KR101451516B1 - Learning control apparatus of rolling process - Google Patents

Abstract

The present invention improves the precision of the predictions in the set calculation of the rolling process. The learning control device 5 includes an instantaneous value calculation means 10, a forward / backward point comparison means 13, and an update gain calculation means 17. [ The instantaneous value calculating means 10 calculates the learning coefficient using the actual data at the target point of the rolled material 1 and the actual data at the auxiliary target point set near the target point. The forward / backward comparison means 13 compares the learning coefficient of the target point calculated by the instantaneous value calculation means 10 with the learning coefficient of the auxiliary target point to determine the presence or absence of noise included in the learning coefficient of the target point. The update gain calculating means 17 calculates the update gain of the learning coefficients based on the presence or absence of noise determined by the forward / backward point comparing means 13. [

Description

[0001] LEARNING CONTROL APPARATUS OF ROLLING PROCESS [0002]

The present invention relates to a learning control device used in control of a rolling process.

In the control of the rolling process, each facility is driven so that the rolled material at the completion of the production has a desired dimension and a desired temperature. The control of the rolling process generally consists of setting control and dynamic control. The setting control is performed particularly at the beginning of rolling. Further, when rolling is started and sufficient performance data is obtained, the transition from the setting control to the dynamic control is performed.

In the setting control, the rolling phenomenon is predicted by a model equation. That is, in the setting control, based on the predicted value obtained by the model equation, the set value of the actuator of the rolling mill etc. is determined. Examples of the set values to be determined include, for example, the rolling speed, the quantity of cooling water, and the roll gap of the rolling mill.

The physical phenomena occurring in the rolling process can not be expressed completely by the model equation. In addition, there is also a case where the model expression is simplified by giving priority to ease of adjustment and reduction in calculation load. Therefore, a deviation occurs between the predicted value obtained by the model equation and the actually obtained value (actual value).

Conventionally, a learning term is provided in the model expression, and the learning coefficient to be reflected in the learning term is updated based on the performance data. In such a learning control, it is important to follow the value (used value) of the learning coefficient to be used with high precision and high response. However, it is difficult to accurately grasp what factors are caused by errors in the model equation. Incidentally, the actual data itself used in the learning control also includes an error. For this reason, the obtained learning coefficient (instantaneous value) can not be used as it is as a used value of the learning coefficient.

Therefore, in general, the learning coefficient (instantaneous value) is reflected in the updated value through the smoothing filter. An example is shown below.

Znew = Zuse 占 (1 -? ₁ ) + Zcur 占? ₁

here,

Znew: learning coefficient (renewal value)

Zuse: learning coefficient (previous time value)

Zcur: learning coefficient (instantaneous value)

α ₁ : Update coefficient of learning coefficient (eg, time constant of filter)

to be.

The update gain (alpha ₁ ) of the learning coefficient is generally stored as an adjustable integer. The balance of the tolerance against noise (or inner / outer resistance) in the learning control and the follow-up performance is adjusted by the update gain α ₁ . That is, if the update gain α ₁ is set to a value close to 0, the noise resistance improves, but the followability deteriorates. If the update gain (alpha ₁ ) is set to a value close to 1, the noise resistance is lowered but the followability is improved.

On the other hand, Patent Document 1 proposes an apparatus for simultaneously improving noise resistance and followability.

Japanese Patent Application Laid-Open No. 2009-116759

In the apparatus described in Patent Document 1, a renewal value of a learning coefficient is calculated by using a smoothing filter in order to improve followability of learning control.

Zrp (i) = (1 -? ₂ ) x Zrp (i-1) +? ₂ x E (i-

here,

α ₂ : Time constant parameter

Zrp (i-1): the previous output in the same section

E (i-1): error

to be.

Further, in order to improve the noise resistance of the learning control, the tendency of the learning coefficient is determined by regression calculation based on the learning coefficients of a plurality of previously manufactured materials, and is made an update value of the learning coefficient.

Znz (i) = p1 x (i) + p2

here,

p1: Coefficient when a plurality of past learning coefficients (instantaneous values) are returned in a linear equation

p2: coefficient (intercept) when a plurality of past learning coefficients (instantaneous values)

to be.

Further, the behavior of the learning coefficient in the past is monitored, and as the weight coefficient wk, priority is given to noise resistance and follow-up in response to the behavior. The update value of the final learning coefficient is obtained by the following equation.

