JP5759206B2 - Learning coefficient controller - Google Patents

Description

  The present invention relates to a process for rolling metal or the like, and more particularly to a learning coefficient control device for correcting parameters of a model expressing the rolling process.

  Normally, in process control, a model expressing a phenomenon of a process is created, and a control system is constructed for this model and control is performed. Therefore, the control accuracy of the final product is largely related to how accurately the model represents the process. Therefore, in order to fill in the error between the process and the model, the model is learned to improve the accuracy of the model.

で予測精度を向上させる学習方法がある。 In general, in order to improve the prediction accuracy of the rolling model, learning control is performed to fill in the error between the actual rolling value and the predicted rolling model value. There is a load model as a representative rolling model, and a learning coefficient as shown in Equation (1) as a control method of the load model.

There is a learning method that improves prediction accuracy.

ここで、

である。

here,

It is.

  The learning coefficient is expressed as a ratio between the rolling actual value and the rolling model predicted value, and is learned for each layer using a layered table divided by factors that are likely to cause a rolling model error. Specifically, as described in Patent Document 1, by stratifying the learning classification by steel type, rolling size and work roll, learning the error of the rolling model due to the difference between the rolling material and the rolling stand, The accuracy of the rolling model is improved.

  In addition, the learning coefficient may be an abnormal value due to an abnormality in the actual value. This does not mean that the model has resulted in an abnormal value, so such an abnormal value must not be learned. For this reason, a learning coefficient is generally provided with a limit value. When the learning coefficient exceeds the limit value, the learning coefficient is ignored or learned by being replaced with a predetermined value within the limit value. .

JP-A-10-180321

  As described above, the learning accuracy is stratified by learning and storing the learning coefficients in layers, but the learning coefficients for each layer have the following problems.

  If the parameters of the model formula are not appropriate, or if a preset limit value is not appropriate, the limit value may be exceeded even if the learning coefficient calculated for a certain stratification is a correct value. Also in these cases, the learning coefficient is ignored or learned by being replaced with a predetermined value within the limit value. For this reason, since the learning coefficient is not appropriate for the next rolled material, the prediction accuracy of the rolling model cannot be improved.

  The present invention has been made to solve the above problem, and even when the learning coefficient exceeds the limit value, the stratified learning coefficient is corrected, and the learning coefficient is controlled within the limit value, thereby efficiently. An object of the present invention is to provide a learning coefficient control device capable of correcting a model error.

The invention according to claim 1 is based on the difference between the predicted value of the rolling load and the actual value for the process model that finally predicts one rolling load based on the load model expressing the rolling process and the deformation resistance model. A learning calculation unit for calculating the learning coefficient of the model, a storage unit for storing the learning coefficient of each model in a stratified table prepared for each model, a load model and a deformation resistance model constituting the prediction calculation In order to obtain a correction value for changing the predicted value of the deformation resistance model so that the predicted value of the rolling load finally obtained when the learning coefficient of the load model is corrected is not changed , The predicted value of the load model and deformation resistance model calculated based on the actual values including tension and the predicted value of the deformation resistance model are changed from the results calculated based on the actual values including the plate thickness and tension. Conversion coefficient calculation unit that calculates the conversion coefficient so that the ratio of the predicted value of the load model and the deformation resistance model before and after the change is the same from the predicted value of the load model and the deformation resistance model after change obtained when A conversion coefficient storage unit that stores the conversion coefficient calculated by the conversion coefficient calculation unit, and when the learning coefficient of the load model stored in the storage unit exceeds a preset first limit value, A learning coefficient correction value calculation unit that calculates a correction value of a learning coefficient, a conversion coefficient average value calculation unit that calculates an average value of conversion coefficients stored in the conversion coefficient storage unit, and a learning coefficient correction value calculation unit Based on the corrected value of the calculated learning coefficient and the average value of the conversion coefficient calculated by the conversion coefficient average value calculation unit, the amount of change in the predicted value generated by correcting the learning coefficient of the load model is transformed into the deformation resistance model. Complement And a model correction value calculation unit that calculates a correction value of the model for compensation.

