorrectly, and the prediction accuracy of the setting calculation decreases.

Furthermore, since the inaccurate learning results during that time are sequentially written in the learning table classified by the rolling conditions, the influence differs for each learning division due to the difference in update frequency. Even if the effect could be identified, it would be difficult to repair the learning table. In other words, there is a problem that inaccurate learning results are continuously used, resulting in poor quality.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and provides a learning control device for a rolling process that can correct subsequent learning even if an error factor occurs due to sudden fluctuations. for the purpose.

A learning control device for a rolling process according to an aspect of the present invention updates learning coefficients of a mathematical model used to calculate set values for a rolling process, using a learning coefficient table configured by a plurality of cells classified according to rolling conditions. a prediction value calculation unit for calculating a prediction value based on actual values measured in the rolling process; and a prediction value calculated by the prediction value calculation unit. , an instantaneous value calculating unit for calculating an instantaneous value of the learning coefficient based on the difference from the actual value of the rolling process; the instantaneous value of the learning coefficient calculated by the instantaneous value calculating unit; an updating unit for calculating an updated value of the learning coefficient based on the previous value of the corresponding cell and updating the learning coefficient of the cell to which the rolling condition of the learning coefficient table corresponds; and the learning coefficient calculated by the instantaneous value calculating unit. instantaneous value of , previous value of learning coefficient, updated value of learning coefficient, date and time information identifying rolling material, rolling conditions, cell coordinates based on rolling conditions in the learning coefficient table, actual values in rolling process, and rolling process Based on the rolling information database that stores the date and time history of events and the instantaneous value of the learning coefficient stored in the rolling information database, the point of occurrence of the sudden change in the learning coefficient is specified, and the time before and after the occurrence of the sudden change of the learning coefficient is determined. a sudden change detection unit that detects a deviation in the level of the instantaneous value of the learning coefficient in the rolling information database as a sudden change component of the learning coefficient; and a re-updating unit for re-updating the learning coefficient of the learning coefficient table by correction based on the sudden change component of the learning coefficient.

Further, the learning control device for a rolling process according to an aspect of the present invention further includes a notification unit that notifies the necessity of maintenance as an event when the sudden change detection unit identifies the point in time when the sudden change in the learning coefficient occurs. .

Further, in the rolling process learning control device according to an aspect of the present invention, the re-updating unit performs maintenance that becomes an event in the rolling information database after the sudden change detection unit identifies the point in time when the sudden change in the learning coefficient occurs. If there is no date and time history, based on the instantaneous value of the learning coefficient in rolling after the occurrence of the sudden change in the learning coefficient, the sudden change component of the learning coefficient, and the previous value of the learning coefficient of the cell based on the rolling conditions of the learning coefficient table, Re-update the learning coefficients in the learning coefficient table.

According to the present invention, subsequent learning can be corrected even if an error factor occurs due to sudden fluctuations.

It is a figure which illustrates the structure of the learning control apparatus of the rolling process concerning 1st Embodiment. 4 is a diagram exemplifying data stored in a rolling information database; FIG. FIG. 4 is a diagram illustrating data stored in a learning coefficient table and an event-by-event learning coefficient table; 6 is a flowchart illustrating processing performed by a sudden change detection unit; FIG. 10 is a diagram exemplifying a result of specifying and detecting a sudden change in a learning coefficient by a sudden change detection unit; (a) is a diagram exemplifying a result of change point detection based on a change in an instantaneous value of a learning coefficient. (b) is a diagram illustrating likelihood trends in a change point detection method using the maximum likelihood and the least squares method. (c) is a diagram exemplifying the trend of the absolute value of the degree of change in change point detection using the cumulative sum. 9 is a flow chart showing a specific example of processing performed by a re-updating unit; FIG. 7 is a diagram illustrating the configuration of a rolling process learning control device according to a second embodiment;

An embodiment of a rolling process learning control device will be described below with reference to the drawings. FIG. 1 is a diagram illustrating the configuration of a rolling process learning control device 1 according to the first embodiment.

