JPH1031505A - Learning control method for process line - Google Patents

Abstract

PROBLEM TO BE SOLVED: To execute precise learning control by separating a learning coefficient into learning coefficients by individual layers and a time sequential learning coefficient so as to learn them. SOLUTION: The calculation value YCAL of a control model is calculated by using result data of this material and the instantaneous value C of the learning coefficient is obtained with a ratio with a result value YACT. The instantaneous value C is divided by the learning coefficient by individual layers CSo(i) of the individual layers (i), to which this material that is read from a table 1 belongs so as to obtain the instantaneous value CT(i) of the time sequential learning deficient. It is smoothed by the time sequential learning coefficient CTo(i) which is read from a table 2, the new time sequential learning coefficient CTn(i) and CTn(i-1) and CTn(i+1) are calculated and are stored in the table 2. The instantaneous value C of the learning coefficient is divided by a new time sequential learning coefficient CTn(i) and the instantaneous value CS(i) of the learning coefficient by individual layers is obtained. It is smoothed by the learning coefficient by individual layers CSo(i) and a new learning coefficient by individual layers CSn(i) is calculates so as to store it in the table 1.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a process line learning control method for calculating a control amount based on a control model in a process line such as a hot rolling line or a cold rolling line.

[0002]

2. Description of the Related Art FIG. 5 is a flowchart of a learning control method generally used in a conventional hot rolling rolling process and cooling process.

The calculated value Y _CAL of the control model is obtained using the actual value of the material this time, and the actual value Y _ACT of the control result such as, for example, the input side thickness, the output side thickness, and the gap between the rolling rolls of the horizontal rolling mill is obtained. The instantaneous value C of the learning coefficient that absorbs the prediction error of the control model is obtained from the equation (1) from the ratio (step ST5-1). C = Y _ACT / Y _CAL (1)

[0004] The stratified learning coefficient CSo (i) of the stratum i to which the present material belongs, which stores the instantaneous value C of the learning coefficient as a table value, is smoothed by a learning smoothing coefficient β, and a new learning of the stratified i is performed. The coefficient CSn (i) is calculated from equation (2) to update the table (step ST5-2). CSn (i) = β · C + (1−β) · CSo (i) (2)

[0005] In the control of the next material of the i-th layer, the learning coefficient CSn (i) updated in the above procedure is read from the table, and the model predictive value Y ' _CAL is corrected by using the learning coefficient CSn (i). Is determined by the following equation (3) (step ST5-3). Ypre = CSn (i) · Y ′ _CAL (3) where Y ′ _CAL : f (α1, α2 ·· αm, x′1,
x'2... x'm) f: function α1 to αm: parameter x'1 to x'm: explanatory variable

The learning control of the control model described above is performed for each stratum, and the model prediction error in each stratum is absorbed to bring the control result of the next material closer to the control target value. Prior arts related to the above-mentioned conventional methods include, for example, JP-A-4-368901 and JP-A-6-223811.
There is an official gazette.

[0007]

Since the conventional process line learning control method is configured as described above, there are the following first to third problems.

[0008] The first problem is that when processing from the current material to the next material for a certain layer is not continuous, the learning coefficient obtained from the actual value of the current material is not an appropriate value in controlling the next material. Because there is no limitation, the control accuracy of the next material may be significantly reduced. This is because the process line changes over time, such as deterioration of the components of the process line, changes in characteristics and environment, etc. between the current and next materials. This is because the learning coefficient for each stratum is collectively updated without separating the error of the actual value into the error of the control model itself for each stratum and the error due to the aging of the process line.

As means for solving this problem, for example, Japanese Patent Laid-Open Publication No. Hei 4-369901 discloses that a learning coefficient is divided into a layer-by-layer learning coefficient and a time-series learning coefficient common to all layers regardless of the layer. Then, a means for determining a learning coefficient is proposed.

However, in the above-mentioned publication, the time series learning coefficient is learned for every stratum on the assumption that the learning coefficient can be completely separated into the stratified learning coefficient and the time series learning coefficient. Actually, the two learning coefficients cannot be completely separated, and it is necessary to consider the fluctuation of the learning coefficient due to the fluctuation of the operation state.

