2. The plant control device according to claim 1, wherein the control method learning device learns the combination of the actual data with the control operation to perform the control, with the use of information on magnitude of the actual data and information for normalizing the actual data and facilitating the execution of the pattern recognition, for the purpose of changing the combination of the actual data with the control operation according to the magnitude of the actual data of the plant to be controlled.

3. The plant control device according to claim 1 or 2, wherein the control rule execution unit holds a determined combination of the actual data of the plant to be controlled with the control operation as a first neutral network, the control rule learning unit holds the combination of the actual data with the control operation as a second neutral network, and uses the second neutral network obtained as a result of learning in the control method learning device as the first neural network in the control rule execution unit.

4. The plant control device according to any one of claims 1 to 3, wherein the control execution device includes a control operation disturbance generation unit that gives a disturbance to the control output, and the control method learning device learns with the inclusion of a case where the disturbance is applied.

5. The plant control device according to any one of claims 1 to 4, wherein the control method learning device obtains a plurality of combinations of the actual data with the control operation by learning under a plurality of predetermined specifications, and the control execution device selects a plurality of combinations of one piece of actual data with the control operation according to an operating state of the plant to be controlled from among a plurality of combinations of the actual data with the control operation, and gives the control output.

6. The plant control device according to claim 3, wherein the control method learning device changes the neutral network for learning the combination of the actual data to be used with an operation method according to the magnitude of the actual data.

7. The plant control device according to any one of claims 1 to 6, wherein the control method learning device changes an acceptability determination criterion of the control result based on the state of the plant to be controlled or an operator's experience of the plant to be controlled, obtains relationships between the actual data and the operation method for the respective plants to be controlled, and stores the obtained relationships in a database, to perform the control by a different control method according to the state of the plant to be controlled or the operator' s experience of the plant to be controlled.

8. The plant control device according to any one of claims 1 to 7, wherein the control method learning device creates the combination of the actual result data with the control operation by simulation with the use of a control model of the plant to be controlled before carrying out the control at the plant to be controlled and reduces a learning period of the combination of the actual data with the control operation in the plant to be controlled.

9. A rolling mill control device to which the plant control device according to any one of claims 1 to 8 is applied, wherein the plant to be controlled is a rolling mill, and the actual data is an exit side shape of the rolling mill.

10. A plant control method that controls a plant to be controlled by recognizing a combination pattern of actual data of the plant to be controlled, including: a control method learning device that learns a combination of actual data of the plant to be controlled with control operation; and a control execution device that executes control of the plant to be controlled according to the learned combination of the actual data and the control operation, wherein the control execution device gives a control output according to a determined combination of the actual data of the plant to be controlled and the control operation, determines whether or not the control output is permitted and notifies the control method learning device that the actual data and the control operation are erroneous, and prevents the control output from being output to the plant to be controlled when it is determined that the actual data of the plant to be controlled deteriorates, and the control method learning device determines whether or not a control result of the actual data as compared with actual data before control is permitted after a time delay until a control effect appears in the actual data when the control execution device outputs the control output to the plant to be controlled, obtains teacher data with the use of the acceptability of the control result in the control result acceptability determination unit and the control output, deletes learning data similar to a combination of the actual data before correcting the created learning data with the control operation, and learns with the actual data and the teacher data as the learning data, learns to obtain the individual combinations of the actual data with the control operation for a plurality of control targets according to a state of the plant to be controlled, uses the obtained combinations of the actual result data with the control operation as determined combinations of the actual data of the plant to be controlled with the control operation in the control rule execution unit, and performs the control with the use of the corrected learning data when the state of the plant to be controlled is similar to the combination of the actual data before correction with the control operation.

11. A rolling mill control method to which the plant control device according to claim 10 is applied, wherein the plant to be controlled is a rolling mill, and the actual data is an exit side shape of the rolling mill.

12. A program that causes a computer system to realize a plant control device that controls a plant to be controlled by recognizing a combination pattern of actual data of the plant to be controlled, wherein the computer system comprises: a control method learning device that learns a combination of actual data of the plant to be controlled with control operation; and a control execution device that executes the control of the plant to be controlled according to the learned combination of the actual data with the control operation, for achieving processing of the control execution device, the program includes: a control rule execution program that gives a control output according to a determined combination of the actual data of the plant to be controlled with the control operation; a control output determination program that determines whether or not the control output output by the control rule execution program is permitted and notifies the control method learning device that the actual data and the control operation are erroneous; and a control output prevention program that prevents the control output from being output to the plant to be controlled if the control output determination program determines that the actual data of the plant to be controlled deteriorates when the control output is output to the plant to be controlled, for achieving processing of the control method learning device, the program includes: a control result acceptability determination unit for achieving processing of a control result acceptability determination which determines the acceptability of a control result as to whether the actual data is improved or lowered as compared with actual data before control after a time delay until a control effect appears in the actual data when the control execution device actually outputs the control output to the plant to be controlled; a learning data creation program that obtains teacher data with the use of the acceptability of the control result in the control result acceptability determination program and the control output; a control rule oblivion unit program that deletes learning data similar to a combination of the actual data before correcting the created learning data with the control operation; and a control rule learning program that learns with the actual data and the teacher data as the learning data, and the control method learning device learns to obtain the individual combinations of the actual data with the control operation for a plurality of control targets according to a state of the plant to be controlled, uses the obtained combinations of the actual result data with the control operation as determined combinations of the actual data of the plant to be controlled with the control operation in the control rule execution program, and performs the control with the use of the corrected learning data when the state of the plant to be controlled is similar to the combination of the actual data before correction with the control operation.

The present invention relates to a plant control device, a control method for the plant control device, and a rolling machine control device, and a control method and a program for the rolling mill control device in real time feedback control performed with the use of an artificial intelligence technique such as a neural network. [Background Art]
Up to now, in various plants, a plant control based on various control theories is carried out in order to obtain a desired control result by the control of the plants.
As one example of the plants, for example, in a rolling mill control, a fuzzy control and a neuro-fuzzy control have been applied as a control theory targeted for a shape control that controls a wavy state of a plate as an example of control. The fuzzy control is applied to a shape control using coolant, and the neuro-fuzzy control is applied to a shape control for a Sendzimir rolling mill. Among those controls, as disclosed in Japanese Patent No.

2804161, the shape control employing the neuro-fuzzy control is performed by obtaining a similarity ratio of a difference between an actual shape pattern detected by a shape detector and a target shape pattern to a preset reference shape pattern, and obtaining a control output amount for an operation end according to the similarity ratio under a control rule expressed by a control operation end operation amount for a preset reference shape pattern. Hereinafter, as the conventional technique, the shape control for the Sendzimir rolling mill using the neuro-fuzzy control is used.
FIG. 5 shows a shape control of the Sendzimir rolling mill disclosed in FIG. 1 of Japanese Patent No. 2804161. In the shape control of the Sendzimir rolling mill, the neuro-fuzzy control is used. In this example, a pattern recognition mechanism 51 performs pattern recognition of a shape according to an actual shape detected by a shape detector 52, and calculates which of preset reference shape patterns is closest to the actual shape. A control calculation mechanism 53 executes a control under a control rule using control operation end operation amounts for preset shape patterns as shown in FIG. 6. More specifically, referring to FIG. 6, the pattern recognition mechanism 51 calculates which of shape patterns (s) of 1

to 8 is closest to a difference (As) between the shape actual performances detected by the shape detector 52 and a target shape (sref), and the control calculation mechanism 53 selects and executes one of the control methods 1 to 8.
However, according to the method of Japanese Patent No. 2804161, there are cases in which an operator is required to perform a manual operation during rolling for verification of the control rule, and the verification of the control rule or the like is performed, but in some cases, a shape change contrary to expectation may be shown. In other words, the control rule determined as described above does not always conform to reality. This is due to lack of consideration of mechanical characteristics and changes in an operating condition and a mechanical condition of the rolling mill, but it is difficult to verify whether or not the preset control rule is the best rule, one by one, because there are a large number of conditions to be considered. For that reason, once the control rule is set, the control rule is often left as it is unless there is a problem.
When the control rule is not based on the reality due to the change in operating condition or the like, since the control rule is fixed, it becomes difficult to achieve control accuracy with a certain level or higher. Also,

once the shape control is activated, the operator does not perform the manual operation (this causes a disturbance for control). Therefore, it is difficult to find a new control rule by an operator's manual intervention. Furthermore, even when rolling a new standard rolled material, it is difficult to set the control rule according to the material.
As described above, in the conventional shape control, there is a problem that it is difficult to correct the control rule because the control is performed with the use of the control rule set in advance.
SUMMARY In order to solve the problem described above, a rule that the control rule is randomly changed to improve the shape while performing the shape control as disclosed in Japanese Patent No. 4003733 is learned, to thereby realize the following matters.
1) A new control rule is discovered while performing the shape control during rolling.
2) Since the new control rule is not predictable in advance and the control rule which could not be predicted at all can be optimized, the control operation end is operated