The learning control device 1 has a function as a computer equipped with a CPU and a memory (not shown). It is a device that controls the rolling process.

Then, the learning control device 1 updates and saves the learning coefficients of the mathematical model used for calculating the set values for the rolling process, using a learning coefficient table composed of a plurality of cells classified according to the rolling conditions, and performs the rolling process. Control. Specifically, the learning control device 1 has a storage unit 2, a learning unit 3, a setting calculation unit 4, and a learning coefficient re-updating unit 5, for example.

The storage unit 2 is a device that stores, for example, a rolling information database (DB) 20, a learning coefficient table 22, and an event-by-event learning coefficient table 24.

For example, as shown in FIG. 2, the rolling information database 20 stores, for each rolled material, the manufacturing number, manufacturing date and time, and rolling conditions, as well as the instantaneous value and previous value of the learning coefficient of the cell to be updated, the updated value, and , is a database that stores coordinate information of cells based on the rolling conditions of the learning coefficient table.

In addition, the rolling information database 20 includes production date information (or the production number of the rolled material) that identifies the rolled material, a history of event occurrence dates such as maintenance, periodic inspection, or equipment replacement for the rolling process, and target rolling phenomena. You may save the actual value etc. in the rolling process used at the time of prediction.

As shown in FIG. 3, the learning coefficient table 22 is composed of a plurality of cells divided by rolling conditions, and is a table that records (saves) learning coefficients in cells corresponding to the rolling conditions.

As described above, the instantaneous value of the learning coefficient varies greatly due to various disturbances. Therefore, the updated value of the learning coefficient is obtained using the smoothed instantaneous value of the learning coefficient. For example, the updated value of the learning coefficient is calculated by the following formula (1).

Although the coordinates of the cell in the learning coefficient table 22 are described as two variables here, the number of variables does not depend on this. That is, for example, when steel grades are used as classifications in addition to classifications of product strip target width values and product strip thickness target values, the coordinates of the cell are determined by three variables.

The obtained update value of the learning coefficient is recorded so as to update the learning coefficient of the update target cell in the learning coefficient table 22 . Further, the rolling information database 20 records the cell coordinate information for the rolling conditions in the learning coefficient table 22 , the previous value of the learning coefficient, and the actual value used when calculating the predicted value of the prediction target.

In the process of updating the learning coefficient, multiple cells adjacent to the cell to be updated may be updated at the same time to promote updating of the learning coefficient table 22 as a whole. For example, the learning coefficients of adjacent cells may be updated as shown in Equation (2) below.

Also, the learning coefficients recorded in the learning coefficient table 22 are copied to the event-by-event learning coefficient table 24 for each event such as maintenance such as regular inspection of equipment and equipment replacement. That is, the configuration of the event-by-event learning coefficient table 24 is the same as the configuration of the learning coefficient table 22 shown in FIG.

The learning unit 3 (FIG. 1) is a device having a predicted value calculating unit 30, an instantaneous value calculating unit 32, and an updating unit 34.

The predicted value calculation unit 30 calculates a predicted value to be predicted based on the actual values measured in the rolling process for the mathematical model used for the setting calculation, and outputs the calculated value to the instantaneous value calculation unit 32 .

The instantaneous value calculation unit 32 calculates the instantaneous value of the learning coefficient based on the actual value to be predicted, outputs the calculated instantaneous value to the updating unit 34, and stores the instantaneous value together with the rolling conditions in the rolling information database 20. Save to For example, the instantaneous value calculator 32 calculates the instantaneous value of the learning coefficient based on the difference between the predicted value calculated by the predicted value calculator 30 and the actual value of the rolling process.

The updating unit 34 calculates updated values of the learning coefficients based on the instantaneous values calculated by the instantaneous value calculating unit 32, and outputs the calculated updated values to the rolling information database 20 and the learning coefficient table 22. For example, the updating unit 34 updates the learning coefficient based on the instantaneous value of the learning coefficient calculated by the instantaneous value calculating unit 32 and the learning coefficient (previous value) of the update target cell to which the rolling condition in the learning coefficient table 22 corresponds. A value is calculated, and the learning coefficient of the cell corresponding to the rolling condition in the learning coefficient table 22 is updated.

