KR102440698B1 - Plant control device and controlling method thereof, rolling mill control device and controlling method and program thereof - Google Patents

Abstract

A plant control device and a control method thereof, a rolling mill control device and a control method thereof, which can be controlled, for example, an increase in the machine life of the rolling mill or an improvement in control precision by avoiding ineffective control; and program is provided.
The operation stage of the control target plant is controlled according to the output of the neural network learned by the control method learning apparatus for learning the combination of performance data and control operation of the control target plant according to the control effect, and forming a plurality of neural networks with different control effects. A plant control device comprising a control execution device that gives a control output to There is only the output of the neural network formed by controlling the operation stage of , and learning the case where the control effect is low, and the output of the neural network formed by learning the case where the control effect is low when there is room at the operation stage in the operation stage. control the operating stage of the plant to be controlled according to It is characterized in that it does not control.

Description

A plant control device and its control method, a rolling mill control device and its control method, and a program

The present invention relates to a plant control apparatus and its control method, a rolling mill control apparatus and its control method, and a program in real-time feedback control performed using artificial intelligence technology such as a neural network.

Conventionally, in various plants, in order to obtain a desired control result by the control, plant control based on various control theories is implemented.

As an example of a plant, for example, in rolling mill control, fuzzy control and neurofuzzy control have been applied as a control theory targeting shape control that controls the wave pattern state of a plate as an example of control. The fuzzy control is applied to shape control using a coolant, and neuropurge control is applied to shape control of the senzier rolling mill. Among these, shape control to which neurofuzzy control is applied, as disclosed in Patent Document 1, obtains the similarity ratio between the difference between the actual shape pattern detected by the shape detector and the target shape pattern, and a preset reference shape pattern, and the similarity ratio From , this is also performed by obtaining the control output amount for the operating stage according to the control rule expressed by the control operation stage operation amount with respect to the preset reference shape pattern. Hereinafter, as a prior art, the shape control of the senjimere rolling mill using neuro-fuzzy control will be used.

In FIG. 5, the shape control of the senjimere rolling mill described in FIG. 1 of patent document 1 is shown. Neuro-fuzzy control is used in shape control of the senjimere rolling mill. In this example, the pattern recognition mechanism 51 performs pattern recognition of a shape from the thread shape detected by the shape detector 52, and calculates which of the preset reference shape patterns the thread shape is closest to. The control arithmetic mechanism 53 performs control using the control rule comprised by the control operation stage operation amount with respect to the preset shape pattern as shown in FIG. 6 , in the pattern recognition mechanism 51 , the difference Δε between the shape performance detected by the shape detector 52 and the target shape εref is 1 to 8 shape patterns ε Which one of them is closest is calculated, and the control calculation mechanism 53 selects and executes any of the control methods 1 to 8 .

However, in the method of Patent Document 1, there are cases where the operator performs manual operation during rolling for verification of the control rule to perform verification of the control rule, etc., however, an unexpected shape change may be exhibited. That is, the case where the control rule determined as mentioned above does not conform to reality arises. This is caused by a lack of examination of mechanical properties or a change in the operating state or mechanical conditions of the rolling mill. Therefore, once a control rule is set, it is often left in that state as long as there are no defects.

When the control rule is no longer conforming to reality due to a change in operating conditions or the like, it becomes difficult to achieve control accuracy to a certain extent or higher because the control rule is fixed. Also, once the shape control is operated, the operator does not manually operate (it becomes a disturbance in the control), so it is difficult to find a new control rule by the manual intervention of the operator. Moreover, even when rolling a rolled material of a new standard, it is difficult to set a control rule according to the material.

As described above, in the conventional shape control, there is a problem in that it is difficult to correct the control rule because the control is performed using the preset control rule.

In order to solve this problem, as shown in Patent Document 2, by changing the control rule at random while performing shape control and learning the rule to improve the shape,

1) Discovering new control rules while performing shape control during rolling,

2) A new control rule is not predictable in advance, and in some cases a control rule that could not be predicted at all becomes optimal.

realizing that

Japanese Patent No. 2804161 Japanese Patent No. 4003733

In the prior art, a representative shape is previously set as a reference shape pattern, and control is performed based on a control rule indicating the relationship between the control operation stage operation amount with respect to the reference waveform pattern. The learning of the control rule is also related to the amount of operation of the control operation stage with respect to the reference waveform pattern, and the predetermined representative reference shape pattern is used as it is.

However, with respect to the same shape pattern, the amount of manipulation of the control operating end varies, and when the control method is learned according to the evaluation criteria that the shape deviation is reduced by the manipulation of the operating end, the manipulated amount of a plurality of control operating ends is learned, so the average value thereof It may become a negative control output, and the control effect may become small. Although it is also possible to select and learn only the operation method with a large control effect, in that case, the control output with a small control effect is not learned, and this may also reduce the control effect.

Further, in the prior art, the control operation stage operation amount with respect to the reference waveform pattern is determined, but the mechanical life of the control operation stage when the control operation is performed according to the determined control operation stage operation amount is not taken into consideration.

In view of the above, in the present invention, a plant control device and a control method thereof that can be controlled, for example, an increase in the machine life of a rolling mill or an improvement in control precision by avoiding ineffective control, etc., and a control method thereof; An object of the present invention is to provide an apparatus for controlling a rolling mill, a method for controlling the same, and a program.

In view of the above, in the present invention, "according to the output of the neural network learned by the control method learning apparatus which learns the combination of performance data and control operation of the control target plant according to the control effect, and forms a plurality of neural networks with different control effects, A plant control device comprising a control execution device that provides a control output for controlling an operating stage of the plant to be controlled, wherein the control execution device has an output of a neural network formed by learning a case in which the control effect is high, and the control effect is high. In this case, only the output of the neural network formed by controlling the operating stage of the control target plant according to the output and learning the case where the control effect is low exists, and the control effect is not effective when the position of the operating stage in the operating stage has room. The control stage of the control target plant is controlled according to the output of the neural network formed by learning the low case, and only the output of the neural network formed by learning the case where the control effect is low exists, and the position of the operating stage in the operation stage is free A plant control device characterized in that the operating stage of the plant to be controlled is not controlled when there is no control.”

In addition, in the present invention, "a combination of performance data and control operation of the control target plant is learned according to the control effect, and according to the output of the neural network learned by the learning unit which forms a plurality of neural networks with different control effects, A plant control method comprising: a control unit for providing a control output for controlling an operating stage There is only the output of the neural network formed by controlling the operation stage of , and learning the case where the control effect is low, and the output of the neural network formed by learning the case where the control effect is low when there is room at the operation stage in the operation stage. control the operating stage of the plant to be controlled according to A plant control method characterized by not controlling

Further, in the present invention, "a program for realizing a plant control device that performs control by recognizing a combination pattern of performance data of a control target plant with respect to a control target plant by a computer system, wherein the computer system is a control target plant A control method learning program that learns the combination of performance data and control operation of and, the control execution program controls the operating stage of the control target plant according to the output when the output of the neural network formed by learning the case where the control effect is high and the control effect is high, and when the control effect is low When there is only the output of the neural network formed by learning, and the control effect is low when there is room for the position of the operating stage in the operating stage, the operating stage of the control target plant is controlled according to the output of the neural network formed by learning, and When there is only the output of the neural network formed by learning the case where the control effect is low, and there is no leeway in the position of the operating stage at the operating stage, the operating stage of the plant to be controlled is not controlled.”

By using the present invention, by automatically changing the control rule in consideration of the operating end position and the control effect during control, the increase of the machine life of the rolling mill by preventing the operating end position from reaching the limit value in advance, or ineffective By avoiding control, effects such as improvement of control precision can be expected.

