JPH04367901A - Learning control method in process lens - Google Patents

Abstract

PURPOSE:To execute the processing of a process line with high control precision by correcting appropriately a shift of a phenomenon and a model at the time of controlling the process line. CONSTITUTION:With regard to a material processed by a process line, a table for storing a shift of an actual result value and a model calculation value in each senate group is generated, and based on actual result operation data of the present material, a model calculation value Ycal is calculated, a shift F of Ycal and an actual result value Yact of the present material is calculated, a value Fg obtained by eliminating a smooth value Fto of a shift like a time series of an actual result to the previous material and the model calculation value from the shift F is calculated, that which is obtained by smoothing a table value Fgo of a separate group learning coefficient of the present material by a value Fg is stored as a new table value, a value obtained by smoothing the table value Fgo of the separate group learning coefficient of the present material from the shift F is calculated, that which is obtained by smoothing the smooth value Fto by the value Ft is stored as a new time series learning coefficient, and at the time of setting and calculating the next material, the model calculation value is corrected by the time series learning coefficient and the separate group learning coefficient.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a learning control method in process lines such as rolling lines and cooling lines.

0002

[Prior Art] Generally, in process lines such as rolling lines and cooling lines, models for computer control are constructed based on theoretical or empirical formulas, but they do not completely represent the phenomena in the process line. Since it is almost impossible to construct a model, a method of controlling the model by modifying it based on data processed in the process line up to the previous time, that is, a learning control method is used.

For example, typical examples include models for predicting the rolling load in a rolling line and for predicting the amount of temperature drop in a cooling line.

[0004] As described above, the deviation between the actual value measured on the process line and the predicted value of the model occurs because the phenomenon on the process line is not completely represented in the model.

[0005] The reasons for this are: (1) the phenomenon itself has not yet been elucidated; (2) the phenomenon has been elucidated but the representation of the model has become too complex for online calculations, so it has been simplified; ■Incomplete collection of data required for model calculations, etc.

[0006]

[Problems to be Solved by the Invention] By the way, the discrepancies between the above-mentioned model and phenomena can be broadly classified into chronological fluctuations in the process line due to changes in equipment or environment, and changes in the type of material processed on the process line. There are also group-specific variations due to differences in processing patterns.

[0007] However, in conventional learning control in process lines, these two types of fluctuations are not properly separated and taken in, so there has been a problem that satisfactory control results cannot be obtained.

[0008] The present invention aims to solve these problems, and it is an object of the present invention to appropriately correct the discrepancy between a phenomenon and a model when controlling a process line, so that processing on the process line can be performed with high control accuracy. The purpose of this study is to provide a learning control method based on the following methods.

[0009]

[Means for Solving the Problems] In order to achieve the above object, the learning control method in a process line of the present invention includes the following steps:
In a system that controls a process line based on the model while modifying a setting calculation model based on the previous actual values, the type, size, control target value, processing pattern, etc. of the material to be processed in the process line, etc. For each group, create a table that stores the discrepancy between the actual value and the model calculation value calculated by the model as a learning coefficient. ■Difference F between the model calculation value Ycal and the actual material value Yact
Calculate, ■ From the deviation F, calculate the value Fg that is obtained by removing the smoothed value Fto of the time-series deviation between the actual value up to the previous material and the model calculation value, and ■ Perform group-based learning for the group to which the current material belongs. The table value Fgo of the coefficient is smoothed by the value Fg and stored in the table as a new table value, and the table value Fgo of the learning coefficient for each group of the group to which the current material belongs is removed from the deviation F. 2. Calculate the value Ft of
■When calculating the setting of the next material, the model calculation value calculated from the model is corrected based on the group learning coefficient and the new time-series learning coefficient of the group to which the next material belongs, stored in the table. It is a feature.

0010

[Operation] In the process line learning control method of the present invention described above, when performing learning control to control the process line while modifying the setting calculation model based on the previous actual values, it is necessary to The deviation F between the model calculation value Ycal calculated using the model and the actual value Yact of the current material is calculated, and from this deviation F, the time-series deviation between the actual value up to the previous material and the model calculation value is smoothed. A value Fg is calculated by removing the value Fto.

[0011] This value Fg causes a discrepancy between the model and the phenomenon due to variations due to group differences in the type, size, control target value, processing pattern, etc. of the material processed in the process line. Smooth the table value Fgo of the learning coefficient for each group of the group to which it belongs,
The results are stored as new group learning coefficients.

On the other hand, from the deviation F, a value Ft is calculated by removing the table value Fgo of the group learning coefficient for the group to which the current material belongs. This value Ft is a deviation due to time-series fluctuations, and the smoothed value Fto is smoothed by this value Ft, and the result is stored in the table as a new time-series learning coefficient.

[0013] Then, when calculating the settings for the next material, the predicted value (model calculation value) calculated from the model is corrected by the group-specific learning coefficient and the time-series learning coefficient of the group to which the next material belongs, and the phenomenon and the model are The discrepancy will be corrected appropriately.