This variation can be basically regarded as a variation of the time series learning coefficient. However, when time series learning is performed for all layers as in the above method, the learning coefficient becomes unstable. Therefore, it is necessary to perform time-series learning only between adjacent strata.

The second problem is that the updating of each learning coefficient does not consider the instantaneous value of the learning coefficient of the current material and the magnitude of the change amount of the learning coefficient of the previous material with respect to the learning smoothing coefficient. That is, the follow-up ability and the quick effect of learning are not guaranteed.

[0013] The third problem is that the learning coefficient is unconditionally updated in spite of the fact that the learning coefficient used for the material this time is appropriate and the control result sufficiently achieves the control target value. The control accuracy of the material may be reduced. This is because in the collection of the actual value used for calculating the learning coefficient of the material this time, a collection timing error, a sensor error, and the like may occur, and appropriate actual data may not be obtained.

The present invention has been made in order to solve the above-mentioned problem. Even when processing of the current material and the next material in the same layer is not continuous, the next material is accurately controlled, and the accuracy is sufficiently high. After being secured in the next time, the next time, one after another,
It is an object of the present invention to obtain a learning control method for a process line that keeps good accuracy successively after the material is successively turned.

[0015]

According to a first aspect of the present invention, there is provided a learning control method for a process line, wherein a learning coefficient is calculated based on a ratio between a calculated value of a control model calculated using an actual value of a present material and the actual value. Is obtained, and the instantaneous value of the learning coefficient is divided by the layer-by-layer learning coefficient of each layer to which the present material belongs to obtain the instantaneous value of the time-series learning coefficient. Smoothing with the coefficient, calculate a new time series learning coefficient and store it in the table, and at the same time, update the time series learning coefficient in the same manner for each layer adjacent to the layer to which the material belongs, and further, The instantaneous value is divided by the new time-series learning coefficient to obtain an instantaneous value of the stratified learning coefficient, the instantaneous value of the stratified learning coefficient is smoothed by the stratified learning coefficient, a new stratified learning coefficient is calculated, and It is something to memorize.

According to a second aspect of the present invention, a learning control method for a process line calculates a deviation between an instantaneous value of a learning coefficient for each layer and a stored read value, and uses the deviation in the learning coefficient for each layer according to the amount of the deviation. The learning smoothing coefficient is determined.

According to a third aspect of the present invention, a process line learning control method calculates a deviation between an instantaneous value of a stratified time series learning coefficient and a stored read value, and calculates a time series learning coefficient according to the deviation amount. Is to determine a learning smoothing coefficient used in the update calculation of.

According to a fourth aspect of the present invention, in the process line learning control method, a threshold is provided for an error between a control target value and a control actual value in updating a learning coefficient for each layer, and the learning coefficient is determined based on the threshold. The learning coefficient is updated only when the instantaneous value of is large.

[0019]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. FIG. 1 is a flowchart for explaining a process line learning control method according to Embodiment 1 of the present invention. In FIG. 1, first, a calculated value Y _CAL of a control model is calculated by using equation (4) using actual data of the current material. _{calculated, Y CAL = f (α1,} α2 ·· αm, x1, x2 ·· xm) by (4) the instantaneous value C of the learning coefficient by the ratio between the actual value Y _ACT by the measurement equation (5) It is obtained (step ST1-1). C = Y _ACT / Y _CAL (5)

Next, the instantaneous value C of the time-series learning coefficient is divided by the layer-by-layer learning coefficient CSo (i) of the layer i to which the current material belongs, which is read from the table 1 and divided by the instantaneous value C of the learning coefficient.
(I) is obtained by equation (6) (step ST1-
2). CT (i) = C / CSo (i) (6)

Thereafter, the instantaneous value CT of the time series learning coefficient is calculated.
(I) Time-series learning coefficient CT read from Table 2
o (i) is smoothed by equation (7), and CTn (i) = αCT + (1−α) · CTo (i) (7) α: Learning smoothing coefficient A new time series learning coefficient CTn (i) It is calculated and stored in Table 2. At the same time, the time series learning coefficients CTn (i-1) and CTn (i + 1) are similarly calculated for the stratum i-1 and i + 1 adjacent to the stratum to which the material belongs and stored in the table 2. (Step ST1-3).