1 is a view showing an outline of a plant control apparatus according to an embodiment of the present invention;
2 is a view showing a specific configuration example of the control rule execution unit 10 according to an embodiment of the present invention.
3 is a view showing a specific configuration example of the control rule learning unit 11 according to an embodiment of the present invention.
Fig. 4 is a diagram showing the configuration of a neural network when the present invention is used for shape control of a senjimere rolling mill.
5 is a view showing shape control of the senjimere rolling mill described in FIG. 1 of Patent Document 1. FIG.
It is a figure which shows the control rule in shape control of the senjimere rolling mill described in FIG. 1 of patent document 1. FIG.
Fig. 7 is a diagram showing an outline of the input data creation unit 2;
Fig. 8 is a diagram showing the outline of the control output calculating section 3;
Fig. 9 is a diagram showing an outline of the control output determining unit 5;
Fig. 10 is a diagram showing a shape deviation and a control method;
Fig. 11 is a diagram showing the outline of the control result good or bad judgment unit 6;
Fig. 12 is a diagram showing the relationship between sub data and symbols in the control output calculating section 3 collectively;
Fig. 13 is a diagram showing processing steps and processing contents in the learning data creation unit 7;
Fig. 14 is a diagram showing an example of data stored in a learning data database DB2;
Fig. 15 is a diagram showing an example of a neural network management table TB;
Fig. 16 is a diagram showing an example of a learning data database DB2;
Fig. 17 is a diagram showing an example of determination contents of the control output selection unit 107;
Fig. 18 is a diagram showing a relationship between a shape evaluation result and a control output;
Fig. 19 is a diagram showing the relationship between the position of the operating end at the control operating stage and the degree of margin;

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described in detail with reference to the drawings. Before that, the knowledge in the present invention and the process leading to the present invention will be described by taking shape control of the rolling mill as an example.

First, in order to solve the above problem in the present invention,

1) Instead of setting the reference shape pattern and the control operation for it separately in advance and learning the control operation method, a combination of the shape pattern and the control operation is plural according to the control effect when the control operation is performed on the same shape. It learns, selects an optimal operation according to the state of the control operation stage from among the plurality of control operations calculated from the plurality of control rules, and performs the control operation by using it. As an optimal operation selection, when the operation range of the control operation stage has a margin, the control operation is performed even if the control effect is small, and when there is no margin, the control operation with a small control effect is not performed

2) A new control rule is not predictable in advance, and in some cases a control rule that could not be predicted at all becomes optimal.

something is needed

In order to realize this, it is necessary to change the combination of the shape pattern used for shape control and the control operation, and change the control operation so that the control result is improved. For that purpose, it is necessary to construct a neural network that can learn the combination of shape patterns and control operations, and to change the output of the neural network control operation for the shape pattern generated by the rolling mill according to the quality of the control result.

If the above is implemented while performing shape control with respect to the rolling mill in operation, since an erroneous control output may be emitted, a shape may deteriorate and operation abnormality, such as a plate|board breakage, may generate|occur|produce. When a plate fracture|rupture generate|occur|produces, it takes time to replace|exchange the roll used by a rolling mill, and the to-be-rolled material in rolling becomes useless, etc. damage is large. Therefore, it is necessary to avoid outputting an erroneous control output to the rolling mill as much as possible.

In view of the above, in the present invention, in order to realize this, the quality of the control operation output by the neural network is verified using, for example, a simple model of the rolling mill, etc. The shape deterioration is prevented by not outputting to the control operation stage. At this time, with respect to the neural network, it is assumed that the control operation for the shape pattern is an error, and learning is performed.

Since there is a possibility that the verification method itself of the quality of the control operation is wrong, the control operation output of the neural network that is judged to be wrong with a certain probability is also output to the control operation stage of the rolling mill, so that the combination of the shape pattern and the control operation is not expected Learning becomes possible

In order to make a neural network learn a control rule, a lot of learning data which is a combination of shape performance and its control method are needed. In this specification, a plurality of training data to be used for learning of a certain neural network is referred to as a training data group. A control rule that is a learning result varies depending on which training data group is used. When the control method for the shape pattern is changed by the above method, a method of creating new learning data and adding it to the existing learning data group is used. The training data in the training data group only increases, and the time required for learning the neural network also increases. Therefore, although it is conceivable to delete learning data based on a period or to delete|delete randomly, it also arises in which the control rule of a learning result is renewed by this. Therefore, it is ideal to add new learning data while leaving the learning data in the learning data group of the present situation.

When new training data is generated, there must be training data that is a combination of the shape performance determined to have no control effect and the control method thereon in the training data group, and even if new training data is added to the training data group is a result of competing, so that the control rule that is the learning result does not become the intended one (the control rule equivalent to the new learning data). In order to solve this problem, while adding new learning data, it is necessary to delete the learning data judged as having no control effect from the learning data group.

Thereby, it becomes possible to limit the number of learning data included in a learning data group. As the learning data increases, the time required for learning increases accordingly. Since the plant control device learns according to a predetermined schedule, it is ideal that the learning time is substantially constant. By deleting the learning data as described above, it becomes possible to limit the increase in the learning data and to learn using the learning data in a certain range.

In addition, the control rules of the neural network are not changed until learning using the new training data group. Therefore, there is a case where control is executed again using the control rule determined to have no control effect. Since learning of the neural network requires time, until a learning result using a new learning data group is obtained, the control method is performed using the modified learning data without performing control using the control rule determined to have no control effect. By doing so, the control effect can be increased.

[Example]

1 shows an outline of a plant control device according to an embodiment of the present invention. The plant control device of FIG. 1 inputs the control target plant 1 and performance data Si from the control target plant 1, and outputs SO determined according to the control rule (neural network) illustrated in FIG. 6 . is given to the control target plant 1 and controlled, and performance data Si from the control target plant 1 is inputted to perform learning, and the learned control rule is transferred to the control execution device 20 . It is comprised by the control method learning apparatus 21 reflected in the control rule in , a plurality of database DBs (DB1 to DB3), and the management table TB of the database DB.

Among them, the control execution device 20 includes a control input data creation unit 2 , a control rule execution unit 10 , a control output calculation unit 3 , a control output suppression unit 4 , a control output determination unit 5 , The control operation disturbance generating unit 16 is constituted as a main element.

In the control execution device 20 , first, the control input data S1 to be given to the control rule execution unit 10 is generated from the performance data Si of the rolling mill which is the control target plant 1 , using the control input data creation unit 2 . write In addition, here, the performance data Si of a rolling mill is a state quantity of a rolling mill, and the state quantity to be given to the control rule execution part 10 among the state quantities of a rolling mill is distinguished as control input data S1. Therefore, it can be said that control input data S1 is also performance data Si of a rolling mill.

The control rule execution unit 10 uses two neural networks (control rules) expressing the relationship between the control target performance data Si (the control input data S1 of the control rule execution unit 10 ) and the control operation stage operation command S2 . Thus, two operating end operation commands of the neural networks are determined from the performance data Si to be controlled, and an optimal operating end operation command is created as the control operating end operation command S2. In the control output calculating part 3, based on the control operation stage operation command S2, the control operation amount S3 with respect to a control operation stage is calculated. Thereby, according to the performance data Si of the control target plant 1, the control operation amount S3 is created using a neural network.

Further, in the control output determination unit 5 in the control execution device 20 , the control operation stage is controlled using the performance data Si from the control target plant 1 and the control operation amount S3 from the control output calculation unit 3 . The manipulated variable output or not data S4 is determined. In the control output suppression unit 4 , it is determined whether or not to output the control MV S3 to the control operation stage according to the control MV output availability data S4 , and the control MV S3 is applied to the control target plant 1 . Output as manipulated variable output SO. As a result, the control operation amount S3 determined to be abnormal is not output to the control target plant 1 . In addition, the control operation disturbance generating unit 16 generates disturbance to the control target plant 1 for the purpose of verifying the plant control device.

The control execution device 20 configured as described above refers to the control rule database DB1 and the output determination database DB3 for execution of the processing as will be described later. The control rule database DB1 is connected so as to be accessible to both the control rule execution unit 10 in the control execution device 20 and the control rule learning unit 11 in the control method learning device 21 to be described later. A control rule (neural network) as a learning result in the control rule learning unit 11 is stored in the control rule database DB1, and the control rule execution unit 10 refers to the control rule stored in the control rule database DB1. The output determination database DB3 is accessible to the control output determination unit 5 in the control execution device 20 .