L (i) = wk x Zrp (i) + (1-wk) x Znz (i)

In the technique described in Patent Document 1, it is judged whether or not noise is included in the learning coefficient of the currently rolling ash based on the learning coefficients of a plurality of ashes manufactured previously. Therefore, when noise is already included in the learning coefficient of the past, noise resistance can not be improved.

In addition, since the time constant parameter? ₂ is constant, followability is constant. There is a problem that sufficient follow-up can not be obtained when the error of the control result is large and the learning coefficient is greatly changed.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems, and an object of the present invention is to provide a learning control of a rolling process capable of improving at least the accuracy of predictions in the setting calculation, Device.

Another object of the present invention is to provide a learning control apparatus for a rolling process capable of improving accuracy of predicted values in a setting calculation by improving both noise tolerance and follow-up performance of learning control.

A learning control apparatus for a rolling process according to the present invention is a learning control apparatus for updating a learning coefficient used for improving the accuracy of a set value in a setting calculation of a rolling process. The learning control apparatus includes performance data at a target point of a rolled material, The learning coefficient of the auxiliary point and the learning coefficient of the target point calculated by the first calculation means are compared with each other and included in the learning coefficient of the target point And second calculation means for calculating the update gain of the learning coefficient based on the presence or absence of the noise determined by the first comparison means.

Further, the learning control device of the rolling process according to the present invention further comprises second comparing means for comparing the learning coefficient of the target point of the rolled material and the corresponding learning coefficient of the past, and the second calculating means includes: The learning coefficient update gain is set based on the comparison result of the second comparison means and the learning coefficient of the target point calculated by the first calculation means is reflected in the updated value.

With the learning control apparatus according to the present invention, the precision of the predicted value in the setting calculation of the rolling process can be improved.

BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 is a diagram showing a configuration of a learning control apparatus for a rolling process according to Embodiment 1 of the present invention. Fig.
2 is a diagram showing a configuration of a learning control apparatus of a rolling process according to Embodiment 2 of the present invention.

The present invention will be described in detail with reference to the accompanying drawings. In the drawings, the same or equivalent parts are denoted by the same reference numerals. The redundant description is appropriately simplified or omitted.

Embodiment Mode 1.

1 is a diagram showing a configuration of a learning control apparatus for a rolling process according to Embodiment 1 of the present invention.

In Fig. 1, reference numeral 1 denotes a rolled material flowing through the rolling line. In the rolling line, the rolled material 1 is rolled by a rolling mill to a desired dimension and a desired temperature. As described above, in the control of the rolling process, the setting control and the dynamic control are performed.

The setting calculation means (2) performs the setting calculation of the rolling process. That is, in the setting control, the setting calculation means 2 uses the predetermined model formula to determine the set value of the actuator of the rolling facility or the like. The setting values determined by the setting calculation means 2 include, for example, a rolling speed, a quantity of cooling water, and a roll gap of the rolling mill. A deviation occurs between the set value (predicted value) obtained from the model equation and the value (actual value) obtained when actually rolling is performed. Therefore, in the setting calculation, the learning coefficient is used to improve the precision of the setting value. Thereby, the prediction accuracy of the rolling phenomenon can be improved and the operation can be stabilized.

The reading means 3 reads the learning coefficient (used value) from the storage means 4. In the storage means 4, a learning coefficient (used value) for each type of rolled material 1 is stored. The reading means 3 reads the appropriate learning coefficient from the storage means 4 in accordance with the type of the rolled material 1 (hereinafter also referred to as " current material " The setting calculation means 2 calculates the set value by reflecting the learning coefficient read by the reading means 3 to the learning term of the model expression.

The learning control device 5 is a device for updating learning coefficients stored in the storage means 4. [ The learning control device 5 is provided with the performance data collecting means 6, the performance data editing means 7, the storage means 8, the performance recalculation means 9, the instantaneous value computing means 10, the normal determination means 11, the storage means 12, the forward and backward comparison means 13, the storage means 14, the time series comparison means 15, the storage means 16, , And an update value calculation means (18).

The performance data collection means (6) collects the performance data at the target point and auxiliary target point of the rolled material (1). The performance data includes process data during rolling and data (result) after completion of rolling.

In the rolled material 1, a target point and an auxiliary target point for collecting the performance data are set in advance. The target point is set, for example, at the tip of the rolled material 1. [ The position of the target point may be set at the middle portion or the tail end portion of the rolled material 1. A target point may be set at a plurality of positions of the rolled material 1. [ However, considering that the setting control is performed at the beginning of rolling, it is appropriate to set the target point on the leading end of the rolled material 1. [

The auxiliary target point is set in the vicinity of the target point. Fig. 1 shows an example in which two auxiliary target points are respectively set on the front and rear of the target point (the front end side and the front end side of the rolled material 1). The position of the auxiliary target point is not limited to the above. The auxiliary target point may be set only on the front side (or only on the rear side) of the target point. The auxiliary target point may be set one before the target point. However, considering that updating of the learning coefficients is performed optimally, it is appropriate to set a plurality of auxiliary target points before and after the target point, respectively.