The setting calculation unit 4 has a learning coefficient reading unit 40 and a setting calculation unit 42, and is a device that determines setting values for each piece of equipment using a mathematical model that reproduces the phenomenon of the rolling process.

The learning coefficient reading unit 40 reads the learning coefficient of the cell corresponding to the rolling condition in the learning coefficient table 22 and outputs it to the setting calculation unit 42 in order to improve the prediction accuracy.

The setting calculation unit 42 corrects the setting value for each piece of equipment using the learning coefficient output by the learning coefficient reading unit 40, and outputs the corrected setting value to each piece of equipment.

The learning coefficient re-updating unit 5 includes a sudden change detection unit 50, a determination unit 52, and a re-updating unit 54, detects a sudden change in the learning coefficient, specifies the time of occurrence, calculates the sudden change component, After the sudden change of the learning coefficient occurs, the learning coefficient stored in the learning coefficient table 22 is corrected using the sudden change component (learning coefficient sudden change component).

For example, the sudden change detection unit 50 has a change point detection function for detecting a sudden change in the learning coefficient based on the instantaneous value of the learning coefficient stored in the rolling information database 20. The point of occurrence of the sudden change is identified, and the deviation of the level of the instantaneous value of the learning coefficient before and after the point of occurrence of the sudden change is calculated as the sudden change component of the learning coefficient.

FIG. 4 is a flowchart illustrating processing performed by the sudden change detection unit 50. FIG. First, after the rolling of the material to be rolled is finished, the sudden change detection unit 50 acquires the instantaneous values of the learning coefficients corresponding to the rolling conditions I and J from the rolling information database 20 (S100).

Then, the sudden change detection unit 50 determines whether or not the number of acquired instantaneous value data has reached N (S102). is reached (S102: Yes), the process proceeds to S104.

That is, the sudden change detection unit 50 acquires the instantaneous values of the learning coefficients for the rolling number N from the rolling information database 20 over the past. At this time, the sudden change detection unit 50 may acquire the instantaneous values of the learning coefficients not only for one section corresponding to the rolling condition, but also for adjacent sections.

In this case, the condition for acquiring the instantaneous value of the learning coefficient is expressed as the following formula (3).

It is desirable that the rolling number N from which the sudden change detection unit 50 acquires the instantaneous value of the learning coefficient includes about several hundred rolling rolls past the event occurrence point _nE . Here, the acquisition conditions are the relevant rolling conditions and their adjacent divisions. This is for time-series analysis of the learning coefficients of the same level when the learning coefficients of the sections that are not adjacent to the rolling conditions are completely different values. If the learning coefficients are at the same level regardless of the rolling conditions, the sudden change detection unit 50 may acquire all the learning coefficients in chronological order.

Based on the instantaneous value of the learning coefficient thus obtained, the sudden change detection unit 50 detects a sudden change in the learning coefficient. In addition, the sudden change detection unit 50 uses a general change point detection method to identify the presence or absence of a sudden change in the learning coefficient and the point in time when it occurs. Methods of detecting change points include, for example, a method using maximum likelihood and least squares method, a method using cumulative sum, and the like. Here, the two methods described above will be described.

The change point detection method using the maximum likelihood and the least squares method is that when the transition of the instantaneous value of the learning coefficient of the rolling condition and its adjacent conditions is divided into τth intervals, the likelihood in the interval before and after τ is This is a method of finding the maximum or minimum point in time.

The specific contents are shown below. Here, the change point τ and the transition of the instantaneous value of the learning coefficient of the rolling condition and its adjacent conditions are defined as in the following equation (4).

At this time, μ1, μ2, and τ are determined so as to minimize the likelihood U by the method of least squares as follows. Although the residual sum of squares is used as the likelihood here, it is not limited to this. Note that yk indicates the _k -th ZCURRENT.