2 shows a specific configuration example of the control rule execution unit 10 according to an embodiment of the present invention. The control rule execution unit 10 inputs the control input data S1 created by the control input data creation unit 2 , and gives the control operation stage operation command S2 to the control output calculation unit 3 .

The control rule execution unit 10 includes neural networks 101 and 102 . Here, the neural networks 101 and 102 are neural networks learned according to the specification A regarding the control effect to be described later, and the neural network 101 is a specification related to the case of the control effect that the shape deviation is largely corrected as a result of the control operation. The neural network learned according to A1, the neural network 102, is a neural network learned according to the specification A2 regarding the case of the control effect that, as a result of the control operation, the change in shape deviation was small but could be corrected. In the neural networks 101 and 102, basically, by using the control input data S1 created by the control input data creation unit 2 by the method of Patent Document 1 as illustrated in Fig. 6, the neural network operating stage operation command N1, N2 is determined.

The neural network operation stage operation commands N1 and N2 are input to the output presence/absence determination unit 105 . The output presence/absence determination unit 105 determines whether an operation stage operation command N1 having a large control effect and an operation stage operation command N2 having a small control effect is output from the neural network 101, and if not outputted, a flag of no output turns on, and when output is present, the flag of no output is turned off.

In the output presence/absence determining unit 105, when the control effect is high, the output of the neural network formed by learning the case where the control effect is high is present, and when the control effect is low, it is formed by learning the case where the control effect is high. It is determined that there is an output of the neural network formed by learning the case where the output of the neural network does not exist and the case where the control effect is high.

The control rule execution unit 10 according to the embodiment of the present invention, in addition to the control input data S1 created by the control input data creation unit 2 , sets the performance position of the control operation stage as the state quantity Si of the rolling mill to be controlled, the operation margin of the operation stage. It is input to the determination unit 106 . In the operation end operation margin determination unit 106, it is determined whether the control operating end has sufficient margin for operation by control, and if there is room, the margin flag is turned ON, and when there is no margin, the margin flag is to OFF.

In the control output selection unit 107 , the operation end operation command of the neural network 101 is used using the no output flag from the output presence/absence determination unit 105 and the presence flag from the operation stage operation margin determination unit 106 . It is determined which of N1 and the operation stage operation command N2 of the neural network 102 is to be used, and the control output S3 is output to the control output calculating part 3 .

Specifically, as shown in FIG. 17, the control output selection unit 107 determines the content of the judgment by the control output selection unit 107 when the no output flag is OFF (there is an operation stage operation command with a large control effect). Selects the operator-end operation command N1 for , and selects the operator-end operation command N2 when the no output flag is ON (there is no operator-level operation command with a large control effect) and the slack flag is ON, otherwise selects the operator-end operation command N2 That is, when the no output flag is ON (there is no operation command for the operation stage having a large control effect) and the slack flag is OFF, the control output is set to 0.

As a result, if there is an output of the neural network formed by learning the case where the control effect is high and the control effect is high, the control stage of the control target plant is controlled according to the output and the neural network formed by learning the case where the control effect is low. When there is only the output of , and the control effect is low when there is room in the operating stage at the operating stage, the operating stage of the control target plant is controlled according to the output of the neural network formed by learning, and the control effect It is assumed that the control stage of the plant to be controlled is not controlled when only the output of the neural network formed by learning the case where is low exists and there is no leeway in the position of the operation stage in the operation stage.

This point is a way of thinking that, when it is considered that the control effect is small, the life of the control operation stage is secured by not operating the control operation stage if there is not enough room for the control operation stage. On the other hand, if the control effect is large, it can be said that the control method is judged by giving priority to the control effect over the life.

The control rule execution unit 10 further includes neural network selection units 103 and 104, and by referring to the control rule stored in the control rule database DB1, an optimal control rule as a control rule in the neural networks 101 and 102. Select to run.

As described above, in the control rule execution unit 10 of FIG. 2 , a necessary neural network is selected and used from a plurality of neural networks divided for operator groups and control purposes. It is good that the control rule database DB1 also includes, as data from the control target plant 1, performance data (eg, data of the operation team) Si from which a neural network and a quality judgment criterion can be selected. In addition, since there is a relationship that a neural network becomes a control rule when it is executed, in this specification, a neural network and a control rule are not distinguished and are used in the meaning of agreement.

Returning to FIG. 1 , in the control method learning apparatus 21 , the neural networks 101 and 102 used in the control execution apparatus 20 are learned. When the control execution device 20 outputs the control manipulated variable output SO to the control target plant 1 , it takes time for the control effect to actually appear as a change in the performance data Si. For this reason, learning is performed using the data delayed by that time. In Fig. 1, Z ^-1 represents an appropriate time delay function for each data.

The control method learning apparatus 21 is comprised using the control result quality determination part 6, the learning data creation part 7, the control rule learning part 11, and the quality determination database DB4 as main elements.

Among them, the control result quality determination unit 6 uses the performance data Si from the control target plant 1 and the previous performance data value SiO, and the quality determination data S5 stored in the quality determination database DB4, the performance data Si It is determined whether is changed in a good direction or a bad direction, and as a result of the control, good or bad data S6 is output.

In the learning data creation unit 7 in the control method learning apparatus 21 , input data such as the control operation stage operation command S2 , the control operation amount S3 , the control operation amount output permission data S4 created by the control execution apparatus 20 are respectively identical to each other. New teacher data S7a to be used for learning the neural network is created and given to the control rule learning unit 11 by using the time-delayed data and the control result good or bad data S6 from the control result good or bad judgment unit 6 . In addition, the teacher data S7a corresponds to the control operation stage operation command S2 output from the control rule execution unit 10 , and the learning data creation unit 7 provides the control result good or bad quality given by the control result good or bad judgment unit 6 . It can be said that the data obtained by estimating the control operation stage operation command S2 output by the control rule execution unit 10 using the data S6 was obtained as the new teacher data S7a.

3 shows a specific configuration example of the control rule learning unit 11 according to the embodiment of the present invention. The control rule learning unit 11 has an input data creation unit 114 , a teacher data creation unit 115 , a neural network processing unit 110 , and a neural network selection unit 113 as main components. Further, the control rule learning unit 11 includes data S8a obtained by delaying the control input data S1 from the input data creation unit 2 as input from the outside, and new teacher data S7a from the learning data creation unit 7 . , and also refer to the data accumulated in the control rule database DB1 and the learning data database DB2.

In the control rule learning unit 11, the control input data S1 is introduced into the neural network processing unit 110 through the input data generating unit 114 after appropriate time delay compensation.

In the control rule learning unit 11, the new teacher data S7a from the learning data creation unit 7 is also the past teacher data S7b stored in the learning data database DB2 in the teacher data creation unit 115. It is given to the neural network processing unit 110 as teacher data S7c of the included total. These teacher data S7a and S7b are appropriately stored in the learning data database DB2 and used.

Similarly, the input data S8a from the control input data creation unit 2 is input data S8c of the total including the past input data S8b stored in the learning data database DB2 in the input data creation unit 114 as the neural network processing unit ( 110) is given. These input data S8a and S8b are suitably stored in the learning data database DB2 and used.

The neural network processing unit 110 includes a neural network 111 and a neural network learning control unit 112 , and the neural network 111 receives input data S8c from the input data creation device 114 and the teacher data creation unit 115 . of teacher data S7c, the neural network selection unit 113 introduces the selected control rule (neural network), and stores the finally determined neural network in the control rule database DB1.

The neural network learning control unit 112 controls the input data generating unit 114, the teacher data generating unit 115, and the neural network selecting unit 113 at appropriate timing to obtain the input of the neural network 111, and further processing results is stored in the control rule database DB1.

Here, the neural networks 101 and 102 in the control execution device 20 in Fig. 2 and the neural network 111 in the control method learning device 21 in Fig. 3 are both neural networks having the same concept, but to be used Hereinafter, the differences in the basic concepts will be described. First, the neural networks 101 and 102 in the control execution device 20 are neural networks with predetermined contents, and when the control input data S1 is given, the control operation stage operation command S2 as a corresponding output is obtained, that is to say, in one direction. It is a neural network used for processing In contrast, when the neural network 111 in the control method learning apparatus 21 sets the input data S8c and teacher data S7c for the control input data S1 and the control operation stage operation command S2 as learning data, this input/output relationship It is to find a neural network that satisfies .