The performance data collection means 6 may acquire the performance data of any one point of the rolled material 1 and use the value as the performance data of the target point. The performance data collection means 6 acquires performance data of a plurality of points included in a certain range of the rolled material 1 and acquires a value obtained from the plurality of performance data (for example, an average value) Data may be used.

The same is true for the performance data of the auxiliary target point.

The performance data editing means (7) edits the performance data acquired by the performance data collection means (6). For example, the performance data editing unit 7 removes the measurement noise from the performance data acquired by the performance data collection unit 6. The performance data editing means 7 removes the outliers (abnormal values) from the performance data acquired by the performance data collection means 6. An abnormality determination threshold value is stored in the storage means (8). The performance data editing means 7 compares the performance data acquired by the performance data collection means 6 with the abnormality determination threshold stored in the storage means 8 to determine whether the performance data corresponds to an outliers . The performance data editing means 7 performs the same editing of the performance data collected at the target point and the performance data collected at each auxiliary target point.

The function of the performance data editing means 7 may be added in response to necessity.

The performance recalculation (recalculation) means 9 calculates an actual recalculation value. The performance re-calculation means 9 uses the performance data edited by the performance data editing means 7 and the model formula used in the setting calculation means 2 when calculating the re-calculated values.

The setting calculation means 2 substitutes the predicted value into the model equation and obtains a set value (predicted value) for use in the setting control. On the other hand, the performance recalculation means 9 substitutes the performance data into the model expression to obtain the predicted value. A value obtained by substituting the performance data into the model expression (predicted value) is called the performance recalculation value. The performance re-calculation means 9 calculates the re-calculated value of the target point based on the performance data collected at the target point. In addition, the performance re-calculation means 9 calculates the re-calculated value of each auxiliary target point based on the performance data collected at each auxiliary target point.

The instantaneous value calculating means 10 (first calculating means) calculates a learning coefficient (instantaneous value). The instantaneous value calculating means 10 uses the actual data and the actual recalculated value when calculating the learning coefficient (instantaneous value).

For example, the instantaneous value calculating means 10 first obtains the reevaluation value of the target point obtained by the calculation from the reevaluating means 9. The instantaneous value calculating means 10 calculates the actual data corresponding to the actual recalculated values calculated by the actual recalculating means 9 among the actual data collected by the actual data collecting means 6 at the target point, (6). The instantaneous value calculating means 10 compares the obtained recalculated value of the target point with the actual data of the target point and calculates the learning coefficient (instantaneous value) Zmain of the target point. The learning coefficient is displayed using, for example, the ratio of the two values. The difference between the two values may be used to represent the learning coefficient.

The instantaneous value calculating means 10 calculates the learning coefficient (instantaneous value) Zsub (i) by using the performance data and the reevaluated value for each auxiliary target point. i represents the number of the current auxiliary target point. In the present embodiment, since four auxiliary target points are set, i = 1, 2, 3, and 4 are obtained.

The normal determination means 11 determines whether or not the learning coefficient (instantaneous value) Zmain of the target point is a normal value. Upper and lower limit values of learning coefficients are stored in the memory means 12. The normal determination means 11 compares the learning coefficient (instantaneous value) Zmain calculated by the instantaneous value calculation means 10 with the upper and lower limit values stored in the storage means 12 to make the determination.

For example, when the learning coefficient (instantaneous value) Zmain deviates from the range defined by the upper and lower limit values, the normal determination means 11 determines that the learning coefficient (instantaneous value) Zmain is abnormal. The normal decision means 11 replaces the learning coefficient (instantaneous value) Zmain with the upper limit value when the learning coefficient (instantaneous value) Zmain exceeds the upper limit value. The normal decision means 11 replaces the learning coefficient (instantaneous value) Zmain with the lower limit value when the learning coefficient (instantaneous value) Zmain is below the lower limit value.

On the other hand, when the learning coefficient (instantaneous value) Zmain is included in the range defined by the upper and lower limits, the normal determining means 11 determines that the learning coefficient (instantaneous value) Zmain is normal. In this case, the normal determination means 11 does not rewrite the learning coefficient (instantaneous value) Zmain. The normal determination means 11 outputs the learning coefficient (instantaneous value) Zmain input from the instantaneous value calculation means 10 as it is.