- 6 -
at random, the new control rule is found out while viewing a control result of the operation.
In the conventional art described above, a representative shape is set as a reference shape pattern in advance, and the control is performed on the basis of a control rule indicating a relationship between the reference waveform pattern and the control operation end operation amount. The control rule learning is also related to the control operation end operation amount for a reference waveform pattern, and the predetermined representative reference shape pattern is used as it is. For that reason, there is a problem that the shape control reacts on only a specific shape pattern.
The reference shape pattern is determined based on knowledge about the rolling mill targeted for a human in advance, and the experience in which a shape actual performance and manual intervention operation are accumulated, but it is difficult to cover all of the shapes generated in the target rolling mill and the rolled material. For that reason, when a shape different from the reference shape pattern occurs, the control by the shape control is not executed, the shape deviation remains without being reduced, or the shape is erroneously recognized as a similar

- 7 -
reference shape pattern, erroneous control operation is performed, and conversely the shape may be deteriorated.
For that reason, in the conventional shape control, there is a problem that an improvement of the control accuracy is limited because the control rule is learned with the use of the preset reference shape pattern and the control rule for the reference shape pattern to execute the control.
In order to solve the above problem, for example, it is assumed to provide a plant control device that controls a plant to be controlled by recognizing a combination pattern of actual data of the plant to be controlled, including: a control method learning device that learns a combination of actual data of the plant to be controlled with control operation; and a control execution device that executes control of the plant to be controlled according to the learned combination of the actual data and the control operation, in which
the control execution device includes: a control rule execution unit that gives a control output according to a determined combination of the actual data of the plant to be controlled with the control operation; a control output determination unit that determines whether or not the control output output by the control rule execution

- 8 -
unit is permitted and notifies the control method learning device that the actual data and the control operation are erroneous; and a control output prevention unit that prevents the control output from being output to the plant to be controlled if it is determined that the actual data of the plant to be controlled deteriorates when the control output determination unit outputs the control output to the plant to be controlled,
the control method learning device includes: a control result acceptability determination unit that determines the acceptability of a control result as to whether the actual data is improved or lowered as compared with actual data before control, after a time delay until a control effect appears in the actual data, when the control execution device actually outputs the control output to the plant to be controlled; a learning data creation unit that obtains teacher data with the use of the acceptability of the control result in the control result acceptability determination unit and the control output; and a control rule learning unit that learns with the actual data and the teacher data as the learning data, and the control method learning device learns to obtain the individual combinations of the actual data with the control operation for a plurality of control targets

- 9 -
according to a state of the plant to be controlled, and uses the obtained combinations of the actual result data with the control operation as determined combinations of the actual data of the plant to be controlled with the control operation in the control rule execution unit.
In this example, the control rule learning unit learns with the use of the learning data including multiple pieces of actual data and the teacher data. However, since the control rule learning unit learns with the inclusion of the past learning data as the learning data, if the learning data that causes the control result to be poor is left, conflict occurs and learning becomes insufficient.
In addition, since learning of the control rule is executed after the learning data is accumulated to some extent, or is executed during stoppage of operation or the like, there is a possibility that the erroneous control output is output again until then. For that reason, there is a problem that it takes time to correct the control rule and optimize the control rule.
Therefore, the present invention aims at providing a plant control device, a control method for the plant control device, a rolling mill control device and a control method and a program for the rolling mill control device, which are capable of automatically correcting a control

- 10 -
rule of a control pattern and an operation method used for shape control and the like during the control, and optimizing the control rule in the shortest time.
According to the present invention, there is provided a plant control device that controls a plant to be controlled by recognizing a combination pattern of actual data of the plant to be controlled, including:
a control method learning device that learns a combination of actual data of the plant to be controlled with control operation; and a control execution device that executes control of the plant to be controlled according to the learned combination of the actual data and the control operation,
in which the control execution device includes: a control rule execution unit that gives a control output according to a determined combination of the actual data of the plant to be controlled and the control operation; a control output determination unit that determines whether or not the control output output by the control rule execution unit is permitted and notifies the control method learning device that the actual data and the control operation are erroneous; and a control output prevention unit that prevents the control output from being output

- 11 -
to the plant to be controlled when it is determined that the actual data of the plant to be controlled deteriorates,
the control method learning device includes: a control result acceptability determination unit that determines whether or not a control result of the actual data as compared with actual data before control is permitted after a time delay until a control effect appears in the actual data, when the control execution device outputs the control output to the plant to be controlled; a learning data creation unit that obtains teacher data with the use of the acceptability of the control result in the control result acceptability determination unit and the control output; a control rule oblivion unit that deletes learning data similar to a combination of the actual data before correcting the created learning data with the control operation; and a control rule learning unit that learns with the actual data and the teacher data as the learning data, and
the control method learning device learns to obtain the individual combinations of the actual data with the control operation for a plurality of control targets according to a state of the plant to be controlled, uses the obtained combinations of the actual result data with the control operation as determined combinations of the

- 12 -
actual data of the plant to be controlled with the control operation in the control rule execution unit, and performs the control with the use of the corrected learning data when the state of the plant to be controlled is similar to the combination of the actual data before correction with the control operation.
According to the present invention, the control pattern of the shape pattern and the operation method used in the shape control during the control can be automatically corrected and optimized in the shortest time
BRIEF DESCRIPTION OF THE DRAWINGS [0012]
FIG. 1 is a diagram showing an outline of a plant control device according to an embodiment of the present invention;
FIG. 2 is a diagram showing a specific configuration example of a control rule execution unit 10 according to the embodiment of the present invention;
FIG. 3 is a diagram showing a specific configuration example of a control rule learning unit 11 according to the embodiment of the present invention;

- 13 -
FIG. 4 is a diagram showing a neural network configuration when the present invention is used for a shape control of a Sendzimir rolling mill;
FIG. 5 is a diagram showing a shape control of a Sendzimir rolling mill disclosed in FIG. 1 of Patent Reference 1;
FIG. 6 is a diagram showing a control rule in the shape control of the Sendzimir rolling mill disclosed in FIG. 1 of Patent Reference 1;
FIG. 7 is a diagram showing an outline of an input data creation unit 2;
FIG. 8 is a diagram showing an outline of a control output calculation unit 3;
FIG. 9 is a diagram showing an outline of a control output determination unit 5;
FIG. 10 is a diagram showing a shape deviation and a control method;
FIG. 11 is a diagram showing an outline of a control acceptability determination unit 6;
FIG. 12 is a diagram showing a relationship between data and symbols of various parts in the control output calculation unit 3 in an orderly manner;
FIG. 13 is a diagram showing processing stages and processing contents in a learning data creation unit 7;

- 14 -
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 reflecting processing of a control rule oblivion unit 200; and
FIG. 17 is a diagram showing stored contents of an oblivion rule database DB5.
DETAILED DESCRIPTION
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Prior to the description of the embodiments, knowledge of the present invention and a process leading to the present invention will be described by taking a shape control of a rolling mill as an example.
First, in order to solve the above problem in the present invention, the following matters are required. 1) A combination of a shape pattern and control operation is learnt and the control operation is executed with the use of the learnt combination, instead of a configuration in which a reference shape pattern and control operation

- 15 -
for the reference shape pattern are set, separately, to learn a control operation method.
2) Since a new control rule is not predictable in advance and a control rule which cannot be predicted at all can be optimal, a control operation end is operated at random, and the optimal control rule is found while viewing the control result of operating the control operation end.
In order to realize the above configuration, there is a need to change the combination of the shape pattern used for the shape control and the control operation to change the control operation so as to improve the control result. To that end, there is a need to configure a neural network capable of learning the combination of the shape pattern and the control operation, and change an output of the control operation of the neural network for the shape pattern generated by the rolling mill according to the acceptability of the control result.
When the above control is executed while carrying out the shape control on the rolling mill during operation, an erroneous control output may be issued, resulting in deterioration in the shape and occurrence of operational abnormality such as plate breakage. If the plate breakage occurs, it takes time to replace a roll used in the rolling mill, and a rolled material is wasted during rolling. Thus,

- 16 -
damage is great. For that reason, there is a need to avoid outputting the erroneous control output to the rolling mill as much as possible
Therefore, in the present invention, the acceptability of the control operation output by the neural network is verified with the use of, for example, a simple model of the rolling mill, and the output apparently deteriorated in the shape is not output to the control operation end of the rolling mill to prevent the shape from being deteriorated. At that time, in the neural network, learning is performed assuming that the control operation for that shape pattern is erroneous.
Since there is a possibility that the method per se of verifying the acceptability of the control operation is erroneous, even the control operation output of the neural network determined to be erroneous with a certain probability is output to the control operation end of the rolling mill, thereby being capable of learning the unexpected combination of the shape pattern and the control operation.
In order to cause the neural network to learn the control rule, a large number of pieces of learning data which is the combinations of the shape actual performance and the control method for the shape actual performance