The way of thinking of the basic processing in the control method learning apparatus 21 comprised as mentioned above is as follows. First, when the content of the control MV output/failure data S4 is "A", the control MV output SO is output to the control target plant 1, and the content of the control result good or bad data S6 is "positive" (the direction in which the performance data Si improves) ), it is determined that the control operation stage operation command S2 output by the control rule execution unit 10 is correct, and training data is created so that the output of the neural network becomes the control operation stage operation instruction S2.

On the other hand, the content of the control MV output/failure data S4 is “negative”, or the control MV output SO is output to the control target plant 1 and the content of the control result good or bad data S6 is “negative” (in the direction in which the performance data Si deteriorates). change), it is determined that the control operation stage operation command S2 output by the control rule execution unit 10 is incorrect, and learning data is created so that the output of the neural network is not generated. At this time, the neural network output is configured so that two types of outputs, positive and negative, are emitted to the same control operation stage as control outputs, and training data is created so that the control operation stage operation command S2 on the output side is not output. do.

In addition, in the control rule learning part 11 illustrated in FIG. 3, as a result of data processing by the neural network learning control part 112, it processes as follows. Here, first, by using the learning data that is a combination of S8c in which the control input data S1 to the control execution device 20 is delayed, and the teacher data S7c created by the teacher data creation unit 115, the control rule execution unit 10 is used. learning of the neural networks 101 and 102 used for In reality, the same neural network 111 as the neural networks 101 and 102 of the control rule execution unit 10 is provided in the control rule learning unit 11, operation tests are conducted under various conditions, the response is learned at that time, and the learning result It is to obtain a confirmed control rule that produces better results. Since it is necessary to perform learning using a plurality of learning data, a plurality of past learning data is retrieved from the learning data database DB2 in which the previously created learning data is accumulated, and a plurality of past learning data are retrieved, learned and processed; The learning data of this time is stored in the learning data database DB2. In addition, the learned neural network is stored in the control rule database DB1 for use in the control rule execution unit 10 .

At this time, the past learning data must contain the learning data that causes the output of the control operation stage operation command S2 that is the basis of the learning data updated this time, and even if the learning data updated this time are added as it is and learned, As a result of learning from conflicting learning data, it interferes with the neural network learning new control methods. Therefore, it is preferable to execute the process of deleting the past learning data most similar to the combination of the original control input data S1 of the learning data updated and added this time, and the control operation stage operation command S2.

Neural network learning can be learned by using the past learning data together whenever new learning data is created, or learning by using the past learning data together after the learning data is accumulated to some extent (for example, 100 pieces). do.

Further, in the control result quality judgment unit 6, a good judgment is made based on the good or bad judgment standard from the good or bad judgment database DB4. Since the judgment result is different depending on the control purpose, the judgment result of the control result is different depending on the control purpose. Therefore, even if the input data is the same, by creating a plurality of neural networks for a plurality of control purposes, and learning by creating and learning each teacher data according to the control purpose even if the input data is the same, one time It is possible to simultaneously learn a plurality of neural networks corresponding to a plurality of control purposes by creating a plurality of teacher data for the input data and using it for learning a neural network corresponding to each teacher data. Here, the plurality of control purposes means, for example, in the case of shape control, which part (plate edge, center, asymmetrical part, etc.) is to be preferentially controlled in the plate width direction, and a plurality of control target items (eg, plate thickness, tension, rolling load, etc.) which one is to be controlled preferentially.

In the case of the above configuration, once the neural network 101 used in the control rule execution unit 10 learns, a new control operation is not performed. Therefore, the control operation disturbance generating unit 16 generates a new operation method at a timely random number, and the new control method is learned by executing the control operation in addition to the control operation amount S3.

Hereinafter, the detail of this plant control method is demonstrated targeting the shape control in a senjimere rolling mill as described in patent document 1. Note that the shape control will be described assuming that the following specifications A and B are adopted.

Specification A is a specification related to the control effect. As a result of the control operation, the case where the shape deviation is largely corrected is referred to as A1, and the case where the change in the shape deviation is small but can be corrected is referred to as A2.

Specification B is a specification for correspondence to a condition that has been determined in advance. To give an example, since the relationship between the shape pattern and the control method changes under various conditions, for example, it is considered necessary to divide the specification B1 into plate width and specification B2 as the steel type. By changing each of the above, the degree of influence on the shape of the shape manipulation end is changed.

In this example, the control target plant 1 is a senjimere rolling mill, and the performance data is the shape performance. Moreover, a senjimere rolling mill is a rolling mill which has cluster rolls for cold rolling hard materials, such as stainless steel. In the senjimere rolling mill, a small-diameter work roll is used for the purpose of applying a pressure drop to a hard material. For this reason, it is difficult to obtain a flat steel plate. As a countermeasure, the structure of the cluster roll and various shape control units are employed. The senjimere rolling mill is generally equipped with 6 divided rolls and 2 rolls referred to as AS-U, in addition to the upper and lower first intermediate rolls having a knitting taper and shiftable. In the example described below, as the shape performance data Si, the detection data of the shape detector is used, and as the control input data S1, a shape deviation that is a difference from the target shape is used. In addition, as control operation amount S3, it is set as the roll shift amount of the AS-U of #1 - #n, and an upper and lower 1st intermediate roll.

Fig. 4 shows the configuration of a neural network when used for shape control of a senjimere rolling mill. Here, the neural network represents the neural networks 101 and 102 for the control rule execution unit 10 and the neural network 111 for the control rule learning unit 11, but both have the same structure.

In the case of shape control of the senjimere rolling mill shown in Fig. 4, the performance data Si from the control target plant 1 includes the shape detector data (here, it is assumed that the shape deviation that is the difference between the actual shape and the target shape is output). It is performance data of the senjimere rolling mill, and the control input data creation unit 2 obtains a standardized shape deviation 201 and a shape deviation step 202 as the control input data S1. Accordingly, the input layer of the neural networks 101 , 102 , 111 is constituted by the normalized shape deviation 201 and the shape deviation step 202 . In Fig. 4, the shape deviation step 202 is input to the neural network input layer, but the neural network may be switched according to the steps.

Moreover, the output layer is comprised by the AS-U operation degree 301 and the 1st intermediate operation degree 302 according to the AS-U which is a shape control operation stage of a senjimere rolling machine, and a 1st intermediate roll. For each operation degree, for AS-U, the AS-U opening direction (the direction in which the roll gap (the gap between the upper and lower work rolls of the rolling mill) widens) and the AS-U closing direction (the direction in which the roll gap narrows) are determined for each AS - has for U. Also, for the first intermediate roll, the first intermediate roll opening direction (the direction in which the first intermediate roll operates toward the outside from the center of the rolling mill), the first intermediate roll closing direction (the direction in which the first intermediate roll operates toward the center of the rolling mill) direction) with respect to the upper and lower first intermediate rolls. For example, when the shape detector is 20 zones and the shape deviation step 202 is set to 3 steps (large, medium, small), the input layer is 23 inputs. In addition, if the AS-U saddles are 7 and the upper and lower first intermediate rolls are shiftable in the plate width direction, the output layer has a total of 18 AS-U operation precision 301 of 14 and 4 first intermediate operation degrees. do. Set the number of layers in the middle layer and the number of neurons in each layer in a timely manner. In addition, as will be described later with reference to FIG. 8 , the neural network output is configured so that two types of outputs, a + direction and a - direction, are generated for each control operation stage for the shape control operation stage of the sensory mill as the output layer.

Fig. 10 shows the shape deviation and the control method. Here, the control method when the shape variation is large is shown in the upper part of FIG. 10, and the control method when the shape variation is small is shown in the lower part of FIG. In addition, the height direction is the magnitude|size of a shape deviation, the horizontal axis direction is the board width direction, the both sides of the board width show the board edge part, and the center shows the board center part. As shown in the upper part of Fig. 10, when the shape variation is large, correcting the overall shape takes priority over the local shape variation in the plate width direction. On the other hand, when the shape variation is small as shown in the lower part of Fig. 10, making the local shape variation small has priority.