Next, the instantaneous value C (i) of the learning coefficient for each stratum is obtained by equation (8) by dividing the instantaneous value C of the learning coefficient by the new time series learning coefficient CTn (i) (step ST1-
4). CS (i) = C / CTn (i) (8) Layer-by-layer learning of layer-by-layer i, which is the instantaneous value CS (i) of the layer-by-layer learning coefficient, which is a readout value from table 1 of the layer-by-layer learning coefficient. Smoothing with coefficient CSo (i), using learning smoothing coefficient β,
A new stratified learning coefficient CSn (i) is calculated by equation (9) and stored in Table 1 (step ST1-5). CSn (i) = β · CS (i) + (1−β) · CSo (i) (9) where the stratified learning coefficient updates only stratified i of the material,
The learning for the adjacent stratified i and i + 1 is not performed.

[0023] In the rolling next member, reads the sequence learning coefficient and stratification learning coefficient when updated by said from Table 1 and Table 2, to correct the prediction error by multiplying the model predicted value Y _{'CA L} using this Next material prediction Y _PRE
Ask for.

As described above, according to the first embodiment, the learning coefficient is separated into the stratified learning coefficient and the time-series learning coefficient. By performing learning limited to stratified i-1 and i + 1, a stable time-series learning coefficient can be obtained, and a stratified learning coefficient is also stabilized, so that more accurate learning control can be performed.

Embodiment 2 FIG. 2 is a flowchart for explaining a process line learning control method according to the second embodiment of the present invention. In the first embodiment in FIG. 1, a calculation step of calculating a learning smoothing coefficient of a stratified learning coefficient (step ST2- 6) is added. Steps ST2-1 to ST2-5 are the same as operation steps ST1-1 to ST1-5 of the first embodiment shown in FIG.

In the above calculation step (step ST2-6), the instantaneous value CS (i) of the learning coefficient for each stratum and the table 1
Learning coefficient CS for each stratum i, which is a stored read value from
The deviation ΔCS from o (i) is calculated by equation (10), and ΔCS = | CSo (i) −CS (i) | (10) The deviation ΔCS is large in accordance with the deviation ΔCS. At this time, a large learning smoothing coefficient β is read from the table 3 when it is small. Thereby, the responsiveness of the learning control of the stratified learning coefficient can be improved.

Embodiment 3 FIG. 3 is a flowchart for explaining a process line learning control method according to the third embodiment of the present invention. In the second embodiment in FIG. 2, a calculation step of calculating a learning smoothing coefficient of a time-series learning coefficient (step ST3- 7) is added. Steps ST3-1 to ST3-6 are the same as operation steps ST2-1 to ST2-6 of the second embodiment shown in FIG.

In the above calculation step (step ST3-7), the deviation ΔCT between the instantaneous value CT of the time series learning coefficient and the time series learning coefficient CTo which is a stored and read value from the table 2 is calculated by the equation (11). ΔCT = | CTo−CT | (11) According to the deviation amount ΔCT, a large learning smoothing coefficient α is obtained from the table 4 when the deviation amount ΔCT is large, and a small learning smoothing coefficient α is obtained when the deviation amount ΔCT is small. read out. Thereby, the responsiveness of the learning control of the time series learning coefficient can be improved.

Embodiment 4 FIG. 4 is a flowchart for explaining a process line learning control method according to the fourth embodiment of the present invention. In the third embodiment shown in FIG. 3, when the learning coefficient is updated, the difference between the control target value and the actual control value is large. A judgment step (step ST4-8) for judging whether or not learning is possible according to the judgment and a step (step ST4-9) for not updating the learning coefficient based on the judgment result are added. Steps ST4-1 to ST4-7 correspond to the operation steps ST3-1 to ST3-1 of the third embodiment shown in FIG.
Same as ST3-7.

First, the calculated value _YCAL of the control model is calculated using the actual data of the current material, and the instantaneous value C of the learning coefficient is obtained from the ratio with the actual value _YACT . Here, it is determined whether or not the control result is within a threshold value of an error from the control target value (step ST4-8).
Neither the stratified learning coefficient is updated (step ST4-9).
If NO, that is, if it is equal to or greater than the threshold value, each learning coefficient is calculated according to the procedure of the first embodiment, the second embodiment or the third embodiment. Thus, it is possible to prevent the prediction accuracy from being lowered due to erroneous learning of the learning coefficient.