- 17 -
are required. The large number of pieces of learning data used for learning of a certain neural network is expressed as a learning data group in the present specification. The control rule as the learning result varies depending on what learning data group is used. According to the above method, when the control method for the shape pattern is changed, a method is used in which new learning data is created and added to the existing learning data group. The learning data in the learning data group only increases, and a time required for learning the neural network also increases. For that reason, it is conceivable to delete the learning data on the basis of a period, or delete the learning data at random, but there are cases where the control rule of the learning result is renewed in some cases. For that reason, it is ideal to add new learning data while the learning data in the existing learning data group is left.
When the new learning data is generated, there should be learning data which is the combination of the shape actual performance determined to be no control effect and the control method for the shape actual performance in the learning data group. Even if the new learning data is added to the learning data group while the above learning data is left, a conflict resultantly occurs between those pieces

- 18 -
of learning data, and the control rule as the learning result is not intended (the control rule equivalent to the new learning data).
Also, as the learning data increases, a time required for learning increases accordingly. Since the plant control device performs learning according to a predetermined schedule, it is ideal that the learning time is kept almost constant.
Therefore, in the present invention, at the same time when adding the new learning data, the learning data determined to have no control effect is deleted from the learning data group. As a result, an increase in the learning data is restricted and learning can be performed with the use of a certain range of learning data.
Also, the control rule of the neural network does not change until learning is performed with the use of the new learning data group. For that reason, there are cases where the control is executed again with the use of the control rule determined to have no control effect. Since it takes time to learn the neural network, the control is not performed with the use of the control rule determined to have no control effect until the learning result using the new learning data group is obtained, and the control

- 19 -
method based on corrected learning data is used, thereby being capable of enhancing the control effect. [First Embodiment]
FIG. 1 shows an outline of a plant control device according to an embodiment of the present invention. The plant control device in FIG. 1 includes a control target plant 1, a control execution device 20 that receives actual data Si from the control target plant 1 and gives a control operation amount output SO determined according to a control rule (neural network) exemplified in FIG. 6 to the control target plant 1 for controlling the control target plant 1, a control method learning device 21 that receives the actual data S1 from the control target plant 1 to learn, and reflects the learnt control rule on the control rule in the control execution device 20, and multiple database DB (DB1 to DB3) as well as a management table TB of the database DB.
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 prevention unit 4, a control output determination unit 5, a control operation disturbance generation unit 16, and a replacement control rule execution unit 100 as main elements.

- 20 -
Among those components, the control execution device 20 first creates input data S1 of the control rule execution unit 10 from actual data Si of the rolling mill which is the control target plant 1 with the use of the control input data creation unit 2. The control rule execution unit 10 creates a control operation end operation command S2 from the actual data Si to be controlled with the use of a neural network (control rule) expressing a relationship between the actual data Si to be controlled and the control operation end operation command S2. The control output calculation unit 3 calculates a control operation amount S3 to a control operation end based on the control operation end operation command S2. As a result, the control output calculation unit 3 creates the control operation amount S3 with the use of the neural network according to the actual data Si of the control target plant 1.
The control output determination unit 5 in the control execution device 20 determines a control operation amount output availability data S4 to the control operation end with the use of the actual data S1 from the control target plant 1 and the control operation amount S3 from the control output calculation unit 3. The control output prevention unit 4 determines whether or not to allow the control operation amount S3 to be output to the control

- 21 -
operation end, according to the control operation amount output availability data S4, and outputs the control operation amount S3 determined to be available as the control operation amount output SO to be given to the control target plant 1. As a result, the control operation amount S3 determined to be abnormal is not output to the plant target plant 1. Incidentally, for the purpose of verifying the plant control device, the control operation disturbance generation unit 16 generates a disturbance and gives the disturbance to the control target plant 1.
In the replacement control rule execution unit 100, a control rule oblivion unit 200 in the control method learning device 21, which will be described later, extracts information on the learning data determined to be deleted and the corrected learning data from an oblivion rule database DB5 and compares the extracted information with the input data S1 and the output data S2 of the control rule execution unit 10 to output the control output. As a result, when the control output is to deteriorate a state of the control target plant 1, the control rule oblivion unit 200 outputs the control output based on the information on the corrected learning data instead of the control output of the control rule execution unit 10.

- 22 -
For the purpose of executing the above processing, the control execution device 20 configured as described above refers to the control rule database DB1 and the output determination database DB3, as will be further described later. The control rule database DB1 is accessibly connected to both of the control rule execution unit 10 in the control execution device 20 and a control rule learning unit 11 in the control method learning device 21 to be described later. A control rule (neural network) as the 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 accessibly connected to the control output determination unit 5 in the control execution device 20.
FIG. 2 shows a specific configuration example of the control rule execution unit 10 according to the embodiment of the present invention. The control rule execution unit 10 receives the input data S1 created by the control input data creation unit 2 and gives a control operation end operation command S2 to the control output calculation unit 3. The control rule execution unit 10 includes a neural network 101, and the neutral network 101 basically determines the control operation end operation command S2

- 23 -
by the method of Patent Reference 1 as exemplified in FIG. 6. In the present invention, the control rule execution unit 10 further includes a neural network selection unit 102, and the neutral network selection unit 102 selects and executes an optimum control rule as a control rule in the neutral network 101 with reference to the control rule stored in the control rule database DB1. As described above, the control rule execution unit 10 of FIG. 2 selects a necessary neural network from multiple neural networks separated for each operator team and each control purpose and uses the selected neutral network. It is preferable that the control rule database DB1 also includes actual data (data of operation team or the like) Si from which the neutral network and the acceptability criteria can be selected as the data from the control target plant 1. Meanwhile, since there is a relationship that the executed neutral network results in the control rule, the neutral network and the control rule are used synonymously without being distinguished from each other in the present specification.
Returning to FIG. 1, the control method learning device 21 executes the learning of the neural network 101 used in the control execution device 20. When the control execution device 20 outputs the control operation amount

- 24 -
output SO to the control target plant 1, it takes time for a control effect to actually appear as a change in the actual data Si. For that reason, the control method learning device 21 executes the learning with the use of the data delayed by that time. In FIG. 1, reference symbol Z-1 represents an appropriate time delay function for each data.
The control method learning device 21 mainly includes a control result acceptability determination unit 6, a learning data creation unit 7, the control rule learning unit 11, an acceptability determination database DB4, the control rule oblivion unit 200, and the oblivion rule database DB5.
Among those components, the control result acceptability determination unit 6 determines whether the actual data Si changes so as to be acceptable or unacceptable with the use of the actual data Si and an actual data previous value Si0 from the control target plant 1 and the acceptability determination data S5 stored in the acceptability determination database DB4. Then, the control result acceptability determination unit 6 outputs control result acceptability data S6.
The learning data creation unit 7 in the control method learning device 21 creates new teacher data S7a to

- 25 -
be used for learning of the neutral network with the use of data obtained by delaying the input data such as the control operation end operation command S2, the control operation amount S3, and the control operation amount output availability data S4, which are created by the control execution device 20 by the same time, and the control result acceptability data S6 from the control result acceptability determination unit 6. The learning data creation unit 7 then gives the created new teacher data S7a to the control rule learning unit 11. Incidentally, the teacher data S7a corresponds to the control operation end operation command S2 output from the control rule execution unit 10 and the learning data creation unit 7 obtains data obtained by estimating the control operation end operation command S2 output from the control rule execution unit 10 with the use of the control result acceptability data S6 given by the control result acceptability determination unit 6, as the new teacher data S7a.
FIG. 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 includes an input data creation unit 114, a teacher data creation unit 115, a neural network processing unit 110,

- 26 -
and a neural network selection unit 113 as main components. In addition, the control rule learning unit 11 obtains data S8a obtained by delaying in time the input data S1 from the input data creation unit 2 as an input from the external and the new teacher data S7a from the learning data creation unit 7. In addition, the control rule learning unit 11 refers to the data stored in the control rule database DB1 and the learning data database DB3.
In the control rule learning unit 11, the input data S1 is taken into the neural network processing unit 110 through the input data creation unit 114 after an appropriate time delay compensation.
In the control rule learning unit 11, the new teacher data S7a from the learning data creating unit 7 is given to the neutral network processing unit 110 as total teacher data S7c including past teacher data S7b stored in the learning data database DB3 in the teacher data creation unit 115. Those pieces of teacher data S7a and S7b are stored in the learning data database DB3 as appropriate and used.
Likewise, the input data S8a from the control input data creation unit 2 is given to the neutral network processing unit 110 as total input data S8c including past input data S8b stored in the learning data database DB3

