The invention relates to a plant controller and its controlling method, and a rolling mill controller and its controlling method and program, in a feedback control at real time performed by using an artificial intelligence technology such as a neural network.
Hitherto, plant control based on various kinds of control theories has been performed in various kinds of plants in order to obtain desired control results through their control.
As one example of the plant, in a rolling mill, a fuzzy control or a neuro-fuzzy control is applied, for example, as a control theory targeted for a shape control of controlling a waviness state of a sheet. The fuzzy control is applied to the shape control using a coolant and the neuro-fuzzy control is applied to the shape control in a Sendzimir mill. The shape control by using the neuro-fuzzy control is performed by acquiring a similarity ratio of a difference between an actual shape pattern detected by a shape detector and a target shape pattern and a reference shape pattern previously set and from the similarity ratio, requiring a control output amount as for an operating end, according to a control rule represented by a control operating end operation amount for the previously set reference shape pattern, as shown in Japanese Patent No. 2804161. Hereinafter, a shape control in the Sendzimir mill

using the neuro-fuzzy control is used as the conventional technique.
Fig. 5 shows a shape control in the Sendzimir mill described in Fig. 1 of Japanese Patent No. 2804161. The neuro fuzzy control is used for the shape control in the Sendzimir mill. In this example, a pattern recognition mechanism 51 recognizes the shape pattern according to the actual shape detected by a shape detector 52 and calculates which one of the previously set reference shape patterns is closest to the actual shape. The control calculation mechanism 53 performs control using a control rule formed by the control operating end operation amount for the shape patterns previously set as shown in Fig. 6. Speaking more specifically as for Fig. 6, the pattern recognizing mechanism 51 calculates which one of the shape patterns (s) from 1 to 8 is closest to a difference (As) between the actual shape detected by the shape detector 52 and the target shape (sref), and the control calculation mechanism 53 selects one of the control methods 1 to 8 and performs the above.
In the method of Japanese Patent No. 2804161, however, there is a case of checking a control rule through an operator' s manual operation during rolling in order to check the control rule, and an unexpected shape change sometimes occurs in some cases. In other words, there occurs the case where the control rule determined as mentioned above does not reflect the reality. This is because of the inadequate examination of the machine characteristics and a change in the rolling mill operation state and the machine condition; however, it is difficult to examine the previous set control rules as the respective best rules one by one because the conditions to be considered are too many. Once

the control rule is set, it often remains as it is unless some problem occurs.
When the control rule becomes incapable of coping with the reality because of an operation condition change, it is hard to achieve a more accurate control because of the fixed control rule. Further, once the shape control stops, an operator does not do the manual operation (it becomes a disturbance for control) and a new control rule is difficult to find through the operator's manual interruption. Further, also in rolling a material of a new standard, it is difficult to set a control rule in accordance to the material.
As mentioned above, in the conventional shape control, the previously set control rules are used and the control rules are hard to modify disadvantageously.
In order to solve the problem, by changing the control rule at random while performing the shape control and learning a rule for better shape as shown in Japanese Patent No. 4003733,
1) a new control rule is found while performing the shape control during rolling, and
2) because a new control rule is not always predictable but a control rule not predictable may be the best in some cases, a control operating end is cooperated at random and its control result is watched to find the best rule.
SUMMARY In the conventional technique, a typical shape is previously set as a reference shape pattern, to perform control based on a control

rule indicating a relation with a control operation end operation amount for the reference wave pattern. The control rule is learned about the control operating end operation amount for the reference wave pattern, while using the predetermined typical reference shape pattern as it is. This disadvantageously results in a shape control just reacting to only a specified shape pattern.
The reference shape pattern is determined by man according to the knowledge about a target rolling mill and the experience having the actual shapes and the manual intervention operations accumulated, and it is difficult to cover all the shapes occurring in a target rolling mill and a material to be rolled. When a shape different from the reference shape pattern is generated, control by the shape control is not performed but a shape deviation is left without being suppressed, or it is recognized as the similar reference shape pattern by mistake and a wrong control operation is performed further to deteriorate the shape on the contrary.
As mentioned above, the conventional shape control is defective in improvement of control accuracy because of learning a control rule using the previously set reference shape pattern and the corresponding control rule and performing control.
Therefore, the invention provides "a plant controller of performing control on a control target plant, in recognition of a combination pattern of the performance data thereof, including a controlling method learning device that learns a combination of performance data and control operation of the control target plant, and a control executing device that performs control on the control