  According to the present invention, one model is calculated based on the correction value of the learning coefficient calculated by the learning coefficient correction value calculation unit and the average value of the conversion coefficient calculated by the conversion coefficient average value calculation unit. The correction value of the model is calculated to compensate for the amount of change in the predicted value caused by the correction of the learning coefficient of the other model, so even if the learning coefficient exceeds the preset limit value, stratification The model error can be efficiently corrected by correcting the learned learning coefficient and controlling the learning coefficient within the limit value, thereby improving the accuracy of the model.

It is the block diagram which applied the structure of the learning coefficient control apparatus of Example 1 to the cold tandem rolling mill. 3 is a detailed diagram of a conversion coefficient storage unit in the learning coefficient control device of Embodiment 1. FIG. It is a figure which shows the effect of the learning coefficient correction | amendment part in the learning coefficient control apparatus of Example 1. FIG. It is a figure which shows the effect of the learning coefficient correction | amendment part in the learning coefficient control apparatus of Example 1. FIG. FIG. 3 is a detailed diagram of a correction value storage unit in the learning coefficient control device according to the first embodiment. It is the figure which applied the structure of the learning coefficient control apparatus of Example 2 to the cold tandem rolling mill.

  An example in which the learning coefficient control apparatus according to the first embodiment of the present invention is applied to a rolling model of a cold tandem rolling mill will be described.

  FIG. 1 is a configuration diagram in which the configuration of the learning coefficient control device of the first embodiment is applied to a cold tandem rolling mill. In FIG. 1, the rolling mill is configured by arranging five machines in a tandem from a first stand 1 to a fifth stand 5. The rolled material 6 is rolled into a desired product in the direction of arrow 7 and wound up by a coiler (not shown). Each rolling mill is provided with a device for measuring a rolling load or the like (not shown). Further, a plate thickness detection device 8 for measuring the plate thickness of the rolled material 6 is provided on the exit side of the fifth stand 5.

  The set value of each actuator of the cold rolling mill is calculated so as to obtain a desired product from rolling information given in advance using a rolling model. For example, the rolling load can be predicted by expressing the load model as equation (2) and the deformation resistance model as equation (3).

ここで、

である。なお、式（２）には張力の影響関数が含まれているため、荷重と変形抵抗の間には非線形な関係がある。通常、圧延モデルは誤差をもっているため、次に述べる学習制御を行うことで、圧延モデルの精度向上が図られている。

here,

It is. In addition, since the influence function of tension is included in Equation (2), there is a non-linear relationship between the load and the deformation resistance. Usually, since a rolling model has an error, the accuracy of the rolling model is improved by performing learning control described below.

  The learning coefficient control device 9 includes a learning calculation unit 10, a storage unit 11, a conversion coefficient calculation unit 14, a conversion coefficient storage unit 15, a learning coefficient monitoring unit 16, a learning coefficient correction value calculation unit 17, a learning coefficient correction unit 18, and a conversion coefficient. An average calculation unit 19, a model correction value calculation unit 20, and a correction value storage unit 21 are provided.

  A performance data collection device (not shown) collects rolling performance values during rolling. The learning calculation unit 10 calculates a ratio or difference between the rolling record value and the rolling model predicted value as a learning coefficient using the rolling record value collected by the record data collecting device, and performs a smoothing process. The storage unit 11 stores the learning coefficient smoothed by the learning calculation unit 10 in a stratified learning coefficient table. In the next material setting calculation, the initial value of each actuator is calculated using the latest learning coefficient (learning value) stored in the storage unit 11.