- 27 -
in the input data creation unit 114. Those pieces of input data S8a and S8b are stored in the learning data database DB 3 as appropriate and used.
The neural network processing unit 110 includes a neural network 111 and a neural network learning control unit 112. The neural network 111 receives the input data S8c from the input data creation device 114, the teacher data S7c from the teacher data creation unit 115, and the control rule (neural network) selected by the neural network selection unit 113, and stores the finally determined neural network in the control rule database DB1.
The neural network learning control unit 112 controls the input data creation unit 114, the teacher data creation unit 115, and the neural network selection unit 113 at appropriate timings, obtains an input of the neural network 111, and stores the processing result in the control rule database DB 1.
The control rule oblivion unit 200 retrieves learning data similar to the original input data S8a and the original output S2 used for creating the new teacher data from the learning data database DB2 according to the input data S8a, the new teacher data S7a, and the output S2 of the control rule execution unit 20. Then, the control

- 28 -
rule oblivion unit 200 implements a process of replacing the learning data with the new teacher data.
In this example, the neural network 101 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 neural networks of the same concept. A difference in basic concept in utilization between the neural network 101 and the neural network 111 will be described below. First, the neural network 101 in the control execution device 20 is a neural network having predetermined contents, and obtains the control operation end operation command S2 as a corresponding output when the input data S1 is given, which is a neural network used for so-called one-way processing. On the other hand, the neural network 111 in the control method learning device 21 is configured to obtain the neutral network that satisfies an input to output relationship by learning when setting the input data S8c and the teacher data S7c for the input data S1 and the control operation end operation command S2 as the learning data.
The concept of basic processing in the control method learning device 21 configured as described above will be described as follows. First, when the content of the control operation amount output availability data S4 is

- 29 -
"available", the control method learning device 21 outputs the control operation amount output SO to the control target plant 1, and when the content of the control result acceptability data S6 is "acceptable" (the actual result data Si is changed to be acceptable), the control method learning device 21 determines that the control operation end operation command S2 output from the control rule execution unit 10 is correct, and creates the learning data so that the output of the neural network becomes the control operation end operation command S2.
On the other hand, when the content of the control operation amount output availability data S4 is "unavailable" or when the control method learning device 21 outputs the control operation amount output SO to the control target plant 1 and the content of the control result acceptability data S6 is "unacceptable" (the actual data Si is changed to be unacceptable), the control method learning device 21 determines that the control operation end operation command S2 output by the control rule execution unit 10 is erroneous, and creates the learning data so as not to issue the output of the neural network. At that time, as the control output, the neural network output is configured so as to output two types of outputs of + direction and - direction to the same control operation

- 30 -
end, and the learning data is created so as not to output the control operation end operation command S2 on the output side.
Also, in the control rule learning unit 11 illustrated in FIG. 3, as a result of data processing by the neural network learning control unit 112, the processing is performed as follows. In this case, first, the control rule learning unit 11 executes the learning of the neural network 101 used in the control rule execution unit 10 with the use of the learning data which is a combination of S8c obtained by delaying in time the input data S1 to the control execution device 20 and the teacher data S7c created by the teacher data creation unit 115. Actually, the same neural network 111 as the neural network 101 in the control rule execution unit 10 is provided in the control rule learning unit 11, and the control rule learning unit 11 learns a response at that time as an operation test under various conditions, and obtains a control rule that is confirmed to generate a better result as the learning result. Since there is a need to perform learning with the use of plural pieces of learning data, plural pieces of past learning data are extracted from the learning data database DB2 storing the learning data created in the past to execute the learning process and

- 31 -
store the current learning data in the learning data database DB2. Also, the learnt neural network is stored in the control rule database DB1 for use by the control rule execution unit 10.
At that time, the learning data which causes the control operation end operation command S2 as a source of the learning data updated this time to be output should be included in the past learning data. Even if the learning data updated this time is added as it is and learnt, the learning is resultantly executed with conflicting learning data, which prevents the neural network from learning a new control method. For that reason, the control rule oblivion unit 200 is configured to execute a process of deleting the past learning data most similar to the combination of the original input data S1 and the control operation end operation command S2 of the learning data updated and added this time.
Learning of the neural network may be performed with the use of the past learning data together every time the new learning data is created or the learning may be performed with the use of the past learning data together after the learning data has been accumulated to a certain extent (for example, 100 pieces). In that case, the number of learning data included in a learning data group used

- 32 -
for learning may increase every time the new learning data is created, and the learning effect may decrease due to the existence of conflicting learning data.
In the present patent application, as described above, an increase in the learning data can be prevented by deleting the erroneous learning data. For example, the learning is set in advance so as to be executed with the use of 1000 pieces of learning data, and when the new learning data is created, learning data most similar to the combination of the original input data S1 and the control operation end operation command S2 is deleted. This makes it possible to learn the new control method for the neural network without increasing the learning data. As a result, a time required for learning and computer resources (memory, hard disk, and so on) required as a space for saving the learning data can be saved. In addition, the learning efficiency is improved by deleting the conflicting learning data. For that reason, the neural network can be performed in the shortest time.
In addition, in the control result acceptability determination unit 6, the acceptability determination is carried out on the basis of the acceptability criteria from the acceptability determination database DB 4. In the acceptability determination of the control result, since

- 33 -
the determination result is different depending on the control purpose, the multiple neural networks
corresponding to the multiple control purposes are created. The respective teacher data is created according to the control purpose even if the input data is identical, and learnt to create the plural pieces of teacher data for one input data, and used for learning the neutral networks corresponding to the respective teacher data. As a result, the neutral networks corresponding to the multiple control purposes can be learnt at the same time. In the present specification, the multiple control purposes means, for example, in the case of shape control, which part (plate end part, center part, asymmetric part, or the like) is to be preferentially controlled in the plate width direction, which of multiple control target items (for example, plate thickness, tension, rolling load, or the like) is to be preferentially controlled, and the like. In the configuration as described above, once the neural network 101 used in the control rule execution unit 10 learns, a new control operation is not executed. For that reason, the control operation disturbance generation unit 16 generates a new operation method at random in a timely manner, and executes the control operation in

- 34 -
addition to the control operation amount S3 to learn the new control method.
Hereinafter, details of the plant control method will be described with reference to the shape control in the Sendzimir rolling mill as shown in Patent Reference 1. Meanwhile, the shape control will be described assuming that the following specifications A and B are adopted.
The specification A is a specification about priority, and has information on the priority in a plate width direction. For example, in the shape control, it is often difficult to control a shape to a target value over an entire area in the plate width direction from the viewpoint of mechanical characteristics. For that reason, specifications A1 and A2 for the following two priorities are provided in the plate width direction. In those specifications, the specification A1 for priority is "to give priority to a plate end", the specification A2 for priority is "to give priority to the center part", and the control is performed according to the two priorities A1 and A2. When the control is executed, any one of the specifications A1 and A2 for priority is considered.
The specification B is a specification for handling conditions that are known beforehand. As an example, since a relationship between the shape pattern and the control

- 35 -
method changes under various conditions, it is conceivable that there is a need to divide the specification B1 by the plate width and specification B2 by the type of steel, for example. As each of the above specifications changes, the degree of influence of the shape operation end on the shape changes
In this case, the control target plant 1 is the Sendzimir rolling mill, and the actual data is the shape actual performance. The Sendzimir rolling mill is a rolling mill having a cluster roll for cold rolling a hard material such as stainless steel. In the Sendzimir rolling mill, a small-diameter work roll is used for giving a strong pressure to a hard material. For that reason, it is difficult to obtain a flat steel sheet. As a
countermeasure, a structure of a cluster roll and various shape control units are adopted. In general, in the Sendzimir rolling mill, upper and lower first intermediate rolls have a single taper and can be shifted, and also have six divided rollers and two AS-U rolls at the top and bottom. In an example to be described below, detection data of a shape detector is used as the actual data Si of the shape, and a shape deviation, which is a difference from a target shape, is used as the input data S1. The control operation amount S3 is assumed to be the roll shift amount of the

- 36 -
AS-U of # 1 to #n, and the upper and lower first intermediate rolls.
FIG. 4 shows a neural network configuration for use in the shape control of the Sendzimir rolling mill. In this example, the neural network refers to the neural network 101 for the control rule execution unit 10 and the neural network shown in the neural network 111 for the control rule learning unit 11. The structures of those neural networks 101 and 111 are identical in structure with each other.
In a case of the shape control of the Sendzimir rolling mill shown in FIG. 4, the actual data Si from the control target plant 1 is actual data of the Sendzimir rolling mill including the data of a shape detector (in this example, it is assumed that the shape deviation which is the difference between the actual shape and the target shape is output), and the control input data creation unit 2 obtains a normalized shape deviation 201 and a shape deviation stage 202 as the input data S1. As a result, input layers of the neural networks 101 and 111 are configured by the normalized shape deviation 201 and the shape deviation stage 202. In FIG. 4, the shape deviation stage 202 is an input to the neural network input layer, but a neural network may be switched according to the stage.