As described above, since it is necessary to change the control method according to the size of the shape deviation, as shown in Fig. 4, the shape deviation step 202 is provided and given to the neural networks 101, 102, 111 to determine the size of the shape deviation. . Regarding the shape variation, it is preferable to use, for example, a value normalized to 0 to 1 regardless of the magnitude of the shape variation. This is an example, and it is also conceivable to input the shape deviation to the input layer of the neural network as it is without standardizing the shape deviation, and to change the neural network itself according to the magnitude of the shape deviation (eg, prepare two neural networks and when the shape deviation is large) It is also conceivable to divide the neural network to use and the neural network to use in small cases.

For the neural networks 101, 102, 111 having the same configuration as that of FIG. 4 described above, an operation method for a shape pattern is learned, and shape control is performed using the learned neural network. Even a neural network having the same configuration may have different characteristics depending on the learning conditions, so that different control outputs can be generated with respect to the same shape pattern. The condition of the controlled plant by changing the criterion for determining the quality of the control result based on the condition of the controlled plant or the experience of the operator of the controlled plant, obtaining the relationship between the performance data for the controlled plant and the operation method, and storing it in the database , or other control methods may be used according to the experience of the operator of the plant to be controlled.

Therefore, by using a plurality of neural networks separately according to different conditions of shape performance, it is possible to configure optimal control for various conditions. This corresponds to specification B. The configuration of Fig. 2 described earlier shows a specific example in the case of performing such a specification. In the configuration example of FIG. 2 , a separate neural network is prepared for the neural networks 101 and 102 to be used in the control rule execution unit 10 according to the rolling performance, the name of the rolling mill operator, the steel type of the material to be rolled, the plate width, and the like. It is registered in database DB1. The neural network selection units 103 and 104 select a neural network that matches the condition at that point in time and set it as the neural networks 101 and 102 of the control rule execution unit 10 . In addition, as a condition at that point in time in the neural network selection units 103 and 104 , it is preferable to introduce the plate width data from the performance data Si in the control target plant 1 and select the neural network accordingly. In addition, if the plurality of neural networks used herein have an input layer and an output layer as shown in Fig. 4, the number of layers in the intermediate layer and the number of units in each layer may be different.

In Fig. 7 , the control input data creation unit 2 for creating the control input data S1 (standardized shape deviation 201, shape deviation step 202) for input to the input layer of the neural networks 101, 102, 111 is shown in FIG. shows the outline. Here, as the performance data Si, the shape detector data of a shape detector that detects the plate shape at the time of rolling in the senjimere rolling mill as the control target plant 1 is input, and first, the shape deviation PP value calculating device 210 is used. The shape deviation PP value (Peak To Peak value) S _PP which is the difference between the maximum value and the minimum value of the detection result of each shape detector zone is calculated|required. The shape deviation step calculating unit 211 classifies the shape deviation into three stages: large, medium, and small according to the shape deviation PP value S _PP . The shape is the plate width direction distribution of the elongation rate of the rolled material, and I-UNIT representing the elongation rate in units of 10 to 5 is used as a unit. For example, it is classified as follows.

Here, the shape deviation step becomes (large = 1, medium = 0, small = 0) by the establishment of the formula (1), and the shape deviation step becomes (large = 0, medium = 1) by the establishment of the expression (2) . In addition, here, about the shape deviation of each zone, standardization is performed using S _PM set as S _PM =S _PP .

As described above, the normalized shape deviation 201 and the shape deviation step 202 that are input data to the neural networks 101 and 102 are created. The standardized shape deviation 201 and the shape deviation step 202 are control input data S1 of the control rule execution unit 10 .

The outline|summary of the control output calculating part 3 is shown in FIG. The control output calculation unit 3 includes, in the control rule execution unit 10, a control operation stage operation command S2 that is an output from the neural network 101 (AS-U operation degree 301 in the case of shape control of the senjimere rolling machine); From the first intermediate operation degree 302 (corresponding to this), a control operation amount S3 that is an operation command for each shape control operation stage is created. In addition, here, one data example is shown with respect to the AS-U operation degree 301 and the 1st intermediate operation degree 302 where there exist a plurality of numbers, and each data is a pair of an open direction precision and a closing direction precision. It consists of data.

In the control output calculating section 3, since the input AS-U operation degree 301 has the outputs of each AS-U opening direction and closing direction, the difference between them is multiplied by the conversion gain G _ASU to each AS-U. Outputs the operation command for The conversion gain G _ASU is a conversion gain from the precision to the position change amount since the control output for each AS-U is the AS-U position change amount (unit is the length).

Also similarly, since the input first intermediate operation degree 302 has the first intermediate outer and inner outputs, the operation command for each first intermediate roll shift is output by multiplying their difference by the conversion gain G _1ST . The conversion gain G _1ST is a conversion gain from the precision to the position change amount because the control output for each first intermediate roll is the first intermediate roll shift position change amount (unit is the length).

As described above, the control manipulation amount S3 can be calculated. The control manipulation amount S3 is composed of #1 to #n AS-U position change amount (n is according to the number of saddles of the AS-U roll), upper first intermediate shift position change amount, and lower first intermediate shift position change amount has been Also, Fig. 8 shows a system for adding the disturbance data from the control operation disturbance generating unit 16 to the control operation stage operation command S2.

The outline|summary of the control output determination part 5 is shown in FIG. The control output determination unit 5 includes a rolling phenomenon model 501 and a shape correction quality determination unit 502 , and includes performance data Si from the control target plant 1 , and a control operation amount from the control output calculation unit 3 . S3 and information of the output determination database DB3 are obtained, and control operation amount output permission data S4 to the control operation stage is provided. With this configuration, in the control output determination unit 5, the change in shape when the control operation amount S3 calculated by the control output calculation unit 3 is output to the rolling machine which is the control target plant 1 is controlled by known control. It is predicted by inputting into the model of the target plant 1 (rolling phenomenon model 501 in the case of the embodiment of Fig. 9), and when the shape is expected to deteriorate, the control MV output SO is suppressed to greatly deteriorate the shape. to prevent

In more detail, the control manipulation amount S3 is input to the rolling phenomenon model 501 to predict the shape change due to the control manipulation amount S3, and the shape deviation correction amount prediction data 503 is calculated. On the other hand, shape deviation prediction data 505 is obtained by adding shape deviation correction amount prediction data 503 to shape detector data Si (shape deviation performance data 504 at the present time) from the control target plant 1 , and the shape By evaluating the deviation prediction data 505, it is possible to predict how the shape will change when the control manipulation amount S3 is output to the control target plant 1 . From the current shape variation performance data 504 and shape variation prediction data 505, the shape correction good or bad determination unit 502 determines whether the shape will change in an improving direction or a bad direction, and control operation amount Obtain the output false data S4.

The shape correction quality determination unit 502 specifically determines whether the shape correction is good or bad as follows. First, in order to consider the control priority in the board width direction, the weighting coefficient w(i) in the board width direction is set in the output determination database DB3. Using it, for example, the quality of the shape change is determined using the evaluation function J as in the following (4) equation. In the formula (4), w(i) is a weighting factor, ?fb(i) is a shape deviation performance 504, εest(i) is a shape deviation prediction 505, i is a shape detector zone, and rand is a random number term. to be.

(4) When the evaluation function J of the expression is used, the evaluation function J is positive when the shape is good, and the evaluation function J is negative when the shape is bad. Also, rand is a random number term, and the evaluation result of the evaluation function J is changed randomly. As a result, the evaluation function J may become positive even when the shape deteriorates. Therefore, it is possible to learn the relationship between the shape pattern and the control method even when the rolling phenomenon model 501 is incorrect. Here, rand is timely changed so that the maximum value is increased when the model of the control target plant 1 is uncertain as at the beginning of the trial run, and is set to 0 when a stable control is to be performed by learning the control method to some extent.

In the shape correction quality determination unit 502, the evaluation function J is calculated so that when J ≥ 0, control MV output/false data S4 = 1 (a), and when J < 0, control MV output/false data S4 = 0 ( As in negative), the control MV output or not data S4 is output.