  In the case of the rolling model described above, the storage unit 11 is provided with a load learning table 12 and a deformation resistance learning table 13 for two models (a load model and a deformation resistance model) constituting the prediction calculation. For example, the load learning table 12 is stratified into a steel type, a target plate thickness, and a rolling stand. The deformation resistance learning table 13 is stratified into steel types and rolling stands.

  Next, the characteristic configuration of the learning coefficient control apparatus according to the first embodiment will be described with reference to FIGS. In the learning coefficient control device of the first embodiment, the rolling model described above is expressed as follows.

ここで、

である。ここで、

は、本発明を実施するために新しく設けられたモデルの補正値である。

here,

It is. here,

Is a correction value of a model newly provided for carrying out the present invention.

は、荷重モデル（一方のモデル）の荷重学習係数を修正したときに、荷重予測値(物理量の予測値)が変わらないように変形抵抗モデル（他方のモデル）を変更するための補正値である。荷重学習係数を修正する場合、荷重予測値

は荷重学習係数の変更前の荷重予測値と同じでなければならない。このため、変形抵抗モデルを式（５）で表し、荷重学習係数の修正による影響をモデルの補正値

で補正する。以下に、この補正値の求め方と使用方法を説明する。

Is a correction value for changing the deformation resistance model (the other model) so that the load predicted value (the predicted value of the physical quantity) does not change when the load learning coefficient of the load model (one model) is corrected . When correcting the load learning coefficient, the predicted load value

Must be the same as the predicted load value before the change of the load learning coefficient. For this reason, the deformation resistance model is expressed by equation (5), and the effect of the modification of the load learning coefficient is corrected by

Correct with. Hereinafter, how to obtain the correction value and how to use it will be described.

  The conversion coefficient calculation unit 14 obtains a conversion coefficient α for replacement with a change amount of deformation resistance equivalent to the change amount with respect to the load based on Expression (4). As described above, since there is a non-linear relationship between the load and the deformation resistance, it is difficult to analytically determine the correction value of the model when the load learning coefficient is changed. For this reason, the model correction value is calculated through the conversion coefficient described below.

を計算し、変形抵抗計算値

を用いて荷重計算値

を計算する。その後、変換係数演算部１４は、変形抵抗計算値

を例えば１０%変更して変形抵抗変更値

とし、変形抵抗変更値

を用いて変形抵抗を変化させたときの荷重計算値

を計算する。そして、変換係数演算部１４は、荷重と変形抵抗とのそれぞれについて変更前と変更後の比を求め、両者の比が同じになるように変換係数

を導入し、式（６）の演算を行う。 The conversion coefficient calculation unit 14 calculates the deformation resistance value based on the actual rolling value.

Calculate the deformation resistance value

Calculate the load using

Calculate Thereafter, the conversion coefficient calculation unit 14 calculates the deformation resistance value.

For example, change the deformation resistance by changing 10%

Deformation resistance change value

Calculated load when deformation resistance is changed using

Calculate Then, the conversion coefficient calculation unit 14 obtains the ratio before and after the change for each of the load and the deformation resistance, and the conversion coefficient so that the ratio between the two becomes the same.

To calculate the equation (6).

ここで、

である。つまり式（６）の演算により、モデルの補正値

を演算するための変換係数

が求まる。

here,

It is. In other words, the correction value of the model is obtained by the calculation of Equation (6).

Conversion factor for computing

Is obtained.

  The conversion coefficient calculation unit 14 calculates the conversion coefficient of Expression (6), and then outputs the conversion coefficient to the conversion coefficient storage unit 15. By using the above calculation method, the conversion relationship between models can be easily calculated even when the relationship between related models is nonlinear or when the mathematical model is changed.

  Similarly to the load learning table 12, the conversion coefficient storage unit 15 converts the conversion coefficient calculated by the conversion coefficient calculation unit 14 into the conversion coefficient table 22 shown in FIG. 2 stratified by steel type, target plate thickness, and rolling stand. Remember. Thereby, the latest conversion coefficient for every division | segmentation division can be hold | maintained.