The function of the normal decision means 11 may be added in response to the necessity.

The forward / backward comparison means 13 (first comparison means) compares the learning coefficient (instantaneous value) Zmain of the target point calculated by the instantaneous value calculation means 10 with the learning coefficient (instantaneous value) Zsub ).

For example, the forward / backward comparison means 13 first calculates an average value of the learning coefficients (instantaneous values) Zsub (i = 1 to 4) of the auxiliary target points. Next, the forward / backward comparison means 13 calculates the ratio between the learning coefficient (instantaneous value) Zmain of the target point and the average value. At this time, the forward / backward comparison means 13 may calculate the difference instead of the ratio. Upper and lower limit values for performing comparison determination are stored in the memory means 14. The forward / backward comparison means 13 compares the calculated ratio (or difference) with the upper and lower limit values stored in the storage means 14 to determine whether or not noise included in the learning coefficient (instantaneous value) Zmain exists .

For example, when the calculated ratio deviates from the range defined by the upper and lower limit values, the forward / backward comparison means 13 includes noises in the actual data used for calculating the learning coefficient (instantaneous value) Zmain . That is, the forward / backward comparison means 13 determines that the learning coefficient (instantaneous value) Zmain is abnormal.

On the other hand, when the calculated ratio falls within a range defined by the upper and lower limits, the forward / backward comparison means 13 includes noises in the performance data used for calculating the learning coefficient (instantaneous value) Zmain . That is, the forward / backward comparison means 13 determines that the learning coefficient (instantaneous value) Zmain is normal.

The forward / backward comparison means 13, when judging that there is no noise, digitizes the judgment result as? And outputs the result. ? is numerically expressed as follows, for example.

When the learning coefficient (instantaneous value) (Zmain) does not include noise:? = 1.0

When the learning coefficient (instantaneous value) (Zmain) contains noise: γ = 0.0

The time series comparison means 15 (second comparison means) compares the learning coefficient (instantaneous value) Zmain of the present target point and the rolled material (hereinafter also referred to as " past material " (Instantaneous value) (Zmain_old (j)) of the same point (target point) in the target point The learning coefficient (instantaneous value) (Zmain_old (j)) of the past is stored in the storage means 16. j represents the number of the past reassembly. For example, Zmain_old (1) is the learning coefficient (instantaneous value) of the target point of the past which is the same as the current material and rolled immediately before the current material.

For example, the time series comparison means 15 calculates a learning coefficient (instantaneous value) (Zmain_old (j = 1 to m)) of a target point of a plurality of objects rolled when the current material is the closest to the current material, From the storage means (16). The time series comparing means 15 calculates an average value of learning coefficients (instantaneous values) (Zmain_old (j = 1 to m)) of the acquired target points. Next, the time series comparing means 15 calculates the amount of change of the learning coefficient (instantaneous value) Zmain of the target point of the current reed relative to the average value. For example, the time series comparing means 15 calculates the ratio of the learning coefficient (instantaneous value) Zmain of the current target point to the average value. The time series comparing means 15 may calculate the difference instead of the ratio.

The update gain calculation means 17 (second calculation means) calculates the update gain of the learning coefficient. Specifically, the update gain calculating means 17 calculates the smoothing coefficient?. The update gain calculation means 17 calculates the update gain of the smoothing coefficient a based on the presence or absence of noise determined by the forward / backward point comparison means 13 (i.e., the value of?) And the comparison result of the time series comparison means 15 Calculation is performed. For example, the update gain calculation means 17 calculates the smoothing coefficient? By the following equation.

alpha ₀ is a reference value of the smoothing coefficient derived by the conventional method. As α ₀ , for example, a fixed value is adopted. Further,? = 0.0 to 1.0.

The update value calculation means 18 calculates the learning coefficient (updated value) of the target point. For example, the update value calculation means 18 uses the smoothing coefficient? Calculated by the update gain calculation means 17 and performs the calculation by the following equation.

Znew = Zold x (1 -?) + Zcur x?

here,

Znew: learning coefficient (renewal value)

Zold: Learning factor (used value)

Zcur: learning coefficient (instantaneous value)

to be. The update value calculation means 18 acquires a learning coefficient (use value) Zold from the storage means 4. [ The update value calculation means 18 acquires the learning coefficient (instantaneous value) Zcur from the normal determination means 11. When the learning coefficient (updated value) Znew is calculated by the above equation, the updated value calculation means 18 stores the obtained value in the storage means 4 and updates the stored content.