[0014]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, a case in which the present invention is applied to rolling load prediction in a hot strip mill finish rolling line (process line) will be explained with reference to the drawings. In the hot strip mill finish rolling line, in order to control the plate thickness, the rolling load at each stand is predicted, and the rolling position of each stand is set in consideration of the mill elongation according to the predicted rolling load.

FIG. 2 is a flowchart for explaining the overall flow of plate thickness control in a hot strip mill finishing rolling line to which the method of the present invention is applied. Although a detailed explanation will be omitted since this is a simple process, basically there is a flow of steps B1 to B4 for performing setting calculations, and a flow of learning calculations of steps C1 to C4 that are performed in parallel.

In the setting calculation, first, after reading the command information for the next rolled material (step B1), the plate thickness at the exit side of each stand is determined (step B2), and the rolling conditions for each stand are determined based on the rolling conditions corresponding to this outlet side plate thickness. The rolling load is calculated (Step B3), and the roll gap (rolling position) of each stand is set using a gauge meter formula (Step B4).

On the other hand, in the learning calculation, after reading the actual data of the currently rolled material (step C1), the actual data is used to calculate the rolling load using the same model as the setting calculation (model for setting calculation) (step C2). ), the deviation between the actual rolling load and the rolling load based on the model calculation value is learned (step C3), and the learning result is reflected in the calculation of the rolling load of the next rolled material in step B3.

Next, a learning control method as an embodiment of the present invention will be explained with reference to FIG. Here, FIG. 1 shows the procedure of steps C3 and B3 in FIG. 2 in detail.

FIG. 1 is a flowchart for explaining in detail a learning control method as an embodiment of the present invention. In this embodiment, 200 kinds of materials are rolled by steel type and 16 kinds by finished plate thickness. It was divided into two groups. Further, the rolling load model (setting calculation model) is a known one, and is expressed by the following formula (1).

Pcal=σ・B・ld・Qp
(
1) Here, σ is the deformation resistance of the rolled material, B is the plate width, ld is the contact arc length, and Qp is the rolling force function.

[0021] Using the actual data of the rolled material, (1)
Calculated rolling load Pcal (= Yca
l) (Step A1), and based on the actual rolling load Pact (=Yact) of the currently rolled material detected by the load cell attached to each stand, for each stand,
According to the flow in Figure 1, new group learning coefficients FgN and time series learning coefficients FtN are created using equations (2) to (6) below.
is calculated, and the table values and time-series learning coefficients of group learning coefficients for the group to which the currently rolled material belongs are rewritten to the new values.

That is, in the control according to this embodiment,
When modifying the model for setting calculations based on past actual values, the type of rolled material (material) processed on the line,
A table is created in advance to store the discrepancies between actual values and values calculated from the model for each group such as size. Here, we will discuss the case where the deviation is expressed in the form of a ratio.

First, the ratio F of the calculated rolling load (model calculation value) Pcal calculated by the formula (1) using the actual operation data of the rolled material this time and the actual rolling load (actual value) Pact is calculated using the following formula. The deviation is calculated as shown in (2) (step A2).

F=Pact/Pcal(=Yact/Y
(2) This ratio F includes the discrepancy between the model and the phenomenon caused by both time-series fluctuations in the rolling line and fluctuations due to differences in the type of rolled material processed on the rolling line. has been done.

Next, as shown in equation (3) below, a value Fg is calculated by dividing by a smoothed value Fto of the time-series deviation between the actual value up to the previous rolled material and the model calculation value (Step A3
).

Fg=F/Fto

(3) Here, as will be described later, the smoothed value Fto represents the discrepancy between the model and the phenomenon due to time-series fluctuations in the rolling line. This results in a discrepancy between the model and the phenomenon due to variations due to group differences such as the type of wood.

According to the following equation (4), the table value Fgo of the group learning coefficient for the group to which the currently rolled material belongs is smoothed using this deviation Fg, and is stored in the table as a new table value (step A4).

FgN=αFg+(1−α)Fgo
(
4) In this formula (4), α is a smoothing coefficient and satisfies 0<α<1, table value Fgo is the previous learning coefficient, and FgN is a new group-based learning coefficient stored in the table.

Furthermore, as shown in equation (5) below, the deviation F
is divided by the table value Fgo of the learning coefficient for each group of the group to which the currently rolled material belongs (step A5).

Ft=F/Fgo

(5) Here, the value Ft is the difference between the model of the rolled material and the phenomenon after removing the deviation caused by group differences such as the type of rolled material processed on the rolling line,
In other words, it represents the deviation due to time-series fluctuations.

Then, as shown in the equation (6) below, the smoothed value Fto of the time-series deviation between the actual value up to the previous rolled material and the model calculation value is smoothed by the deviation Ft according to the equation (5). , is stored as a new time series learning coefficient (step A6
).

FtN=βFt+(1−β)Fto
(
6) In this formula (6), β is a smoothing coefficient and 0<β<1, Fto is a previous learning coefficient, and FtN is a new time-series learning coefficient.

When calculating the settings for the next rolling material, the rolling load Pcal of each stand is calculated using equation (1) from the rolling command information of the next rolling material, and the new group learning coefficient FgN and time series learning coefficient FtN are calculated. Using
Predicted rolling load (rolling load for setting calculation) P using the following formula (7)
set is calculated (step A7).