In addition, the change point detection method using the cumulative sum accumulates the degree of change between numerical values along time or along the data group arranged in chronological order, and determines an abnormality when the cumulative sum exceeds the threshold. method. The degree of change Sc is calculated by the following formula (10).

At this time, the time of change is the time when the absolute value of Sc(n) is maximum, as shown in the following equation (11).

Furthermore, the learning coefficient sudden change component (difference) in this method is calculated as shown in the following formula (12).

Also, the deviation of the average value of the learning coefficients before and after the change point is detected when the instantaneous value of the learning coefficient changes suddenly, as shown in FIG.

In this way, the sudden change detection unit 50 acquires the point of change of the learning coefficient and the deviation of the average value of the learning coefficients before and after the point of change (S104: FIG. 4).

FIG. 6 is a diagram illustrating the results of change point detection using the data of the hot rolling plant according to the method described above. FIG. 6(a) is a diagram illustrating the result of performing change point detection based on changes in the instantaneous value (average value) of the learning coefficient. FIG. 6B is a diagram exemplifying a likelihood trend in a change point detection method using maximum likelihood and least squares method. FIG. 6C is a diagram exemplifying the trend of the absolute value of the degree of change in change point detection using the cumulative sum. Note that the object of change point detection is the instantaneous value of the learning coefficient in the mathematical model for product width prediction.

In FIG. 6(a), about 10,000 learning coefficients, including not only one section corresponding to arbitrary rolling conditions but also adjacent sections, are acquired and the trend is shown. In other words, FIG. 6(a) shows that the learning coefficient suddenly fluctuates near the center of the trend.

Further, as shown in FIGS. 6B and 6C, the minimum likelihood value in the change point detection method using the maximum likelihood and the least squares method and the change in change point detection using the cumulative sum The maximum absolute value of the degree appears at the same time point. They coincide with the times when sudden fluctuations in the learning coefficient appear, and appropriately capture the sudden change times.

Note that the change point detection method described here is just an example, and the change point detection method that can be applied to the present invention is not limited to this.

In this way, the sudden change detection unit 50 identifies the time point at which the sudden change in the learning coefficient occurs based on the instantaneous value of the learning coefficient stored in the rolling information database 20, and performs learning before and after the time point at which the sudden change in the learning coefficient occurs. The deviation of the level of the instantaneous value of the coefficient is detected as the sudden change component of the learning coefficient.

The determination unit 52 (FIG. 1) determines whether or not the sudden change component of the learning coefficient at the change point detected by the sudden change detection unit 50 is greater than or equal to the sudden change determination threshold value ε.

The re-updating unit 54 re-updates the learning coefficients in the learning coefficient table 22 when the determining unit 52 determines that the sudden change component of the learning coefficient is equal to or greater than the sudden change determination threshold value ε. For example, the re-update unit 54 re-updates the learning coefficient after the sudden change time based on the change time detected by the sudden change detection unit 50 and the sudden change component of the learning coefficient. At this time, the re-updating unit 54 separately stores the sudden change component of the learning coefficient.

FIG. 7 is a flowchart showing a specific example of processing performed by the re-update unit 54. FIG. First, the re-update unit 54 acquires the learning coefficient table 22 from the storage unit 2 (S200).

Next, the re-update unit 54 acquires the instantaneous value of the learning coefficient and the cell coordinate information for the rolling conditions in the learning coefficient table 22 from the rolling information database 20 (S202).

Then, the re-update unit 54 determines whether or not the instantaneous value of the learning coefficient is the instantaneous value after the sudden change based on the event occurrence date and time history (S204). If the instantaneous value is after the sudden change (S204: Yes), the re-update unit 54 proceeds to the process of S206, and if the instantaneous value is not after the sudden change (S204: No), proceeds to the process of S208.

In S206, the re-updating unit 54 corrects the instantaneous value of the learning coefficient by adding the sudden change component of the learning coefficient, and calculates the updated value.

Also, in S208, the re-updating unit 54 calculates an updated value using the instantaneous value of the learning coefficient.