In the control output suppression unit 4 , it is determined whether or not the control MV output SO is output to the control target plant 1 according to the control MV output availability data S4 which is the determination result of the control output determination unit 5 . The control manipulation variable output availability data S4 is a #1 to #n AS-U position change amount output, an upper first intermediate shift position change amount output, and a lower first intermediate shift position change amount output,

IF (Control MV output or not data S4=0) THEN

#1 to #n AS-U position change amount output = 0

Upper first intermediate shift position change amount output = 0

Lower first intermediate shift position change amount output = 0

ELSE

#1 to #n AS-U position change amount output=#1 to #n AS-U position change amount

Upper first intermediate shift position change amount output = Upper first intermediate shift position change amount

Lower first intermediate shift position change amount output = Lower first intermediate shift position change amount

ENDIF

is determined by

In the control execution device 20, the shape control is performed by executing the above calculation from the performance data Si from the control target plant 1 (rolling mill) and outputting the control operation amount output SO to the control target plant 1 (rolling mill). Conduct.

Next, an operation outline of the control method learning device 21 will be described. In the control method learning apparatus 21, time delay data of the data used in the control execution apparatus 20 is used. The time delay Z ^-1 means e ^-TS and indicates delaying by a preset time T. Since the controlled plant 1 has a time response, there is a time delay until the performance data is changed by the control manipulated variable output SO. Therefore, learning is performed using the performance data at the time when only the delay time T has passed after control operation execution. In shape control, after outputting an operation command to the AS-U or the first intermediate roll, it takes several seconds for the shape system to detect a shape change, so it is better to set T=2 to 3 seconds (type of shape detector) However, since the delay time also changes depending on the rolling speed, the optimal time until the change of the control stage changes to the shape change is set as T).

Fig. 11 shows an outline of the operation of the control good or bad judgment unit 6 . In the shape change quality determination part 602, the quality determination evaluation function J _C with respect to the control effect as shown in the following formula is used.

In the formula (5), εfb(i) is the shape variation performance data included in the performance data Si, εlast(i) is the previous value of the shape variation performance data, and wC(i) is the plate width direction for quality determination is the weighting factor. Here, the weighting coefficient wC(i) for the quality judgment is set from the quality judgment database DB4 according to the specification for the control priority in the plate width direction.

In the formula (5), a(a1, a2) is set according to the specifications A1 and A2 regarding the control effect. a1 is when the specification A1 for correcting the shape deviation is large, and a2 is when the specification A2 is used for correcting the shape deviation to be small. S ₂ (j) is a control operation end command to the control device j, and max|S ₂ (j)| is the maximum absolute value of the control operation end command.

The quality of the control result is determined by the evaluation function Jc. Also, even when the control MV output availability data S4 that is the judgment result of the control output judging unit 5 is 0 (control output impossible), the control MV output to the control target plant 1 is actually = 0, but the shape is bad. judge

Here, in the case where the control manipulation amount output success/failure data S4=0, the control result good/failure data S6=-1. Moreover, the threshold upper limit LCU and the threshold lower limit LCL are set in advance based on the threshold condition (LCU≥0≥LCL). At this time, if the result of comparison with the quality judgment evaluation function Jc is Jc>LCU, the control result quality data S6=-1 (shape deteriorates),

If LCU≥Jc≥0, control result good data S6=0 (change in the direction that the shape deteriorates),

If 0>Jc≥LCL, the control result good or bad data S6=1 (change in the direction that the shape improves),

If Jc<LCL, the control result good or bad data S6=0 (the shape improves).

Here, when the control result good or bad data S6 = -1 has a bad shape and thus the output control output is suppressed, the control result good or bad data S6 = 0 has no shape change or has an improved shape, so the output control output is maintained. In this case, the control result good/failure data S6 = 1 is changed in the direction that the shape is good, but there is a possibility that it will become better, so it is a case of increasing the output control amount.

In this way, since the weighting coefficient wC(i) in the plate width direction changes according to the specifications A1 and A2 related to the control effect, the quality judgment evaluation function Jc is different. Therefore, it is considered that the judgment result of the control result good or bad data S6 is also different. Therefore, in the control method learning apparatus 21, the control result quality data S6 is judged about two types of the specifications A1 and A2 regarding a control effect.

Next, the outline|summary of the learning data creation part 7 is demonstrated. As shown in FIG. 1 , in the learning data creation unit 7 , based on the determination result (control result quality or failure data S6 ) from the control result good or bad determination unit 6 , the control operation stage operation command S2 , the control operation amount S3 , Teacher data S7a for the neural network 111 to be used in the control rule learning unit 11 is created from the result of the control output suppression unit's determination (control operation amount output/non-permission data S4).

In this case, the teacher data S7a is an AS-U operation degree 301 and a first intermediate operation degree 302, which are outputs from the output layer of the neural network 111 shown in FIG. 4 . The learning data creation unit 7 includes a control operation stage operation command S2 (AS-U operation degree 301, first intermediate operation degree 302) that is an output of the neural networks 101 and 102, and a control operation amount output SO Using the #1 to #n AS-U position change amount output, the upper first intermediate shift position change amount output, and the lower side first intermediate shift position change amount output, the neural network 111 to be used in the control rule learning unit 11 is used. Create teacher data S7a for

In explaining the outline|summary of the operation|movement of the learning data creation part 7, the relationship between each sub data and preference in the control output calculating part 3 of FIG. 8 is put together in FIG. Here, the AS-U operation degree 301 is representatively shown with respect to the control operation stage operation command S2 that is the output of the neural network 101, and the data on the positive side of the operation degree is OPref, the data on the negative side of the operation degree is OMref, and the control operation disturbance The operation degree randomly generated from the generator 16 will be described with the operation precision random number Oref, the conversion gain as G, and the control operation amount output SO as Cref. As described above, for simplicity, here, as an output from the output layer of the neural network 101 of the control rule execution unit 10, the operation degree positive side and the operation degree negative side side, and randomly generated from the control operation disturbance generating unit 16 The degree of operation to be performed is defined as the operation degree random number. In addition, the control MV output SO to the control operation stage is used as the operation command value.

13 : has shown the process step in the learning data creation part 7, and process content. Here, in accordance with the convention of the symbols in Fig. 12, in the first processing step 71, the operation command value Cref is obtained by the equation (6).

In the next processing step 72, the operation command value Cref is corrected according to the control result good or bad data S6 to be C'ref. Specifically, when the control result good or bad data S6 = -1, the operation command value is based on the formula (7), when the control result good data S6 = 0, (8), and when the control result good data S6 = 1, (9) Let the correction value of Cref be C'ref.

In the processing step 73, the operation precision correction amount ?Oref is obtained from the corrected operation command value C'ref by the equations (10) and (11).

In the processing step 74, the teacher data OP'ref and OM'ref for the neural network 111 are obtained by Equation (12).

In this way, as shown in FIG. 12 , in the learning data creation unit 7 , the operation command value Cref that is actually output to the control target plant 1 is controlled as a result of the judgment in the control result good or bad judgment unit 6 . The operation instruction value correction value C'ref is calculated according to the result good or bad data S6. Specifically, when the control result data S6 = 1, the control direction is OK, but when it is determined that the control output is insufficient, the operation command value is increased by ΔCref in the same direction. Conversely, when the control result is good data S6 = -1, when it is determined that the control direction is wrong, the operation command value is decreased by ΔCref in the reverse direction. Since the conversion gain G is set in advance, since it is a known point, it is possible to calculate|require correction amount (DELTA)Oref if the value of the positive operation precision side and the operation precision negative side are known. Here, ΔCref is set by obtaining an appropriate value in advance through simulation or the like. By the above procedure, in the control rule learning unit 11, OP'ref and OM'ref can be obtained by the above equation (12).

In addition, although the description is given as a simple example in FIG. 13, in reality, the AS-U operation degree 301 for #1 to #n AS-U, and the first intermediate roll shift on the upper side and the first intermediate roll shift on the lower side All of them are performed with respect to one intermediate operation degree 302, and it is set as teacher data (AS-U operation degree teacher data, first intermediate operation degree teacher data) of the neural network 111 to be used in the control rule learning unit 11. .