- 37 -
Further, the output layer is configured by an AS-U operation degree 301 and a first intermediate operation degree 302 according to the AS-U and the first intermediate rolls which are the shape control operation end of the Sendzimir rolling mill. In the respective operation degrees, in the AS-U, each AS-U has an AS-U opening direction (a direction in which the gap (a roll gap (a distance between the upper and lower operation rolls of the rolling mill) opens) and an AS-U closing direction (a direction in which the roll gap is closed).
In addition, in the first intermediate rolls, the upper and lower first intermediate rolls have a first intermediate roll opening direction (a direction in which the first intermediate rolls move toward the outside from a center of the rolling mill), and a first intermediate roll closing direction (a direction in which the first intermediate rolls operate toward the center side of the rolling mill). For example, when the shape detector has 20 zones and the shape deviation stage 202 is 3 stages (large, medium, small), the input layer has 23 inputs. In addition, when it is assumed that the number of saddles of the AS-U are seven and the upper and lower first intermediate rolls can be shifted in a plate width direction, the number of output layers is 18 in total with 14 AS-U operation degrees

- 38 -
301 and four first intermediate operation degrees. The number of intermediate layers and the number of neurons in each layer are set appropriately. As will be described later with reference to FIG. 8, the shape control operation end of the Sendzimir rolling mill, which is the output layer, configures a neural network output such that two kinds of outputs in the plus and minus directions are output to the individual control operation ends.
FIG. 10 shows the shape deviation and control method. In this example, the control method in the case where the shape deviation is large is shown in an upper part of FIG. 10, and the control method in the case where the shape deviation is small is shown in a lower part of FIG. 10. Incidentally, a height direction indicates a magnitude of the shape deviation, and a lateral axis direction indicates the plate width direction. Both sides of the plate width represent plate ends, and the center represents a plate center portion. As shown in the upper part of FIG. 10, when the shape deviation is large, priority is given to correcting the overall shape rather than a local shape deviation in the plate width direction. On the other hand, as shown in the lower part of FIG. 10, when the shape deviation is small, priority is given to reducing the local shape deviation.

- 39 -
In this way, since there is a need to change the control method according to the magnitude of the shape deviation, as shown in FIG. 4, the shape deviation stage 202 is provided and given to the neural networks 101 and 111 to determine the magnitude of the shape deviation. It is preferable to use the shape deviation normalized to, for example, 0 to 1 regardless of the magnitude of the shape deviation. This is just an example, and it is conceivable to input the shape deviation as it is to the input layer of the neural network without normalizing the shape deviation, or it is conceivable to change the neural network per se depending on the magnitude of the shape deviation (for example, two neural networks are prepared, and separated into the neural network to be used when the shape deviation is large and the neural network to be used when the shape deviation is small).
The neural nets 101 and 111 having the configuration as shown in FIG. 4 described above are made to learn how to operate the shape pattern, and the shape control is executed with the use of the learnt neural nets. Even the neural networks of the same configuration have different characteristics depending on the learning conditions and can issue different control outputs for the same shape pattern.

- 40 -
For that reason, optimal control can be configured for diverse conditions by selectively using the multiple neural networks according to other conditions of the shape actual performance. This is compatible with the specification B. The configuration of FIG. 2 described above shows a specific example in which such a specification is performed. In the configuration example of FIG. 2, individual neural networks are prepared for the neural network 101 to be used in the control rule execution unit 10 according to the rolling actual performance, a rolling mill operator name, the steel type of a material to be rolled, the plate width, and so on, and registered in the control rule database DB1 in advance. In the neural network selection unit 102, a neural network matching the condition at that time point is selected and set in the neural network 101 of the control rule execution unit 10. Meanwhile, as the condition at that time in the neural network selection unit 102, it is preferable to take in data of the plate width from the actual data Si in the control target plant 1, and select the neural network according to the taken data. Also, the multiple neural networks to be used in this example may be different from each other in the number of intermediate layers and the number of units in each layer

- 41 -
if the neutral networks have the input layer and the output layer as shown in FIG. 4.
FIG. 7 shows an outline of the control input data creation unit 2 for creating the data S1 (normalized shape deviation 201, shape deviation stage 202) for input to the input layers of the neural networks 101 and 111. In this case, the control unit data creation unit 2 receives, as the actual data Si, shape detector data of the shape detector for detecting the plate shape during rolling in the Sendzimir rolling mill, which is the control target plant 1. First, a shape deviation PP value calculation device 210 obtains a shape deviation PP value (peak to peak value) SPP which is a difference between a maximum value and a minimum value of the detection result of each shape detector zone. A shape deviation stage calculation unit 211 classifies the shape deviation into three stages of large, medium and small based on the shape deviation PP value SPP. The shape is a distribution of an elongation percentage of the material to be rolled in the plate width direction, and I-UNIT expressing the elongation percentage in units of 10 to 5 is used as a unit. For example, the shape deviation is classified as follows.
In this example, it is assumed that the shape deviation stage is classified as (large = 1, medium = 0,

- 42 -
small = 0) by the establishment of Expression (1), the shape deviation stage is classified as (large = 0, medium = 1, small = 0) by the establishment of Expression (2), and the shape deviation stage is classified as (large = 0, medium = 0, small = 1) by the establishment of Expression (3). In this case, the shape deviation of each zone is normalized with the use of SPM with SPM = SPP. [Ex. 1]
SPP> 50I-UNIT [Ex. 2] 50I - UNIT >SPP>10I- UNIT
[Ex. 3] 10I -UNIT >SPP
In the manner described above, the normalized shape deviation 201 and the shape deviation stage 202, which are input data to the neural network 101, are created. The normalized shape deviation 201 and the shape deviation stage 202 are the input data S1 of the control rule execution unit 10.
FIG. 8 shows an outline of the control output calculation unit 3. The control output calculation unit 3 creates the control operation amount S3 which is an operation command to each shape control operation end operation end according to the control operation end

- 43 -
operation command S2 (in the case of the shape control of the Sendzimir rolling mill, the AS-U operation degree 301 and the first intermediate operation degree 302 correspond to the control operation end operation command S2) output from the neural network 101 in the control rule execution unit 10. In this case, one data example is shown for each of the multiple AS-U operation degrees 301 and each of the multiple first intermediate operation degrees 302, and each data is configured by a pair of data having the opening direction degree and the closing direction degree.
In the control output calculation unit 3, since the input AS-U operation degree 301 has outputs of each AS-U opening direction and closing direction, a difference between those outputs is multiplied by a conversion gain GASU, to thereby output an operation command to each AS-U. Since the control output to each AS-U becomes the AS-U position change amount (unit is a length), the conversion gain GASU becomes a conversion gain from the degree to the position change amount.
Since the first intermediate operation degree 302 input similarly has the outputs of the first intermediate outer side and inner side, a difference between those outputs is multiplied by a conversion gain G1ST, to thereby output an operation command to each first intermediate roll

- 44 -
shift. Since the control output to each first intermediate roll is the first intermediate roll shift position change amount (unit is the length), the conversion gain G1ST is the conversion gain from the degree to the position change amount.
With the above configuration, the control operation amount S3 can be calculated. The control operation amount S3 is configured by #1 to #nAS-U position change amounts (n is based on the number of saddles of the AS-U roll), an upper first intermediate shift position change amount, and a lower first intermediate shift position change amount. Incidentally, FIG. 8 shows a system for adding the disturbance data from the control operation disturbance generation unit 16 to the control operation end operation command S2.
FIG. 9 shows an outline of the control output determination unit 5. The control output determination unit 5 includes a rolling phenomenon model 501 and a shape correction acceptability unit 502. The control output determination unit 5 obtains information on the actual data Si from the control target plant 1, the control operation amount S3 from the control output calculation unit 3, and the output determination database DB3, and gives the control operation amount output availability data S4 to

- 45 -
the control operation end. With the configuration described above, in the control output determination unit 5, a change in the shape when the control operation amount S3 calculated by the control output calculation unit 3 is output to the rolling mill, which is the control target plant 1, is input to the model (rolling phenomenon model 501 in the case of the embodiment in FIG. 9) of the known control target plant 1 and predicted. If it is predicted that the shape is to be deteriorated, the control operation amount output SO is inhibited to prevent the shape from deteriorating greatly.
More specifically, the control operation amount S3 is input to the rolling phenomenon model 501 to predict the shape change by the control operation amount S3, and shape deviation correction amount prediction data 503 is calculated. On the other hand, shape deviation prediction data 505 is obtained by adding the shape deviation correction amount prediction data 503 to the shape detector data Si (shape deviation actual data 504 this time) from the control target plant 1. With the evaluation of the shape deviation prediction data 505, it can be predicted how the shape changes when the control operation amount S3 is output to the control target plant 1. The shape correction acceptability determination unit 502