Then, the re-update unit 54 re-updates the corresponding cell of the learning coefficient table 22 using the learning coefficient stored in the event-by-event learning coefficient table 24 as the previous value (S210).

The update procedure performed by the re-update unit 54 satisfies the conditions shown in the following formula (13).

In updating the learning coefficients in the learning coefficient table 22, if multiple adjacent cells are updated at the same time, the multiple adjacent cells are updated by the same procedure as shown in the following formula (14).

That is, the re-updating unit 54 updates the learning coefficient by correction based on the instantaneous value of the learning coefficient stored in the rolling information database 20 after the occurrence of the sudden change in the learning coefficient specified by the sudden change detection unit 50 and the sudden change component of the learning coefficient. Re-update the learning coefficients in Table 22.

In this way, the learning coefficient re-updating unit 5 overwrites the learning coefficient table 22 with the re-updated learning coefficient of the event-by-event learning coefficient table 24 to correct the sudden change component of the learning coefficient.

Next, a second embodiment of the rolling process learning control device will be described. FIG. 8 is a diagram illustrating the configuration of a rolling process learning control device 1a according to the second embodiment.

For example, the learning control device 1a has a storage unit 2, a learning unit 3, a setting calculation unit 4, and a learning coefficient re-updating unit 5a. In the learning control device 1a shown in FIG. 8, the same reference numerals are given to the substantially same configuration as the learning control device 1 shown in FIG.

The learning control device 1a detects a sudden change in the learning coefficient, and when the learning coefficient is re-updated, notifies the operator of it and prompts maintenance. In addition, the learning control device 1a corrects the learning coefficient until the next event such as maintenance occurs when the maintenance is not carried out despite the fact that the maintenance is being urged. The learning control device 1a also has a function of separately storing a history _TN of the date and time when a sudden change in the learning coefficient was detected.

The learning coefficient re-update unit 5a includes a sudden change detection unit 50, a determination unit 52, a re-update unit 54, and a notification unit 56, detects a sudden change in the learning coefficient, identifies the time of occurrence, and detects the sudden change component. is calculated, and the sudden change component is corrected for the learning coefficient stored in the learning coefficient table 22 when maintenance is not performed after the occurrence of the sudden change of the learning coefficient.

The notification unit 56 has a function of notifying the operator of a sudden change in the learning coefficient detected by the sudden change detection unit 50. For example, when the learning coefficient suddenly fluctuates, the notification unit 56 outputs to a human-machine interface (not shown) for operation that maintenance such as equipment inspection and replacement is required. In addition, the notification unit 56 may make an alarm sound using an alarm sound generating device, or may make an announcement to the operator by another method that is easy for the operator to notice.

In other words, the notification unit 56 notifies the operator of the need for maintenance such as inspection and replacement of the equipment that will be the event when the sudden change detection unit 50 identifies the point in time when the sudden change in the learning coefficient occurs.

In this way, the learning control device 1a detects a sudden change in the learning coefficient in the calculation of the updated value of the learning coefficient after the next rolled material, determines whether or not maintenance has been performed since then, and determines whether or not the maintenance has been performed. If not, the instantaneous value of the learning coefficient is corrected based on the sudden change component of the learning coefficient, and the updated value of the learning coefficient is obtained as follows.

Specifically, if the rolling information database 20 does not have a date and time history of maintenance that becomes an event after the sudden change detection unit 50 specified the point in time when the sudden change in the learning coefficient occurred, the re-updating unit 54 detects the sudden change in the learning coefficient. The learning coefficient in the learning coefficient table 22 is re-updated based on the instantaneous value of the learning coefficient in rolling after the occurrence of , the sudden change component of the learning coefficient, and the previous value of the learning coefficient of the cell based on the rolling conditions in the learning coefficient table 22. .

Then, the learning control device 1a updates the learning coefficient table 22 based on the updated value of the learning coefficient. As a result, the learning control device 1a subsequently performs setting calculations using the learning coefficients of the same level.