  The learning coefficient monitoring unit 16 sets a limit value in which the maximum value or the minimum value of the learning coefficient of the target thickness classification stratified by the steel type and rolling stand of the load learning table 12 is set in advance, for example, the upper and lower limit values of the learning coefficient. When it exceeds, an execution command is transmitted to the learning coefficient correction value calculation unit 17 and the conversion coefficient average value calculation unit 19.

  Thereby, the correction timing of the learning coefficient can be set according to the limit value. In the above description, the limit value to be monitored by the learning coefficient monitoring unit 16 is the upper and lower limit values of the learning coefficient, but the present invention is not limited to this. For example, if the upper and lower limits when the upper and lower limits of the learning coefficient are reduced by 20% are set as the limit values to be monitored, the execution command can be transmitted before the learning coefficient exceeds the upper and lower limits of the learning coefficient. The learning coefficient can be corrected early.

  The learning coefficient correction value calculating unit 17 uses the average value of the load learning coefficient of the target sheet thickness section stratified by the steel type and the rolling stand of the load learning table 12 stored in the storage unit 11 as a correction value (7) ).

ここで、

である。その後、学習係数修正値演算部１７は、学習係数修正部１８とモデル補正値演算部２０に学習係数の修正値を出力する。

here,

It is. Thereafter, the learning coefficient correction value calculation unit 17 outputs the correction value of the learning coefficient to the learning coefficient correction unit 18 and the model correction value calculation unit 20.

  Thereby, when the learning coefficient correction unit 18 described later corrects the load learning coefficient, the behavior of the entire learning coefficient of the target plate thickness categories 1 to N can be considered, and the load learning coefficient is within the effective learning range. By correcting to, stable model learning can be performed.

  In addition, the learning coefficient correction value calculation unit 17 has a maximum value and a minimum value of the learning coefficient in the target thickness class stratified by the steel type and the rolling stand of the load learning table 12 stored in the storage unit 11. The result of dividing the addition result by 2 can be calculated as a correction value. Thus, even when only one category of learning coefficient has a unique value, the amount of correction is large, and therefore, stable model learning can be performed by correcting the learning coefficient within the effective learning range.

  The learning coefficient correction unit 18 calculates the equation (8), that is, the learning coefficient correction value, with respect to the learning coefficient of the target sheet thickness classification corresponding to the steel type classification number and rolling stand classification number of the rolled material in the load learning table 12. Each learning coefficient of the target plate thickness classification is divided by the correction value calculated by the calculation unit 17 to update the load learning coefficient in the load learning table 12.

ここで、

である。

here,

It is.

である。荷重学習係数の修正前では目標板厚区分1の荷重学習係数が上限値を超えている。荷重学習係数を修正することにより、各目標板厚区分の荷重学習係数が全体的に下がり、目標板厚区分1の荷重学習係数が有効範囲内に修正されている。 FIG. 3 shows an example of correcting the load learning coefficient when the correction value is the average value of the target thickness group stratified by the steel type and rolling stand of the load learning table 12. The effective range of the learning coefficient is

It is. Before correction of the load learning coefficient, the load learning coefficient of the target thickness category 1 exceeds the upper limit. By correcting the load learning coefficient, the load learning coefficient of each target plate thickness section decreases as a whole, and the load learning coefficient of the target plate thickness section 1 is corrected within the effective range.

  Further, FIG. 4 shows the result of division in which the correction value is added to the maximum value and the minimum value of the learning coefficient of the target sheet thickness classification stratified by the steel type and rolling stand of the load learning table 12, and the addition result is divided by 2. An example of correcting the load learning coefficient in the case of The effective range of the learning coefficient is the same as described above. In this case, since the correction amount of the learning coefficient is large, the load learning coefficient of the target plate thickness category 1 is corrected within the effective range.

  When the conversion coefficient average value calculation unit 19 receives the execution instruction from the learning coefficient monitoring unit 16, the conversion coefficient average value calculation unit 19 is stratified into steel types and rolling stands stored in the conversion coefficient table 22 of the conversion coefficient storage unit 15 illustrated in FIG. The average value of the conversion coefficient of the target plate thickness category is calculated by equation (9).

ここで、

である。その後、変換係数平均値演算部１９は、式（９）で演算された変換係数の平均値をモデル補正値演算部２０に出力する。

here,

It is. Thereafter, the conversion coefficient average value calculation unit 19 outputs the average value of the conversion coefficients calculated by Expression (9) to the model correction value calculation unit 20.

  The model correction value calculation unit 20 is expressed by Expression (10) using the average value of the conversion coefficient calculated by the conversion coefficient average value calculation unit 19 and the learning coefficient correction value calculated by the learning coefficient correction value calculation unit 17. Calculate the correction value of the model.

ここで、

である。これにより、荷重学習係数の修正に伴う変形抵抗モデルの補正値を求めることができる。

here,

It is. Thereby, the correction value of the deformation resistance model accompanying the correction of the load learning coefficient can be obtained.

  The correction value storage unit 21 stores the correction value calculated by the model correction value calculation unit 20 in the correction value table 23 shown in FIG. 5 stratified into steel types and rolling stands similar to the deformation resistance learning table 13. Next time, when rolling under the same conditions is performed, the deformation resistance is calculated by Equation (5), and the influence of the predicted load value accompanying the correction of the load learning coefficient is corrected by the deformation resistance.

  This is nothing but to consider that the overall deviation of the predicted load value due to the target thickness classification is due to an error in the deformation resistance model. For this reason, by calculating the deformation resistance according to the equation (5), the prediction accuracy of the deformation resistance can be improved, and the accuracy of other rolling models using this can be improved.

  FIG. 6 is a diagram in which the configuration of the learning coefficient control device of the second embodiment is applied to a cold tandem rolling mill. The learning coefficient control apparatus according to the second embodiment is also applied to the same target as the learning coefficient control apparatus according to the first embodiment.

The learning coefficient control apparatus according to the second embodiment is different from the learning coefficient control apparatus according to the first embodiment in that a correction value monitoring unit 24 and an operator display screen 25 (corresponding to the display unit) are added. Only the configuration different from the learning coefficient control apparatus of the first embodiment will be described below.

  The correction value monitoring unit 24 monitors whether or not the correction value calculated by the model correction value calculation unit 20 exceeds a preset limit value. When the correction value exceeds the limit value, an alarm is displayed on the operator display screen 25. Output. Thereby, it is possible to notify the operator of the model adjustment time at an early stage and lead to quick model adjustment, so that the accuracy of the model can be improved.

  Further, the correction value monitoring unit 24 can output an alarm to the storage unit 11 and store it when the correction value calculated by the model correction value calculation unit 20 exceeds a preset limit value. Thereby, it is possible to inform the coordinator of the adjustment time of the model. That is, the model coordinator can confirm the correction history of the learning coefficient, which leads to quick model adjustment, so that the accuracy of the model can be improved.

  When the correction value of the model exceeds the limit value, it means that the prediction accuracy of the model is poor. Therefore, it is necessary to readjust the model parameter, and the correction value monitoring unit 24 can know the adjustment time of the model at an early stage.

  In addition, although the correction value of the deformation resistance model is described in a multiplicative manner, the present invention is not limited to this. Moreover, although the learning coefficient of each rolling model was described in a multiplying manner, it is not limited to this. From the above, stable and accurate model learning can be performed.

  In the first and second embodiments, the model is described as a load model and a deformation resistance model. However, the present invention can be applied to other models having different learning categories, such as a load model and a friction coefficient model, a torque model and a deformation resistance model, and the like. Can be implemented.

  Further, the present invention can be applied to all rolling mills such as a hot rolling mill and a single mill. Furthermore, in the present invention, the process is a rolling process, but the present invention is not limited to this, and any process that is controlled using a model such as a heating / cooling process can be applied.

  The present invention is applicable to all rolling mills such as a cold tandem rolling mill, a hot rolling mill and a single mill.

DESCRIPTION OF SYMBOLS 1-5 Rolling machine 6 Rolled material 7 Rolling direction 8 Plate thickness detection apparatus 9, 9a Learning coefficient control apparatus 10 Learning calculation part 11 Storage part 12 Load learning table 13 Deformation resistance learning table 14 Conversion coefficient calculation part 15 Conversion coefficient storage part 16 Learning coefficient monitoring unit 17 Learning coefficient correction value calculation unit 18 Learning coefficient correction unit 19 Conversion coefficient average value calculation unit 20 Model correction value calculation unit 21 Correction value storage unit 22 Conversion coefficient table 23 Correction value table 24 Correction value monitoring unit 25 Operator display screen

Claims (6)

Learning to calculate the learning coefficient of each model based on the difference between the predicted value of the rolling load and the actual value for the process model that finally predicts one rolling load based on the load model and deformation resistance model expressing the rolling process An arithmetic unit;
A storage unit for storing the learning coefficient of each model in a stratified table prepared for each model;
To change the predicted value of the deformation resistance model so that the predicted value of the rolling load that is finally obtained when the learning coefficient of the load model is modified for the load model and deformation resistance model that constitute the prediction calculation does not change in order to obtain a correction value, the thickness, and the predicted value of the load model and deformation resistance model is calculated based on actual values, including tension, the thickness of the predicted value of deformation resistance model, based on the actual績値including tension The conversion coefficient is calculated so that the ratio of the predicted value of the load model before and after the change and the predicted value of the deformation resistance model is the same from the predicted value of the changed load model and the deformation resistance model obtained when changing from the calculated result. A conversion coefficient calculator to
A conversion coefficient storage unit that stores the conversion coefficient calculated by the conversion coefficient calculation unit;
When the learning coefficient of the load model stored in the storage unit exceeds a preset first limit value, a learning coefficient correction value calculation unit that calculates a correction value of the learning coefficient of the model,
A conversion coefficient average value calculation unit for calculating an average value of the conversion coefficients stored in the conversion coefficient storage unit;
Prediction generated by correcting the learning coefficient of the load model based on the correction value of the learning coefficient calculated by the learning coefficient correction value calculation unit and the average value of the conversion coefficient calculated by the conversion coefficient average value calculation unit A model correction value calculation unit for calculating a correction value of the model for compensating the amount of change in the value with the deformation resistance model;
A learning coefficient control apparatus comprising: A learning coefficient monitoring unit that monitors whether the learning coefficient of the load model stored in the storage unit exceeds a preset first limit value;
A learning coefficient correction unit that corrects the learning coefficient of the model using the correction value of the learning coefficient calculated by the learning coefficient correction value calculation unit;
A correction value storage unit that stores the correction value of the model calculated by the model correction value calculation unit;
The learning coefficient control apparatus according to claim 1, further comprising:   The learning coefficient control device according to claim 1, wherein the learning coefficient correction value calculation unit uses an average value of learning coefficients for each layer as a learning coefficient correction value.   The learning coefficient correction value calculation unit adds a maximum value and a minimum value of learning coefficients for each layer, and divides the addition result by 2 to obtain a learning coefficient correction value. The learning coefficient control device according to claim 2.   A correction value monitoring unit that monitors the correction value calculated by the model correction value calculation unit and outputs an alarm to an operator display screen when the correction value exceeds a preset second limit value. The learning coefficient control device according to any one of claims 1 to 4.   6. The learning coefficient control according to claim 5, wherein the correction value monitoring unit further outputs and stores an alarm to the storage unit when the correction value exceeds the preset second limit value. apparatus.

Images