Fig. 14 shows an example of data stored in the learning data database DB2. In order to learn the neural network 111, a combination of a plurality of input data S8a and teacher data S7a is required. Accordingly, the teacher data S7a (AS-U operation degree teacher data, first intermediate operation degree) created by the learning data creation unit 7 is input data S1 input from the control execution device 20 to the control rule execution unit 10 . It is stored in the learning data database DB2 as a set of learning data S11 in combination with the time delay data S8a of (standardized shape deviation 201 and shape deviation step 202).

In addition, in the plant control device of FIG. 1, various databases DB1, DB2, DB3, DB4, and DB5 are used. represents the configuration of The management table TB has a management table of specifications. Specifically, the management table TB is classified according to specifications A1 and A2 for (B1) sheet width, (B2) steel type, and control priority for specifications. (B1) As the sheet width, for example, 4 divisions of 3 feet width, meter width, 4 feet width, and 5 feet width are used, and as the steel type, about 10 divisions of the steel type (1) to the steel type (10) are used. In addition, about the specification A for the priority of control, it is set as two types, A1 and A2. In this case, there are 80 divisions, and 80 neural networks are classified and used according to the rolling conditions.

The neural network learning control unit 112 associates learning data that is a combination of input data and teacher data as shown in FIG. 14 with a corresponding neural network No. and a used neural network according to the neural network management table TB of FIG. 16 , FIG. It is stored in the learning data database DB2 as shown in .

Whenever the control execution device 20 performs shape control with respect to the control target plant 1, two sets of learning data are created. This is because, with respect to the same input data and control output, two types of teacher data are created because the control result good or bad judgment is made using the two evaluation criteria of the specification A1 and the specification A2 for the priority of control. When the teacher data is accumulated to a certain extent (for example, 200 sets) or newly accumulated in the learning data database DB2, the neural network learning control unit 112 instructs the learning of the neural network 111 .

A plurality of neural networks are stored in the control rule database DB1 according to the management table TB as shown in FIG. 15 , and the neural network learning control unit 112 designates a neural network number that needs to be learned, and the neural network selection unit 113 . The neural network is retrieved from the temporary control rule database DB1 and set as the neural network 111 . The neural network learning control unit 112 instructs the input data creation unit 114 and the teacher data creation unit 115 to retrieve input data and teacher data corresponding to the neural network from the learning data database DB2, and uses them to instruct the neural network We perform learning of (111). In addition, as for the learning method of a neural network, various methods have been proposed, and any method may be used.

When the learning of the neural network 111 is completed, the neural network learning control unit 112 writes back the neural network 111 that is the learning result to the position of the neural network No. in the control rule database DB1 to complete the learning.

Learning may be performed simultaneously for all the neural networks defined in FIG. 15 at regular time intervals (for example, every day), and only the neural network of neural network No. in which new learning data has been accumulated to some extent (for example, 100 sets) You may learn at that point.

By the above, without greatly disturbing the shape of the rolling mill which is the control target plant 1,

1) Instead of setting the reference shape pattern and the control operation for it in advance and learning the control operation method, learn the combination of the shape pattern and the control operation and perform the control operation using it,

2) A new control rule is not predictable in advance, and there are cases where a control rule that could not be predicted at all becomes optimal. Therefore, the control operation stage is operated at random and the control result is found while observing the control result.

In addition, the control rule database DB1 stores the neural network to be used by the control execution device 20. However, if the neural network to be stored is merely initial processing with random numbers, the time until the neural network is trained and its own control becomes possible. it takes Therefore, when the control unit is built with respect to the control target plant 1, the control rules are learned in advance by simulation based on the control model of the control target plant 1 which has become clear at that point in time, and the learning in the simulator is performed. By storing the completed neural network in the database, it is possible to perform performance control to some extent from the beginning of the start of the plant to be controlled.

As is clear from the above description, the contents of the control rule database DB1 formed as a result of the learning processing in the neural network learning control unit 112 are the neural network learned about the specification A1 with high control effect, and the specification with low control effect. Contains the neural network learned about A2. The former is reflected in the neural network 101 of the control rule execution unit through the control rule database DB1, and the latter is reflected in the neural network 102 of the control rule execution unit through the control rule database DB1.

The plant control device of the present invention is actually realized as a computer system, but in this case, a plurality of program groups are formed in the computer system.

These program groups are, for example,

In order to achieve processing by the control execution device, a control rule execution program that gives a control output according to a predetermined combination of performance data and control operation of the plant to be controlled, and whether or not the control output output from the control rule execution program is determined , a control output determination program for notifying the control method learning device that the performance data and control operation are in error, and when the control output determination program outputs a control output to the control target plant, the performance data of the control target plant is deteriorating When it is determined, it is a control output suppression program that prevents the control output from being output to the control target plant,

When the control execution device actually outputs the control output to the control target plant for achieving the processing of the control method learning device, after a time delay until the control effect appears in the performance data, the performance data is better than before the control. A control result good or bad judgment program for achieving the control result good or bad judgment processing for judging whether the control result is good or bad, the good or bad of the control result in the control result good or bad judgment program, and the control output using the control output a learning data creation program for obtaining

Then, the control method learning device learns to obtain a combination of separate performance data and control operation for a plurality of control targets according to the state of the plant to be controlled, and applies the combination of the obtained performance data and control operation to the control rule execution program. It is used as a predetermined combination of performance data and control operation of the control target plant in the present invention.

In addition, when the apparatus of the present invention is applied to an actual plant, it is necessary to set the initial value of the neural network. In this regard, the combination of performance data and control operation is used to control the control target plant before performing control in the control target plant. It is good to create by simulation using a model, and to shorten the learning period of the combination of performance data and control operation in the control target plant.

The operation and effect of the present invention described above will be described in detail with reference to FIGS. 18 and 19 . First, the neural networks 101 and 102 in Fig. 2 each obtain the same control input data S1 from the control input data creation unit 2, but each of the control effects is a neural network reflecting the learning results from different viewpoints. Outputs N1 and N2 of different operating stage operation commands are given. The outputs N1 and N2 of the operation stage operation command may be both obtained at the same timing, or may be obtained only from one side.

18 is a diagram showing a relationship between a shape evaluation result and a control output. Here, the shape evaluation result is taken as the vertical axis and time is taken as the horizontal axis, and it is exemplified that the shape evaluation result has decreased with the passage of time.

In this example, in the first section T1, both the operating stage operation command N1 from the neural network 101 and the operating stage operation command N2 from the neural network 102 are output, but the control output selection unit 107 has a large control effect. The operation stage operation command N1 is selected, and in this case, the presence or absence of a control operation margin at the control operation stage is not considered. Whether there is room or not, it selects the operation stage operation command N1, which has a large control effect, and applies it to the control. Further, even when the operation stage operation command N2 is not output from the neural network 102 , the control output selection unit 107 selects the operation stage operation command N1 having a large control effect, and the control operation margin at the control operation stage is selected. does not take into account the presence or absence of

Next, in the second section T2 , the operation stage operation command N1 is not output from the neural network 101 , but only the operation stage operation command N2 is output from the neural network 102 . A section T21 in the first half of this section T2 indicates a state in which there is room for control operation at the control operation stage, and the control output selection unit 107 selects the operation stage operation command N2 having a small control effect and applies it to the operation command N2. control by On the other hand, in the latter section T22 of this section T2, the control operation in the control operation stage has no margin, and the control output selection unit 107 selects the operation stage operation command N2 having a small control effect. I never do that. As a result, this section becomes uncontrolled.

19 : has shown the relationship between the operation end position in a control operation stage, and a margin. 19 , the vertical axis represents the position of the operating end and the margin, and the horizontal axis represents time. In this figure, if the position of the valve is illustrated as the operating end position, for example, the valve is movable in the range of 0 to the control limit LL. In addition, in the present invention, the margin level LM (0 < clearance level LM < control limit LL) is set, and in the case of an operating stage operation command N1 with a large control effect, the valve can be operated in the range of 0 to control limit LL. do.

On the other hand, in the case of the operation stage operation command N2 having a small control effect, the valve makes the movable range from 0 to the margin level LM. Moreover, in the period T0 in which the valve position at the time of the operating stage operation command N2 with a small control effect has reached the margin level LM, control by the operating stage operation command N2 with a small control effect is inhibited. In addition, this figure shows the movable range to the last, and does not show that there exist two operation stage operation commands at the same time. In addition, although the example in which the margin level was set to the upper limit side with respect to a valve position is shown, it is also possible to set to the lower limit side similarly.

18 and 19, according to the present invention, the following effects can be exhibited. First, for example, in the case of a rolling mill, a plurality of valves serving as an operation stage exist. In this case, when one of the plurality of valves reaches the margin level LM, the control output is not output. Thereby, it becomes possible to suppress consumption of the life of a valve mechanism by moving a valve to the limit when a control effect is small.

Moreover, a part of a some valve may show the event which continues moving in an opening direction, and another part of valve continues moving in a closing direction. Since it is generally desirable for the valve to operate near the center position and not to operate in the end position, limiting the valve within the redundancy level LM is, for example, when starting the next or subsequent control. For example, it is useful in the sense of increasing the responsiveness or increasing the manipulated value. In particular, after this, the operation value in the case where the operation stage operation command N1 with the large control effect is output can be left, so that the operation with the high control effect can be given priority.

The present invention relates to, for example, a control method and a part of a rolling mill which is one of rolling equipment, and there is no particular problem in practical application.

1: Plant to be controlled
2: Control input data creation unit
3: Control output arithmetic unit
4: Control output suppression part
5: Control output judgment unit
6: Control result good or bad judgment unit
7: Training data writing unit
10: control rule execution unit
11: Control rule learning unit
20: control execution device
21: control method learning device
DB1: Control rules database
DB2: Learning data database
DB3: output verdict database
Si: performance data
SO: Control MV output
S1: input data
S2: Control panel operation command
S3: Control MV
S4: Control MV output or not data
S5: Good or bad judgment data
S6: Control result good or bad data
S7a, S7b, S7c: Teacher data
S8a, S8b, S8c: input data (for control rule learning unit)

Claims (14)

The operation stage of the control target plant is controlled according to the output of the neural network learned by the control method learning apparatus for learning the combination of performance data and control operation of the control target plant according to the control effect, and forming a plurality of neural networks with different control effects. A plant control device having a control execution device that gives a control output to
The control execution device controls the operating stage of the control target plant according to the output when the output of the neural network formed by learning the case where the control effect is high and the control effect is high, and learns the case where the control effect is low. When there is only the output of the neural network formed and there is room for the operation stage in the operation stage, the operation stage of the control target plant is controlled according to the output of the neural network formed by learning the case where the control effect is low, and A plant control device characterized in that only the output of the neural network formed by learning the case where the control effect is low exists, and the operating stage of the plant to be controlled is not controlled when there is no leeway in the position of the operating stage at the operating stage . According to claim 1,
The control execution device executes a control rule that selects and assigns an optimal control output in consideration of an output from a plurality of neural networks according to a control effect and a position of the operation stage according to a predetermined combination of performance data and control operation of the control target plant a control output determination unit that determines whether the control output outputted from the control rule execution unit is false, and notifies the control method learning apparatus that the performance data and the control operation are in error; a control output suppressing unit for preventing output of a control output to the control target plant when it is determined that the performance data of the control target plant deteriorates when output is output to the control target plant;
The control method learning device is configured to, when the control execution device actually outputs the control output to the control target plant, after a time delay until the control effect appears in the performance data, whether the performance data is better or worse than before the control a control result good or bad judgment unit for judging whether or not the control result is good or bad, the control result good or bad in the control result good or bad judgment unit; When the control method learning apparatus learns, a combination of separate performance data and control operation is obtained for a plurality of control targets according to the state of the control target plant, and the combination of the obtained performance data and control operation is used in the control rule execution unit is used as a predetermined combination of performance data and control operation of the control target plant, and when the state of the control target plant is the same as the combination of performance data before learning and control operation, control is performed using the learning data after learning plant control unit. According to claim 1,
Since the combination of performance data and control operation is changed according to the magnitude of the performance data of the control target plant, information on the magnitude of the performance data and information that standardizes the performance data and makes pattern recognition easier to perform pattern recognition are used. A plant control device, characterized in that it learns and controls a combination of control operations. According to claim 1,
The control execution device holds a predetermined combination of performance data and control operation of the control target plant as a first neural network, the control method learning device holds the combination of performance data and control operation as a second neural network, and the control A plant control device characterized in that a second neural network obtained as a result of learning in the method learning device is used as the first neural network in the control execution device. According to claim 1,
The control execution device includes a control operation disturbance generating unit that applies disturbance to the control output, and the control method learning unit learns even when disturbance is applied. According to claim 1,
The control method learning apparatus obtains a plurality of combinations of performance data and control operation by learning under a plurality of predetermined specifications, and the control execution apparatus operates the control target plant from among a plurality of combinations of performance data and control operation. A plant control apparatus characterized by selecting a plurality of combinations of one performance data and a control operation according to a state and providing the control output. 5. The method of claim 4,
A plant control device characterized by changing the neural network for learning the combination of the performance data to be used and the operation method according to the magnitude of the performance data. According to claim 1,
Based on the state of the control target plant or the experience of an operator of the control target plant, the quality judgment criteria of the control result are changed, the relationship between the performance data for the control target plant and the operation method is obtained and stored in a database, whereby the control target plant A plant control device, characterized in that the control is performed by a different control method according to the state of the plant or the experience of an operator of the plant to be controlled. According to claim 1,
The combination of the performance data and the control operation is created by simulation using the control model of the control target plant before the control in the control target plant is performed, and the combination of the performance data and the control operation in the control target plant is generated. Plant control device, characterized in that it shortens the learning period. According to claim 1,
The control execution device includes an output presence/absence determining unit for judging the presence or absence of outputs from a plurality of neural networks, and the output presence/absence determining unit includes, when the control effect is high, an output of the neural network formed by learning the case where the control effect is high. and when the control effect is low, it is determined that the output of the neural network formed by learning the case where the control effect is high is not present. A rolling mill control device to which the plant control device according to claim 1 is applied,
The plant to be controlled is a rolling mill, and the performance data is an exit shape of the rolling mill. A control that learns a combination of performance data and control operation of the control target plant according to the control effect, and controls the operating stage of the control target plant according to the output of the neural network learned by the learning unit that forms a plurality of neural networks with different control effects It is a plant control method having a control unit that gives an output,
When the control effect is high because the output of the neural network formed by learning the case where the control effect is high, the control unit controls the operating stage of the control target plant according to the output, and the neural network formed by learning the case where the control effect is low When there is only the output of , and the control effect is low when there is room in the operating stage at the operating stage, the operating stage of the control target plant is controlled according to the output of the neural network formed by learning, and the control effect A plant control method according to claim 1, wherein only the output of the neural network formed by learning the case where n is low, and the operating stage of the control target plant is not controlled when there is no leeway in the position of the operating stage at the operating stage. 13. The method of claim 12,
The control unit selects and provides an optimal control output in consideration of the output from a plurality of neural networks according to the control effect and the position of the operation stage according to a predetermined combination of performance data and control operation of the control target plant, and provides the control. When output is output to the control target plant, when it is determined that the performance data of the control target plant deteriorates, outputting the control output to the control target plant is prevented;
The learning unit, when the control unit actually outputs the control output to the control target plant, after a time delay until the control effect appears in the performance data, whether the performance data is better or worse than before the control, whether the control result is good or bad to determine the quality of the control result, learn the performance data and the teacher data as learning data, and obtain a combination of separate performance data and control operation for a plurality of control targets according to the state of the control target plant by learning, , when the obtained combination of performance data and control operation is used as a predetermined combination of performance data and control operation of the control target plant in the control unit, and the state of the control target plant is the same as the combination of performance data and control operation before learning, A plant control method characterized by performing control using learning data after learning. It is a rolling mill control method to which the plant control method according to claim 12 is applied,
The plant to be controlled is a rolling mill, and the performance data is an exit shape of the rolling mill.

Images