As explained above, according to the present invention, even if an error factor occurs due to sudden fluctuations, subsequent learning can be corrected. For example, according to the present invention, even if an abnormality due to a mechanical factor such as equipment calibration failure in maintenance such as periodic inspection of equipment or equipment replacement, or continuous measurement abnormality due to instrument abnormality occurs, before the abnormality occurs Based on the saved learning factors, the learning factors can be modified. By minimizing the influence of the abnormality, the present invention reduces continuous deterioration of product accuracy and enables stable rolling.

The learning control devices 1 and 1a periodically analyze the data of a large number of rolled coils, detect sudden changes in prediction errors (learning values) due to mechanical factors and measurement abnormalities, and specify the points in time. Then, the learning control devices 1 and 1a restore the values stored immediately before the sudden change in the learning table based on the separately stored learning table for each event such as periodic repair or roll change, and Using the learning value difference as an offset, the learning values from the sudden change to the current rolled material are updated again.

Each function provided in the learning control devices 1 and 1a may be partially or wholly configured by hardware such as a PLD (Programmable Logic Device) or FPGA (Field Programmable Gate Array), or a processor such as a CPU. may be configured as a program executed by

1, 1a... Learning control device, 2... Storage unit, 3... Learning unit, 4... Setting calculation unit, 5, 5a... Learning coefficient re-updating unit, 20... Rolling information Database 22 Learning coefficient table 24 Learning coefficient table for each event 30 Predicted value calculator 32 Instantaneous value calculator 34 Update unit 40 Learning Coefficient reading unit 42 Setting calculation unit 50 Sudden change detection unit 52 Judgment unit 54 Re-update unit 56 Notification unit

Claims

A learning control device for a rolling process that controls the rolling process by updating and storing the learning coefficients of the mathematical model used to calculate the set values for the rolling process using a learning coefficient table composed of a plurality of cells classified according to the rolling conditions. in
a predicted value calculation unit that calculates a predicted value based on actual values measured in the rolling process;
an instantaneous value calculator that calculates an instantaneous value of the learning coefficient based on the difference between the predicted value calculated by the predicted value calculator and the actual value of the rolling process;
An updated value of the learning coefficient is calculated based on the instantaneous value of the learning coefficient calculated by the instantaneous value calculation unit and the previous value of the cell to which the rolling condition of the learning coefficient table corresponds, and the rolling condition of the learning coefficient table is an updating unit that updates the learning coefficient of the corresponding cell;
the instantaneous value of the learning coefficient calculated by the instantaneous value calculating unit, the previous value of the learning coefficient, the updated value of the learning coefficient, date and time information specifying the rolled material, rolling conditions, cell coordinates based on the rolling conditions of the learning coefficient table, a rolling information database that stores actual values in the rolling process and the date and time history of events for the rolling process;
Based on the instantaneous value of the learning coefficient stored in the rolling information database, the point of occurrence of the sudden change of the learning coefficient is specified, and the deviation of the level of the instantaneous value of the learning coefficient before and after the occurrence of the sudden change of the learning coefficient is learned. a sudden change detection unit that detects as a sudden change component of the coefficient;
The learning coefficient in the learning coefficient table is re-set by correction based on the instantaneous value of the learning coefficient stored in the rolling information database after the occurrence of the sudden change in the learning coefficient specified by the sudden change detection unit and the sudden change component of the learning coefficient. A learning control device for a rolling process, comprising: a re-updating unit for updating;
2. The learning control device for a rolling process according to claim 1, further comprising a notification unit that notifies the necessity of maintenance as an event when the sudden change detection unit specifies the point in time when the sudden change in the learning coefficient occurs. .
The re-updating unit
If the rolling information database does not have a date and time history of maintenance that becomes an event after the sudden change detection unit specified the time point of occurrence of the sudden change in the learning coefficient, the instantaneous value of the learning coefficient in rolling after the time point of occurrence of the sudden change in the learning coefficient 3. The learning coefficient of the learning coefficient table is re-updated based on the previous value of the learning coefficient of the cell based on the rolling condition of the rolling condition of the learning coefficient sudden change component and the learning coefficient table. Learning controller for rolling process.