With the learning control device 5 having the above configuration, the accuracy of the predicted value in the setting calculation of the rolling process can be improved. Therefore, the quality of the product and the stability of the operation can be improved.

For example, when there is no noise by the forward / backward comparison means 13, the output? Of the forward / backward point comparison means 13 becomes zero. The smoothing coefficient alpha becomes zero, and the update gain calculation means 17 sets the update gain to zero. The learning coefficient (instantaneous value) Zmain calculated by the instantaneous value calculating means 10 is not reflected in the updated value of the learning coefficient. Therefore, learning control excellent in noise resistance can be provided.

On the other hand, if no noise is determined by the forward / backward comparison means 13, the update gain calculation means 17 sets the update gain on the basis of the comparison result of the time series comparison means 15. Specifically, when the learning coefficient changes significantly compared to the past, the value of? Increases and the learning coefficient is updated by weighting on the instantaneous value. That is, as the change amount calculated by the time series comparison means 15 becomes larger, the update gain calculation means 17 calculates the update gain (the instantaneous value) Zmain calculated by the instantaneous value calculation means 10 to the updated value The update gain is set so as to be largely reflected. Therefore, it is possible to provide learning control excellent in followability.

Embodiment 2:

2 is a diagram showing a configuration of a learning control apparatus of a rolling process according to Embodiment 2 of the present invention. The learning control device 5 in the present embodiment is different from the learning control device 5 shown in Fig. 1 in that it does not include the time-series comparison means 15 and the storage means 16. Fig. Other configurations of the learning control device 5 and configurations other than the learning control device 5 are the same as those shown in Fig.

In the learning control device 5 shown in Fig. 2, when the presence / absence of noise is determined by the forward / backward comparison means 13, the same control as that of the learning control device 5 shown in Fig. 1 is performed.

On the other hand, when no noise is determined by the forward / backward comparison means 13, the update gain calculation means 17 sets the update gain to a constant. For this reason, the learning coefficient (instantaneous value) Zmain calculated by the instantaneous value calculating means 10 is reflected to the updated value at a constant rate.

In the learning control device 5 of the present embodiment, it is not necessary to preserve the learning coefficient of the past. Therefore, the storage area in the learning control device 5 can be greatly reduced. The configuration of the learning control device 5 can be simplified and can be easily implemented.

1: rolled material
2: Setting calculation means
3: Reading means
4, 8, 12, 14, 16: storage means
5: Learning control device
6: Performance data collection means
7: Performance data editing means
9: Performance Recalculation Means
10: instantaneous value calculation means
11: Normal judgment means
13: forward / backward comparison means
15: time series comparison means
17: Update gain calculation means
18: update value calculation means

Claims (5)

A learning control device for updating a learning coefficient used for improving a precision of a set value in a setting calculation of a rolling process,
First calculation means for calculating a learning coefficient using actual data at a target point of the rolled material and actual data at an auxiliary target point set near the target point;
First comparison means for comparing the learning coefficient of the target point calculated by the first calculation means and the learning coefficient of the auxiliary target point to determine the presence or absence of noise contained in the learning coefficient of the target point;
And second calculation means for calculating an update gain of the learning coefficient based on the presence or absence of noise determined by the first comparison means. The method according to claim 1,
The second calculation means sets the update gain of the learning coefficient to 0 when the first comparison means determines that there is noise, and updates the learning coefficient of the target point calculated by the first calculation means to the updated value Is not reflected in the rolling process. 3. The method according to claim 1 or 2,
The second calculation means sets the update gain of the learning coefficient to an integer when the noise is determined by the first comparison means and sets the learning coefficient of the target point calculated by the first calculation means to a constant ratio To the update value. 3. The method according to claim 1 or 2,
Further comprising second comparing means for comparing the learning coefficient of the target point of the rolled material and the corresponding learning coefficient of the past,
Wherein the second calculation means sets the update gain of the learning coefficient based on the comparison result of the second comparison means when no noise is determined by the first comparison means, And the learning coefficient of the target point is reflected in the updated value. 5. The method of claim 4,
The second comparison means calculates a variation amount of the learning coefficient of the target point of the rolled material with respect to the past learning coefficient,
Wherein the second calculation means calculates the learning coefficient of the target point calculated by the first calculation means as the variation amount calculated by the second comparison means increases when no noise is determined by the first comparison means Is set so as to be largely reflected in the renewal value.

Images