[0031]

As described above, according to the first aspect of the present invention, a learning coefficient for correcting an error of the control model itself for each stratum, a learning coefficient for each stratum, and a time-dependent change of a process line. Is divided into time-series learning coefficients for correcting the error of the control model caused by the time-series learning coefficient. Since the learning of the stratification of the material is performed after removing the influence of the coefficient, even if the processing of the current material and the next material in the same stratification are not continuous,
Physical phenomena in the next material processing can be predicted with high accuracy, and a high-precision control result can be achieved. Further, after the control accuracy is sufficiently ensured, there is an effect that the control accuracy can be maintained satisfactorily continuously with the next material, the second-time material, and the successive-time material.

According to the second aspect of the present invention, the deviation between the instantaneous value of the learning coefficient for each stratum and the readout value stored in the table is calculated. When the deviation is large, a large learning smoothing is performed. Since the small learning smoothing coefficient is read from the table when the coefficient is small, the responsiveness of the learning control of the stratified learning coefficient can be improved.

According to the third aspect of the present invention, the deviation between the instantaneous value of the time series learning coefficient and the stored read value from the table is calculated, and when the deviation is large, a large learning smoothing is performed. When the coefficient is small, a small learning smoothing coefficient is read from the table, so that the responsiveness of the learning control of the time series learning coefficient can be improved.

According to the fourth aspect of the present invention, whether or not learning is to be performed is determined in accordance with the magnitude of the error between the control target value and the actual control value. There is an effect that can be prevented.

[Brief description of the drawings]

FIG. 1 is a flowchart illustrating a learning control method according to a first embodiment of the present invention.

FIG. 2 is a flowchart illustrating a learning control method according to a second embodiment of the present invention.

FIG. 3 is a flowchart illustrating a learning control method according to a third embodiment of the present invention.

FIG. 4 is a flowchart illustrating a learning control method according to a fourth embodiment of the present invention.

FIG. 5 is a flowchart illustrating a conventional general learning control method.

[Explanation of symbols]

Y _CAL control model calculated value, Y _ACT actual value, C learning coefficient instantaneous value, CSo (i) stratum i stratified learning coefficient, CT (i) time series learning coefficient instantaneous value, CTo
(I) Time series learning coefficient, CTn (i) New time series learning coefficient, CS (i) Instantaneous value of stratified learning coefficient.

Claims (4)

[Claims] 1. A calculation value of a control model is calculated by using an actual value of a current material, and an instantaneous value of a learning coefficient is obtained from a ratio of the calculated value of the control model to the actual value. The instantaneous value of the time-series learning coefficient is obtained by dividing by the layer-by-layer learning coefficient to which the layer belongs, and the instantaneous value of the time-series learning coefficient is smoothed by the time-series learning coefficient, and a new time-series learning coefficient is calculated and calculated. And at the same time, similarly updates the time-series learning coefficient for the stratum adjacent to the stratum to which the material belongs, and further divides the instantaneous value of the learning coefficient by the new time-series learning coefficient. A learning control method for a process line, wherein an instantaneous value of a learning coefficient is obtained, the instantaneous value of the learning coefficient for each layer is smoothed by the learning coefficient for each layer, and a new learning coefficient for each layer is calculated and stored in a table. 2. The method according to claim 1, wherein a deviation between an instantaneous value of the stratified learning coefficient and a stored read value is calculated, and a learning smoothing coefficient used in the stratified learning coefficient is determined according to the deviation amount. A learning control method for a process line according to claim 1. 3. The method according to claim 1, wherein a deviation between the instantaneous value of the time series learning coefficient and the stored read value is calculated, and a learning smoothing coefficient used in the update calculation of the time series learning coefficient is determined according to the deviation amount. 3. The learning control method for a process line according to claim 1, wherein 4. A method for updating a learning coefficient, wherein a threshold is provided for an error between a control target value and a control actual value, and the learning coefficient is updated only when the instantaneous value of the learning coefficient is larger than the threshold. The learning control method for a process line according to any one of claims 1 to 3, wherein: