JP2009116759A - Method and device for learning control model in process line, and production method of steel plate - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a control model learning method and device in a process line, which solve issues by combined use of a learning system high in following property and a learning system high in a noise-proof property, and to provide a production method of a steel plate. <P>SOLUTION: This is the control model learning method in the process line for learning a control model in a learning control method which performs control by correcting the control model based on a track record value in the process line. Respective outputs of the learning system high in following property and the learning system high in noise-proof property are compounded by attaching weights based on an error between the track record value and a control model output value and a new learning value to the control model is computed. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a learning control method and apparatus in process lines such as a rolling line and a cooling line, and a steel plate manufacturing method.

  In a process line such as a rolling line or a cooling line, a model for computer control is generally constructed based on a theoretical formula or an experimental formula. However, since it is almost impossible to build a model that completely represents the phenomenon in the process line, there is a method for performing control by correcting the model based on data processed in the process line until the last time, that is, a learning control method. Has been done. For example, the prediction model of the rolling load in the rolling line and the prediction model of the temperature drop amount in the cooling line are typical examples.

  As described above, the error (deviation) between the actual value measured in the process line and the control model output value occurs because the phenomenon in the process line is not completely expressed in the control model.

  The reason why this is not completely expressed is that (1) the phenomenon itself is unclear, and (2) the phenomenon is elucidated, but the expression of the control model is too complex to be in time for online calculations. (3) Incomplete collection of data necessary for control model calculation.

  In order to deal with the above problem, for example, a first-order lag filter as described below is used for calculating a learning value.

y _k = (1-α) * y _k-1 + α * u _k-1 (1)
y _k : Next learning value
y _k-1 : Learning value used this time
u _k-1 ： Error (deviation) between actual value and control model output value
α: When the coefficient α is close to “1”, an error between the actual value and the control model output value is reflected in the learning value early (high followability), but if noise exists, it causes unnecessary setting fluctuations. Stability is reduced (noise resistance is low). On the other hand, when α is close to “0”, noise resistance is improved, but there is a problem that followability with respect to errors is lowered.

  In order to deal with the above problem, for example, in the technique disclosed in Patent Document 1, the mismatch between the model and the phenomenon is represented by (a) time-series fluctuation of the process line due to a change in equipment or environment, and (b) process. There has been proposed a method for separately dealing with two types of fluctuations, that is, group-specific fluctuations caused by differences in the types of materials processed in a line and processing patterns. That is, a learning method has been proposed in which a time series learning coefficient is obtained in consideration of a change in the learning coefficient for each group, and the mutual characteristics are utilized.

In the technique disclosed in Patent Document 2, the division and structure of the stratified table are divided into a plurality of parts so as to be independent of the behavior of the control model error, and learning is performed with priorities. A learning method that absorbs the remaining part of the control model error at an early stage has been proposed.
JP-A-4-367901 JP 2003-340508 A

  However, in the technique disclosed in Patent Document 1 described above, the learning speed of group-specific learning, the stability of learning, and the measures for obtaining a uniform effect of learning are not described, and are accumulated in actual operation. It is impossible to prevent deterioration of the control model prediction accuracy based on learning bias and non-uniformity.

  Further, although the technique disclosed in Patent Document 2 described above needs to extract factors that are independent from each other with respect to the control model error, the cause of the error between the actual value and the control model predicted value is as described above. In addition, (1) the phenomenon itself has not been elucidated, or (3) there is a fundamental problem that it is difficult to extract factors in the first place when data necessary for calculation of the control model is incomplete.

  The present invention has been made in view of the above circumstances, and a control model learning method and apparatus in a process line, and a steel plate, which solves the problem by using a learning method with high followability and a learning method with high noise resistance. It aims at providing the manufacturing method of.

  The invention according to claim 1 of the present invention is a control model learning method in a process line, in which a control model is learned in a learning control method in which control is performed by correcting the control model based on actual values in the process line. Based on the error between the actual value and the control model output value, each output of the learning method with high followability and the learning method with high noise resistance is weighted and synthesized, and a new learning value for the control model is obtained. It is a control model learning method in a process line characterized by calculating.

  According to a second aspect of the present invention, in the control model learning method in the process line according to the first aspect, the learning method with high followability is realized by a delay filter. This is a model learning method.

According to a third aspect of the present invention, in the control model learning method in the process line according to the first or second aspect, the highly noise-resistant learning method uses the error at a plurality of past time points. A control model learning method in a process line characterized by being realized by a regression model.

  The invention according to claim 4 of the present invention is the process for learning a control model in a process line according to any one of claims 1 to 3, wherein the type / size of the material to be processed in the process line, the operating conditions, And a control model learning method in a process line, characterized in that processing is performed for each group such as a processing pattern.

Further, the invention according to claim 5 of the present invention is a control model learning apparatus in a process line, in which a control model is learned in a learning control apparatus that performs control by correcting the control model based on the actual value in the process line. A control model calculator that calculates a control model output value using a control model based on the actual value, a comparator that calculates an error between the actual value and the control model output value, and a delay filter for the error A high-response follow-up learner that calculates output 1; a high-noise-learning learner that performs smoothing processing on the error and calculates output 2; and a sudden change weight for determining whether or not the error has suddenly changed. An error behavior monitor to be calculated; and a combiner that combines output 1 and output 2 based on the sudden change weight to calculate a new learning value for the control model. A control model learning device in Sesurain.
Furthermore, the invention according to claim 6 of the present invention is a method for manufacturing a steel sheet, wherein the steel sheet is processed using the control model learning method in the process line according to any one of claims 1 to 4. Is the method.

  In the present invention, the learning value is calculated according to the error behavior between the actual value and the control model output value by using the learning method with high followability and the learning method with high noise resistance. And high noise resistance can be achieved at the same time. As a result, the reliability of the setting has increased, the monitoring load on the operator has been reduced, and stable operation and high-quality product manufacturing have become possible.

  Hereinafter, the present invention will be specifically described with reference to the drawings. FIG. 1 is a diagram showing a configuration example of a control model learning apparatus in a process line according to the present invention. In the figure, 1 is a process, 10 is a control device, 20 is a control model calculator, 30 is a comparator, 40 is a high follow-up learner, 50 is a high noise tolerance learner, 60 is an error behavior monitor, and 70 is a composition. And 80 represent databases.

  An example of learning control for each group caused by a difference in type and processing pattern performed by a normal learning control method will be described below. In addition, about the difference | error of a performance value and a control model calculation value, although both can also be used, it demonstrates as what uses both ratio. The overall process flow is first outlined, and then each process is described in detail.

When the material is manufactured, the comparator 30 calculates an error E (i) between the actual value y _act (i) output from the process 1 and the calculated value y _cal (i) output from the control model calculator 20. Then, the calculated error E (i) is input to the high response follow-up learning device 40, the high noise immunity learner 50, and the error behavior monitoring device 60, and based on the output from the error behavior monitoring device 60, the high response follow-up is performed. By combining the output z _rp (i) from the sex learning device 40 and the output z _nz (i) from the high noise resistance learning device 50 in the synthesizer 70, a learning value L (i) for the next material is calculated. And stored in the database 80.

  When the production of the next material is determined, the next material learning value L (i) is extracted from the database 80 and is input to the control device 10, and the control device 10 controls the control command value based on the next material learning value L (i). Is calculated and sent to the process 1 and reflected in the next material setting.

The control model calculator 20 calculates the calculated value y _cal (i) using the control model based on the performance information (factor 1, factor 2,...) From the process as shown in the following equation. .

y _cal (i) = f (factor 1, factor 2,...) (1)
f: Control model The comparator 30 calculates an error E (i) between the actual value y _act (i) and the calculated value y _cal (i) as follows.

E (i) ＝ y _act (i) / y _cal (i) (2)
The high follow-up learning device 40 calculates the output z _rp (i) using, for example, a first-order lag filter. A high-order lag filter may be used.

z _rp (i) = (1-α) * z _rp (i-1) + α * E (i-1) (3)
α: Time constant parameter z _rp (i-1): Previous output in the same segment The high noise proof learner 50 performs smoothing processing. FIG. 2 is a diagram illustrating an example of the smoothing process. Regress the past n-point errors {E (i-1), E (i-2), ..., E (in)} with a linear expression, and output z _nz (i) as shown in the following expression: Calculate

z _nz (i) = p1 * (i) + p2 (4)
p1: Coefficient when regressed with linear equation
p2: Coefficient (intercept) when regressing with linear equation
A high-order equation may be used as the regression equation. Further, as the high noise proof learner 50, a first order lag filter with a small time constant parameter may be adopted. The regression method has an effect of improving the phase lag compared to the first-order lag filter method.

  The error behavior monitor 60 calculates a sudden change weight wk in order to determine whether or not the error has changed suddenly. FIG. 3 is a diagram illustrating an example of calculating the sudden change weight. First, as shown in the figure, an error change amount {E (i) −E (i−1)} is calculated, and “+ change” (= 1) and “no change” (= 0) and “-change” (= 1). Next, the sudden change portion can be determined by correlating with the pattern shown in FIG.

wk = {m (1) * "No change"
+ M (2) * ("+ change" + "-change")
+ M (3) * “No change”} / 3 (5)
In the above example, description has been made using two points (i) and (i-1). Similarly, a slope is calculated using three or more points, and a pattern corresponding to FIG. In the case of three points, it may be checked with FIG. 4 (b)). Also, a plurality of discriminating means such as two points, three points, etc. may be prepared and combined for use in discrimination.

The synthesizer 70 synthesizes the output z _rp (i) of the high follower learning device 40 and the output z _nz (i) of the high noise proof learner 50 based on the weight determination, and obtains the next material learning value L (i). Calculate as shown in the following equation. However, in synthesizing two outputs, in the high noise resistance learner, if there is a sudden change point in the error of the past n points, the accuracy of the next material learning value is lowered. Until then, it is preferable to take measures such as adopting a high follow-up learning output.

L (i) = wk * z _rp (i) + (1-wk) * z _nz (i) (6)
The calculated next material learning value L (i) is stored in the group classification to which the material belongs in the database 80. For the group classification, classification such as the type of material, size, operation conditions (heating temperature, etc.), processing pattern (depending on the passing process), etc. may be appropriately selected.

  The control device 10 calculates the group classification of the material, extracts the learning value L (i) stored in the database, multiplies the learning value by the control model calculation value, and applies it to the subsequent processing.

  As an Example, the example which applied this invention to the rolling load prediction in the hot rolling of a low carbon material is shown. FIG. 5 is a diagram showing a final stand (7 stands) of hot rolling to which the present invention is applied. The figure shows how the rolling load actual value at 7 stands is taken into the control model learning device, the rolling load learning value is calculated and passed to the control device.

  And FIG. 6 is a figure which shows an example of the result of the Example to which this invention is applied. FIG. 6 (a) shows a case where only a first-order lag filter for which high followability is obtained, whereas FIG. 6 (b) shows a learning method with high followability and noise resistance according to the present invention. This is a case where a high learning method is used in combination, and each shows an error between the actual value and the control model output value and the learning value.

  In FIG. 6 (a), although the followability with respect to the change around the 40th is good, the learning value is frequently changed for other rolled materials because the error follows too much. On the other hand, in FIG. 6B, at the time of sudden change, the high followability of the first-order lag filter is selected to ensure followability, and the noise resistance filter ratio < In the figure (right axis) is 1.0), it can be seen that the fluctuation of the learning value can be suppressed by selecting the high noise resistance filter.

It is a figure which shows the example of 1 structure of the control model learning apparatus in the process line which concerns on this invention. It is a figure which shows an example of a smoothing process. It is a figure which shows the example of calculation of the sudden change weight. It is a figure which shows the example of the pattern which takes a correlation. It is a figure which shows the last stand (7 stands) of the hot rolling to which this invention is applied. It is a figure which shows an example of the result of the Example to which this invention is applied.

Explanation of symbols

1 Process 10 Controller 20 Control Model Calculator 30 Comparator 40 High Followability Learner 50 High Noise Resistance Learner 60 Error Behavior Monitor 70 Synthesizer 80 Database

Claims (6)

A control model learning method in a process line, in which a control model is learned in a learning control method in which a control model is modified and controlled based on actual values in the process line,
Based on the error between the actual value and the control model output value, weighting and synthesizing the outputs of the learning method with high tracking performance and the learning method with high noise resistance to calculate a new learning value for the control model A control model learning method in a process line characterized by: The control model learning method in the process line according to claim 1,
The learning method with high followability
A control model learning method in a process line, which is realized by a delay filter. In the control model learning method in the process line according to claim 1 or 2,
The learning method with high noise resistance,
A control model learning method in a process line, which is realized by a regression model using the error at a plurality of past time points. In the control model learning method in the process line according to any one of claims 1 to 3,
A method for learning a control model in a process line, characterized in that processing is performed for each group such as the type / size of the material processed in the process line, operating conditions, and processing patterns. A control model learning device in a process line that learns a control model in a learning control device that performs control by correcting the control model based on the actual value in the process line,
A control model calculator for calculating a control model output value using a control model based on the actual value;
A comparator for obtaining an error between the actual value and the control model output value;
A high-response follow-up learner that calculates an output 1 using a delay filter for the error;
A high noise proof learner that smoothes the error and calculates output 2;
An error behavior monitor for calculating a sudden change weight for determining whether or not the error is suddenly changed;
A synthesizer that synthesizes the output 1 and the output 2 based on the sudden change weight and calculates a new learning value for the control model;
A control model learning device in a process line, comprising: A method for manufacturing a steel sheet, comprising: processing a steel sheet using the control model learning method in the process line according to any one of claims 1 to 4.

Images