- 46 -
determines whether the shape changes so as to be acceptable or changes so as to be unacceptable based on the current shape deviation actual data 504 and shape deviation prediction data 505, to thereby obtain the control operation amount output availability data S4.
Specifically, the shape correction acceptability unit 502 determines the acceptability of the shape correction as follows. First, as shown in the specifications A1 and A2 on the priority of the shape control, in consideration of the control priority in the plate width direction, a weight coefficient w(i) in the plate width direction is set for each of the specification A1 and the specification A2 in advance in the output determination database DB3. With the use of the weight coefficient w(i), it is determined whether or not the shape change is acceptable by using an evaluation function J like the following Expression (4), for example. Incidentally, in Expression (4), w(i) is a weight coefficient, εfb(i) is the shape deviation actual performance 504, εest(i) is the shape deviation prediction 505, i is a shape detector zone, and rand is a random number term. [Ex. 4]
J=1 jr(w(i).^(i))2 _1^(w(i)lfest(i ))2 +rand

- 47 -
In the case where the evaluation function J of Expression (4) is used, when the shape is improved, the evaluation function J becomes positive and when the shape is deteriorated, the evaluation function J becomes negative. In addition, rand is a random number term, and changes the evaluation result of the evaluation function J at random. As a result, even when the shape is deteriorated, there is a case where the evaluation function J becomes positive. Therefore, even when the rolling phenomenon model 501 is incorrect, a relationship between the shape pattern and the control method can be learnt. In this example, as in an initial stage of trial operation, rand is appropriately changed such that a maximum value is increased when the model of the control target plant 1 is uncertain, and the maximum value is set to 0 when the control method is learnt to some extent and the stable control is intended to be executed.
In the shape correction acceptability determination unit 502, the evaluation function J is calculated, and the control operation amount output availability data S4 is output such that when J > 0 is met, the control operation amount output availability data S4 = 1 (available) is set, and when J < 0 is met, the control operation amount output availability data S4 = 0 (unavailable) is set. The control

- 48 -
output prevention unit 4 determines whether or not to output
the control operation amount output SO to the control target
plant 1, according to the control operation amount output
availability data S4 that is the determination result of
the control output determination unit 5. The control
operation amount output availability data S4 is #1 to #n
AS-U position change amount outputs, the upper first
intermediate shift position change amount output, and the
lower first intermediate shift position change amount
output, and determined by:
IF (control operation amount output availability data S4=0)
THEN
# 1 to # nAS - U position change amount outputs = 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 amounts
Upper first intermediate shift position change amount
output = upper first intermediate shift position change
amount

- 49 -
Lower first intermediate shift position change amount output = lower first intermediate shift position change amount ENDIF.
In the control execution device 20, the above calculation is executed based on the actual data Si from the control target plant 1 (rolling mill) and the control operation amount output SO is output to the control target plant 1 (rolling mill), to thereby execute the shape control.
Next, an operation outline of the control method learning device 21 will be described. In the control method learning device 21, a time delay data of data used in the control execution device 20 is used. The time delay Z-1 means e-TS and indicates that time is delayed by a preset time T. Since the control target plant 1 has a time response, there is a time delay until the actual data changes due to the control operation amount output SO. For that reason, the learning is performed with the use of the actual data at the time when the delay time T elapses after execution of the control operation. In the shape control, since it takes several seconds for a shape meter to detect the shape change after outputting the operation command to the AS-U or the first intermediate roll, it is preferable to set

- 50 -
the delay time T to about 2 to 3 seconds (Since the delay
time also varies depending on the type of the shape detector
and the rolling speed, an optimum time until the change
in the control operation end becomes the shape change may
be set as T.).
FIG. 11 shows an outline of the operation of the
control acceptability determination unit 6. In the shape
change acceptability determination unit 602, an
acceptability evaluation function JC like the following
expression is used.
[Ex. 5]
JC = 1 jr(wC(i).£fb(i))2 _1 jr(wC(i).aast (i))2
In Expression (5), εfb(i) is the shape deviation actual data included in the actual data Si, εlast(i) is the shape deviation actual data previous value, and wC(i) is the plate width direction weight coefficient for acceptability determination. In this example, the weight factor wC(i) for the acceptability determination is set according to the specifications A1 and A2 for the priority of control according to the acceptability determination database DB4. The acceptability of the control result is determined according to the acceptability determination evaluation function Jc. Even in the case where the control operation amount output availability data S4 that is the

- 51 -
determination result of the control output determination unit 5 is 0 (control output unavailable), although the control operation amount output to the control target plant 1 is actually zero, the shape is determined to be deteriorated.
In this example, if the control operation amount output availability data S4 = 0 is met, the control result acceptability data S6 = -1 is set. Further, a threshold upper limit LCU and a threshold addition / decrease LCL
are set in advance under a threshold condition (LCU > 0 > LCL).
At this time, as the result of comparison with the acceptability determination evaluation function Jc, if Jc > LCU, it is assumed that the control result acceptability data S6 = -1 (the shape is deteriorated),
if LCU > Jc > 0, it is assumed that the control result acceptability data S6 = 0 (the shape is changed so as to be deteriorated),
if 0 > Jc > LCL, it is assumed that the control result acceptability data S6 = 1 (the shaped is changed so as to be improved), and
if Jc 0THEN C' ref = Cref - ACref
IF Cref < 0THEN C' ref = Cref + ACref
[Ex. 8]
C'ref = Cref
[Ex. 9]

- 56 -
IF Cref > 0THEN C' ref = Cref + ACref
IF Cref < 0THEN C' ref = Cref - ACref
In the processing stage 73, an operation degree
correction amount ΔOref is obtained from the corrected operation command value Cref by Expressions (10) and (11). [Ex. 10] C' ref = G- ((OPref + AOref) - (OMref - AOref))
[Ex. 11]

AOref = - ( C' ref - {OPref - OMref)) 2 G
1
2 In the processing stage 74, the teacher data OP'ref

and OM'ref to the neural network 111 is obtained from Expression (12). [Ex. 12]
O P 'ref = OPref + AOref OM'ref = OMref - AOref
As shown in FIG. 12, the learning data creation unit 7 calculates the operation command value correction value C ref from the operation command value Cref that is actually output to the control target plant 1 according to the control result acceptability data S6 that is the determination result of the control result acceptability determination unit 6. Specifically, when the control result acceptability data S6 = 1, the control direction is acceptable. However, when it is determined that the control output is insufficient, the operation command value

- 57 -
is increased by ΔCref in the same direction. Conversely, if the control result acceptability data S6 = -1, when it is determined that the control direction is incorrect, and the operation command value is decreased by ΔCref in a reverse direction. Since the conversion gain G is set in advance and is already known, if the values on the positive side of the operation degree and the negative side of the operation degree are known, the correction amount ΔOref can be obtained. In this example, ΔCref is obtained in advance by finding an appropriate value by simulation or the like and set. According to the above procedure, the teacher data OP'ref and OM'ref to be used in the control rule learning unit 11 can be obtained by the above Expression (12).
Although a description is made with a simple case in FIG. 13, actually, all of the AS-U operation degree 301 for #1 to #n AS-U, and the first intermediate operation degree 302 for the upper first intermediate roll shift and the lower first intermediate roll shift are carried out to provide the teacher data (AS-U operation degree teacher data, the first intermediate operation degree teacher data) of the neural network 111 to be used in the control rule learning unit 11.

- 58 -
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 large number of input data S8a and teacher data S7a is necessary. Therefore, the teacher data S7a (the AS-U operation degree teacher data, the first intermediate operation degree) created by the learning data creation unit 7 is combined with the time delay data S8a of the input data S1 (the normalized shape deviation 201 and the shape deviation stage) input to the control rule execution unit 10 by the control execution device 20 and stored as a set of learning data S11 in the learning data database DB2.
At this time, if a set of learning data newly created as described above is merely added to the learning data database DB2, the number of sets of learning data just increases. As the learning data increases, not only it takes time to learn the neural network 111, but also the original set of erroneous learning data causing the neural network 111 to learn the control operation end operation command S2 for the input data S1 corrected as described abo remains in the learning data database DB2, as a result of which the learning is unlikely to be performed efficiently.