Pset=FtN・FgN・Pcal

(7) Here, Pset is the model calculation value Pcal modified by a new group learning coefficient FgN and a time series learning coefficient FtN.

Based on the predicted rolling load (rolling load for setting calculation) Pset calculated as described above, the rolling position S at each stand is determined in step B4. This rolling down position S can be determined from the well-known following equation (8).

[0036] S=h−Pset/M+So
(
8) Here, S is the rolling position, h is the outlet side plate thickness of each stand, M is the mill constant, and So is the offset due to roll wear, roll thermal expansion, etc.

In this embodiment, when using the group learning coefficient FgN in equation (7), the representative thickness of the finished product thickness group to which the next rolled material belongs and the adjacent finished product thickness group are used. The group learning coefficient FgN for both groups was interpolated using the following equation (9) according to the thickness of the finished product of the next rolled material.

[0038]

[Math 1]

Here, Hf is the thickness of the finished product (H
fj≦Hf<Hfj+1), Hfj is the representative thickness of the product thickness group to which the next rolled material belongs, Hfj+1 is the representative thickness of the group adjacent to the product thickness group to which the next rolled material belongs, FgNj is the representative thickness of the group to which the next rolled material belongs. The group learning coefficient FgNj+1 for the finished product thickness group is the group learning coefficient for the group adjacent to the finished product thickness group to which the next rolled material belongs.

In addition, in the case of Hf<Hfj, Hfj-1, FgN of the j-1 group instead of the j+1 group
Interpolation was performed using equation (9) using j-1.

FIG. 3 is a graph showing a comparison between the calculated value and the actual value of the rolling load using the learning control according to the present embodiment. This FIG. 3 shows an example of stand No. 6;
It is clear that the calculated values are in good agreement with the actual values.

FIG. 4 is a graph showing the thickness accuracy of the tip of the rolled material at the exit side of the final stand in this example, and the actual thickness of the tip of the rolled material was ±50μ with respect to the target thickness.
It can be seen that the rolling load learning method according to the present example is excellent.

As described above, according to the learning control method in the process line of this embodiment, the rolling load in the rolling line can be estimated with high accuracy, and the thickness control of the rolled material can be significantly improved.

[0044] In the above embodiment, the method of learning and controlling the rolling load in the rolling line was explained, but the method of the present invention includes learning the pass schedule plate thickness and temperature drop amount at each stand in the rolling line, and controlling the rolling load in the cooling line. The present invention is applied to various types of learning control in process lines, such as learning the amount of temperature drop in the process, and the same effects as those of the above embodiments can be obtained.

Furthermore, in the above embodiment, the case where the difference between the actual value and the model calculation value is learned in the form of a ratio was explained, but depending on the target quantity, the difference between the actual value and the model calculation value can be learned in the form of a difference. In this case as well, learning control can be performed in exactly the same manner as in the above embodiment.

[0046]

[Effects of the Invention] As detailed above, according to the process line learning control method of the present invention, the discrepancy between the observed model and phenomenon can be separated into time-series fluctuations and fluctuations due to group differences. By rewriting the learning coefficient and reflecting it in the setting calculation, the discrepancy between the phenomenon and the model is appropriately corrected when controlling the process line, and the process line processing can be performed with extremely high control accuracy. There is.

[Brief explanation of the drawing]

FIG. 1 is a flowchart for explaining a learning control method in a process line as an embodiment of the present invention.

FIG. 2 is a flowchart for explaining the overall flow of plate thickness control in the hot strip mill finishing rolling line of this embodiment to which the method of the present invention is applied.

FIG. 3 is a graph showing a comparison between calculated rolling load values and actual values according to the present example.

FIG. 4 is a graph showing the plate thickness accuracy of the tip of the rolled material at the exit side of the final stand in this example.

Claims (1)

[Claims] Claim 1. A learning control method for controlling a process line based on the model while modifying a setting calculation model based on previous performance values, wherein the type, size, and control of the material processed in the process line are provided. For each group of target values, processing patterns, etc., a table is created that stores the discrepancy between the actual value and the model calculation value calculated by the model as a learning coefficient. The model calculation value Yca by
1, calculate the deviation F between the model calculation value Ycal and the actual value Yact of the current material, and from the deviation F, calculate the smoothed value of the time-series deviation between the actual value up to the previous material and the model calculation value. A value Fg is calculated by removing Fto, and the table value Fgo of the group learning coefficient of the group to which the material belongs this time is smoothed by the value Fg and stored in the table as a new table value. From, the table value F of the learning coefficient by group for the group to which the material belongs this time
A value Ft from which go is removed is calculated, and the smoothed value Fto is smoothed by the value Ft and stored as a new time-series learning coefficient, and the next time the material setting calculation is performed, the model calculation calculated from the model is used. A learning control method in a process line, characterized in that a value is corrected based on a group-by-group learning coefficient of a group to which the next material belongs and a new time-series learning coefficient stored in the table.