- 59 -
Therefore, in the control rule oblivion unit 200, the erroneous learning data is estimated and deleted from the learning database DB2, thereby preventing the learning data from increasing and increasing the learning efficiency.
FIG. 16 shows a configuration example of the learning data database DB2 reflecting the processing of the control rule oblivion unit 200. The control operation end operation command S2, which is a neural network output when the input data is input to the neural network 111 learnt most recently, is paired with the learning data, which is a combination of the input data S1 and the teacher data T1, and stored in the learning data database DB2.
In the case other than the control result acceptability data S6 = 0, since the learning data for correcting the neural network 111 is created, the input data S8a at that time and the time delay S2a of the control operation end operation command S2 are captured, and the input data S1 of each learning data stored in the learning data database DB2 and the control operation end operation command S2 are correlated, and the learning data S10 with the highest correlation is selected. The correlation coefficient for correlating two data series is general, but without regard to the correlation coefficient, a method

- 60 -
using a square mean or an appropriate evaluation function may be adopted.
FIG. 17 shows stored contents of the oblivion rule database DB5. The data S10 to be forgotten and the newly obtained knowledge data S11 are stored in pairs in the oblivion rule database DB5. The data S10 to be forgotten stores the learning data which is the combination of the input data S1 and the teacher data T1 and the newly obtained knowledge data S11 stores the learning data which is the combination of the input data S1 and the corrected data T1. Incidentally, those data S10 and S11 additionally stores the control operation end operation command S2 which is a neural network output when the input data is input to the neural network 11 learnt most recently.
The control rule oblivion unit 200 deletes the selected learning data S10 from the learning data database DB2 and writes the learning data S10 and the corrected learning data S11 in the oblivion rule database DB5. The oblivion rule database DB5 stores the learning data S10 of the neural network 111, which is determined to be corrected by the control result acceptability
determination unit 6, and the corrected learning data S11 in combination in advance. Meanwhile, the data in the oblivion rule database DB 5 is learnt by the neural network

- 61 -
111, and is all deleted when learning by the corrected learning data is completed.
Even if the learning data S11 determined to be required to be corrected by the control result
acceptability determination unit 6 and corrected is created and added to the learning data database DB2, since the corrected learning data S11 is not reflected on the control rule execution unit 10 until the learning of the neural network 111 is executed, the actual control is executed as it is with the old control rule. The replacement control rule execution unit 100 implements the control accuracy early by reflecting the corrected learning data S11 before learning of the neural network 111.
In the replacement control rule execution unit 100, the input data S1 to the control rule execution unit 10 and the control operation end operation command S2 from the control rule execution unit 10 correlate with the learning data to be corrected which is accumulated in the oblivion rule database DB5. If it is determined that the correlation is large, the control operation end operation command S2 of the corrected learning data S11 is output to the control output calculation unit 3.
Incidentally, the plant control device of FIG. 1 uses various databases DB1, DB2, DB3, DB4, and DB5. FIG. 14

- 62 -
shows a configuration of the neural network management table TB for interconnectively managing and operating the respective databases DB1, DB2, DB3, and DB4. The management table TB has a specification management table. Specifically, the management table TB is classified according to the specifications A1 and A2 for a plate width (B1), a steel type (B2), and the control priority in the specification. As the plate width (B1), for example, four sections of 3 feet width, meter width, 4 feet width, and 5 feet width are used. As the steel type, about 10 sections of steel types (1) to (10) are used. Also, as the specification A for the control priority, there are two types of A1 and A2. In that case, there are 80 sections, and 80 neural networks are used separately according to the rolling conditions.
The neural network learning control unit 112 stores the learning data, which is the combination of the input data and the teacher data as shown in FIG. 14, in the learning data database DB2 as shown in FIG. 16 in association with the corresponding neural network No. according to the neural network management table TB of FIG. 15.
Two sets of learning data are created each time the control execution device 20 performs the shape control on

- 63 -
the control target plant 1. This is because two types of teacher data are created since the control result acceptability determination is performed on the same input data and control output with the use of the two evaluation criteria of the specification A1 and the specification A2 for the control priority. 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 learning of the neural network 111.
The multiple neural networks are stored in the control rule database DB1 according to a management table TB shown in FIG. 15, and the neural network learning control unit 112 designates neural network No. requiring learning, and the neural network selection unit 113 extracts the designated neural network from the control rule database DB1 and sets the extracted neural network to 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 extract the input data and the teacher data corresponding to the set neural network from the learning data database DB2, and executes the learning of the neural network 111 with the use of those pieces of extracted data. Various methods have been proposed for the

- 64 -
learning method of the neural network, and any method may be used.
When the learning of the neural network 111 is completed, the neural network learning control unit 112 writes the neural network 111 as the learning result back to a position of the appropriate neural network No. in the control rule database DB1, to thereby complete the learning.
The learning may be performed all at once at regular time intervals (for example, every day) on all of the neural networks defined in FIG. 15. Alternatively, the learning may be performed on only the neural networks of the neural network No. in which new learning data is accumulated to some extent (for example, 100 sets) at that time point.
As described above, without significantly disturbing the shape of the rolling mill which is the control target plant 1, the following configurations can be realized.
1) Instead of setting the reference shape pattern and the control operation for the reference shape patter separately in advance and learning the control operation method, the combination of the shape pattern and the control operation is learnt, and the control operation is executed with the use of the learnt combination.

- 65 -
2) Since the new control rule is not predictable in advance and the control rule which could not be predicted at all can be optimal, the control operation end is operated at random, the new control rule is found out while watching the control result for the operation.
3) Learning of the neural networks which requires time can be made in the shortest time by a reduction in the increase of learning data, and can be controlled with the use of the new control rule even before learning is completed.
The control rule database DB1 stores the neural network to be used by the control execution device 20, but if the stored neural network is just a neural network performing an initial process with a random number, it takes time until the learning of the neural network progresses and a moderate control is enabled. For that reason, when the control unit is built for the control target plant 1, learning of the control rule is executed by simulation in advance based on the control model of the control target plant 1 known at that time, and the neural network that has completed the learning by simulation is stored in the database, thereby being capable of executing the control having some degree of performance from a start of the control target plant.

- 66 -
In addition, when the number of pieces of learning data is restricted to a certain value, a time required for learning can be shortened.
In addition, the number of learning data can be reduced by reducing the similar learning data in the learning data database DB2 with the use of the same method as that of the control rule oblivion unit 200.
The plant control device according to the present invention is actually realized as a computer system, but in this case, multiple program groups are formed in the computer system.
For achieving processing of the control execution device, the program group includes: a control rule execution program that gives a control output according to a determined combination of the actual data of the plant to be controlled with the control operation; a control output determination program that determines whether or not the control output output by the control rule execution program is permitted and notifies the control method learning device that the actual data and the control operation are erroneous; and a control output prevention program that prevents the control output from being output to the plant to be controlled if it is determined that the actual data of the plant to be controlled deteriorates when

- 67 -
the control output is output to the plant to be controlled. For achieving processing of the control method learning device, the program group includes: a control result acceptability determination unit for achieving processing of a control result acceptability determination which determines the acceptability of a control result as to whether the actual data is improved or lowered as compared with actual data before control, after a time delay until a control effect appears in the actual data, when the control execution device actually outputs the control output to the plant to be controlled; a learning data creation program that obtains teacher data with the use of the acceptability of the control result in the control result acceptability determination program and the control output; a control rule oblivion program that deletes learning data similar to a combination of the actual data before correcting the created learning data with the control operation; and a control rule learning program that learns with the actual data and the teacher data as the learning data. The control method learning device learns to obtain the individual combinations of the actual data with the control operation for a plurality of control targets according to a state of the plant to be controlled, uses the obtained combinations of the actual result data

- 68 -
with the control operation as determined combinations of the actual data of the plant to be controlled with the control operation in the control rule execution unit, and performs the control with the use of the corrected learning data when the state of the plant to be controlled is similar to the combination of the actual data before correction with the control operation.
In applying the device of the present invention to a real plant, there is a need to determine an initial value of the neural network. In this regard, it is preferable to create the combination of the actual data and the control operation by simulation with the use of the control model of the control target plant, before carrying out the control in the control target plant, and to shorten the learning period of the combination of the actual data and the control operation in the control target plant.
In the present invention described above, in the plant control system that corrects the control rules by learning, a method of deleting the already created data to be used for learning corresponding to the corrected control rule (forgetting unnecessary actual data) in order to efficiently perform the control, and a method provisionally using the corrected control method until the learning is executed because re-learning of the control

- 69 -
rules are not implemented immediately, are proposed. According to the present invention, learning of the control rules can be efficiently performed, and implementation of the control method with less old control effect can be prevented until the re-learning of the control rules is executed. For that reason, there are advantages that an improvement in the control accuracy, a reduction in a start-up period of the control unit, coping with aging, and the like can be performed.
As a background of the configuration as in the present invention, since it takes time to learn the control rule when deep learning is used (order of hours), a method to add learning data is not preferable, and there may be a need to delete (to forget) the learning data with the use of some method, and there may be a need to manage the learning data which is just growing by another method other than discarding the learning data by delimiting the period. [Industrial Applicability]
The present invention relates to a control method and a control unit of a rolling mill which is, for example, one piece of rolling equipment, and there is no particular problem in practical application. [Description of Reference Signs] 1: CONTROL TARGET PLANT

- 70 -
2: CONTROL INPUT DATA CREATION UNIT
3: CONTROL OUTPUT CALCULATION UNIT
4: CONTROL OUTPUT PREVENTION UNIT
5: CONTROL OUTPUT DETERMINATION UNIT
6: CONTROL RESULT ACCEPTABILITY DETERMINATION UNIT
7: LEARNING DATA CREATION UNIT
10: CONTROL RULE EXECUTION UNIT
11: CONTROL RULE LEARNING UNIT
20: CONTROL EXECUTION DEVICE
21: CONTROL METHOD LEARNING DEVICE
100: REPLACEMENT CONTROL RULE EXECUTION UNIT
200: CONTROL RULE OBLIVION UNIT
DB1: CONTROL RULE DATABASE
DB2: OUTPUT DETERMINATION DATABASE
DB 3: LEARNING DATA DATABASE
DB 4: ACCEPTABILITY DETERMINATION DATA DATABASE
DB 5: OBLIVION RULE DATABASE
Si: ACTUAL DATA
SO: CONTROL OPERATION AMOUNT OUTPUT
S1: INPUT DATA
S2: CONTROL OPERATION END OPERATION COMMAND
S3: CONTROL OPERATION AMOUNT
S4: CONTROL OPERATION AMOUNT OUTPUT AVAILABILITY DATA
S5: ACCEPTABILITY DETERMINATION DATA

S 6: CONTROL RESULT ACCEPTABILITY DATA
S7a, S7b, S7c: TEACHER DATA
S8a, S8b, S8c: INPUT DATA (FOR CONTROL RULE LEARNING UNIT)

We claim:

A plant control device that controls a plant to be controlled by recognizing a combination pattern of actual data of the plant to be controlled, comprising:
a control method learning device that learns a combination of actual data of the plant to be controlled with control operation; and a control execution device that executes control of the plant to be controlled according to the learned combination of the actual data and the control operation,
wherein the control execution device includes: a control rule execution unit that gives a control output according to a determined combination of the actual data of the plant to be controlled and the control operation; a control output determination unit that determines whether or not the control output output by the control rule execution unit is permitted and notifies the control method learning device that the actual data and the control operation are erroneous; and a control output prevention unit that prevents the control output from being output to the plant to be controlled when it is determined that the actual data of the plant to be controlled deteriorates,

the control method learning device includes: a control result acceptability determination unit that determines whether or not a control result of the actual data as compared with actual data before control is permitted after a time delay until a control effect appears in the actual data when the control execution device outputs the control output to the plant to be controlled; a learning data creation unit that obtains teacher data with the use of the acceptability of the control result in the control result acceptability determination unit and the control output; a control rule oblivion unit that deletes learning data similar to a combination of the actual data before correcting the created learning data with the control operation; and a control rule learning unit that learns with the actual data and the teacher data as the learning data, and
the control method learning device learns to obtain the individual combinations of the actual data with the control operation for a plurality of control targets according to a state of the plant to be controlled, uses the obtained combinations of the actual result data with the control operation as determined combinations of the actual data of the plant to be controlled with the control operation in the control rule execution unit, and performs

the control with the use of the corrected learning data when the state of the plant to be controlled is similar to the combination of the actual data before correction with the control operation.
2. The plant control device according to claim 1,
wherein the control method learning device learns
the combination of the actual data with the control operation to perform the control, with the use of information on magnitude of the actual data and information for normalizing the actual data and facilitating the execution of the pattern recognition, for the purpose of changing the combination of the actual data with the control operation according to the magnitude of the actual data of the plant to be controlled.
3. The plant control device according to claim 1 or
2,
wherein the control rule execution unit holds a determined combination of the actual data of the plant to be controlled with the control operation as a first neutral network, the control rule learning unit holds the combination of the actual data with the control operation as a second neutral network, and uses the second neutral network obtained as a result of learning in the control

method learning device as the first neural network in the control rule execution unit.
4. The plant control device according to any one of
claims 1 to 3,
wherein the control execution device includes a control operation disturbance generation unit that gives a disturbance to the control output, and the control method learning device learns with the inclusion of a case where the disturbance is applied.
5. The plant control device according to any one of
claims 1 to 4,
wherein the control method learning device obtains a plurality of combinations of the actual data with the control operation by learning under a plurality of predetermined specifications, and the control execution device selects a plurality of combinations of one piece of actual data with the control operation according to an operating state of the plant to be controlled from among a plurality of combinations of the actual data with the control operation, and gives the control output.
6. The plant control device according to claim 3,
wherein the control method learning device changes
the neutral network for learning the combination of the

actual data to be used with an operation method according to the magnitude of the actual data.
7. The plant control device according to any one of
claims 1 to 6,
wherein the control method learning device changes an acceptability determination criterion of the control result based on the state of the plant to be controlled or an operator's experience of the plant to be controlled, obtains relationships between the actual data and the operation method for the respective plants to be controlled, and stores the obtained relationships in a database, to perform the control by a different control method according to the state of the plant to be controlled or the operator' s experience of the plant to be controlled.
8. The plant control device according to any one of
claims 1 to 7,
wherein the control method learning device creates the combination of the actual result data with the control operation by simulation with the use of a control model of the plant to be controlled before carrying out the control at the plant to be controlled and reduces a learning period of the combination of the actual data with the control operation in the plant to be controlled.

9. A rolling mill control device to which the plant
control device according to any one of claims 1 to 8 is
applied,
wherein the plant to be controlled is a rolling mill, and the actual data is an exit side shape of the rolling mill.
10. A plant control method that controls a plant to
be controlled by recognizing a combination pattern of
actual data of the plant to be controlled, including: a
control method learning device that learns a combination
of actual data of the plant to be controlled with control
operation; and a control execution device that executes
control of the plant to be controlled according to the
learned combination of the actual data and the control
operation,
wherein the control execution device gives a control output according to a determined combination of the actual data of the plant to be controlled and the control operation, determines whether or not the control output is permitted and notifies the control method learning device that the actual data and the control operation are erroneous, and prevents the control output from being output to the plant to be controlled when it is determined that the actual data of the plant to be controlled deteriorates, and

the control method learning device determines whether or not a control result of the actual data as compared with actual data before control is permitted after a time delay until a control effect appears in the actual data when the control execution device outputs the control output to the plant to be controlled, obtains teacher data with the use of the acceptability of the control result in the control result acceptability determination unit and the control output, deletes learning data similar to a combination of the actual data before correcting the created learning data with the control operation, and learns with the actual data and the teacher data as the learning data, learns to obtain the individual combinations of the actual data with the control operation for a plurality of control targets according to a state of the plant to be controlled, uses the obtained combinations of the actual result data with the control operation as determined combinations of the actual data of the plant to be controlled with the control operation in the control rule execution unit, and performs the control with the use of the corrected learning data when the state of the plant to be controlled is similar to the combination of the actual data before correction with the control operation.

11. A rolling mill control method to which the plant
control device according to claim 10 is applied,
wherein the plant to be controlled is a rolling mill, and the actual data is an exit side shape of the rolling mill.
12. A program that causes a computer system to
realize a plant control device that controls a plant to
be controlled by recognizing a combination pattern of
actual data of the plant to be controlled,
wherein the computer system comprises: a control method learning device that learns a combination of actual data of the plant to be controlled with control operation; and a control execution device that executes the control of the plant to be controlled according to the learned combination of the actual data with the control operation,
for achieving processing of the control execution device, the program includes: a control rule execution program that gives a control output according to a determined combination of the actual data of the plant to be controlled with the control operation; a control output determination program that determines whether or not the control output output by the control rule execution program is permitted and notifies the control method learning device that the actual data and the control operation are

erroneous; and a control output prevention program that prevents the control output from being output to the plant to be controlled if the control output determination program determines that the actual data of the plant to be controlled deteriorates when the control output is output to the plant to be controlled,
for achieving processing of the control method learning device, the program includes: a control result acceptability determination unit for achieving processing of a control result acceptability determination which determines the acceptability of a control result as to whether the actual data is improved or lowered as compared with actual data before control after a time delay until a control effect appears in the actual data when the control execution device actually outputs the control output to the plant to be controlled; a learning data creation program that obtains teacher data with the use of the acceptability of the control result in the control result acceptability determination program and the control output; a control rule oblivion unit program that deletes learning data similar to a combination of the actual data before correcting the created learning data with the control operation; and a control rule learning program that learns

with the actual data and the teacher data as the learning data, and
the control method learning device learns to obtain the individual combinations of the actual data with the control operation for a plurality of control targets according to a state of the plant to be controlled, uses the obtained combinations of the actual result data with the control operation as determined combinations of the actual data of the plant to be controlled with the control operation in the control rule execution program, and performs the control with the use of the corrected learning data when the state of the plant to be controlled is similar to the combination of the actual data before correction with the control operation.