target plant depending on the learned combination of performance data and control operation, in which the control executing device includes a control rule executing unit that gives a control output according to a predetermined combination of performance data and control operation of the control target plant, a control output determining unit that determines permission or inhibition of the control output supplied by the control rule executing unit and notifies the controlling method learning device that the above performance data and control operation are an error, and a control output suppressing unit that inhibits supplying a control output to the control target plant when the control output determining unit determines that in being supplied to the control target plant, the control output is possible to deteriorate the performance data of the control target plant;
the controlling method learning device includes a control result quality determining unit that, when the control executing device actually supplies a control output to the control target plant, determines quality of a control result whether the performance data has become better or worse than before the control at a delayed time until the control effect appears in the performance data, a learning data creating unit that obtains teacher data, using the quality of the control result in the control result quality determining unit and the control output, and a control rule learning unit that learns the performance data and the teacher data as learning data; and through learning by the controlling method learning device, other combinations of performance data and control operation are obtained on a plurality of control targets according to a state of the control target plant

and the obtained combinations of performance data and control operation are used as predetermined combinations of performance data and control operation of the control target plant in the control rule executing unit."
Further, the invention relates to "a rolling mill controller with the plant controller applied, in which the control target plant is a rolling mill and the performance data is of a shape on an output side of the rolling mill."
Further, the invention relates to "a plant controlling method of performing control on a control target plant, in recognition of a combination pattern of the performance data thereof, in which a controlling method learning device that learns a combination of performance data and control operation of the control target plant, and a control executing device that performs control on the control target plant depending on the learned combination of performance data and control operation are provided,
the control executing device gives a control output according to a predetermined combination of performance data and control operation of the control target plant, determines permission or inhibition of the control output, notifies the controlling method learning device that the above performance data and control operation are an error, and inhibits supplying a control output to the control target plant when determining that if being supplied to the control target plat, the control output is possible to deteriorate the performance data of the control target plant, when the control executing device actually supplies a control output to the control target plant,

the controlling method learning device determines quality of a control result whether the performance data has become better or worse than before the control at a delayed time until the control effect appears in the performance data, obtains teacher data, using the quality of the control result and the control output, learns the performance data and the teacher data as learning data, hence to obtain other combinations of performance data and control operation on a plurality of control targets according to a state of the control target plant, and uses the obtained combinations of performance data and control operation as predetermined combinations of performance data and control operation of the control target plant in the control rule executing unit.
Further, the invention relates to "a rolling mill controlling method with the plant controlling method applied, in which the control target plant is a rolling mill and the performance data is of a shape on an output side of the rolling mill."
Further, the invention relates to "a program for realizing a plant controller of performing control on a control target plant, in recognition of a combination pattern of the performance data thereof, by a computer system having a controlling method learning device that learns a combination of performance data and control operation of the control target plant, and a control executing device that performs control on the control target plant depending on the learned combination of performance data and control operation,
the above program including:
for the purpose of achieving the processing of the control

executing device, a control rule executing program for giving a control output according to a predetermined combination of performance data and control operation of the control target plant; a control output determining program for determining permission or inhibition of the control output supplied by the control rule executing program and notifying the controlling method learning device that the above performance data and control operation are an error; and a control output suppressing program for inhibiting supplying a control output to the control target plant when the control output determining program determines that in being supplied to the control target plant, the control output is possible to deteriorate the performance data of the control target plant; and
for the purpose of achieving the processing of the controlling method learning device, a control result quality determining program for, when the control executing device actually supplies the control output to the control target plant, achieving control result quality determination processing of determining the quality of a control result whether the performance data has become better or worse than before the control at a delayed time until the control effect appears in the performance data; a learning data creating program for obtaining teacher data, using the quality of the control result in the control result quality determining program and the control output; and a control rule learning program for learning the performance data and the teacher data as learning data, and
through learning by the controlling method learning device, other combinations of performance data and control operation are obtained

on a plurality of control targets according to a state of the control target plant and the obtained combinations of performance data and control operation are used as predetermined combinations of performance data and control operation of the control target plant in the control rule executing program."
According to the invention, the control rule of shape pattern and operation method can be automatically modified as the optimum one for use in the shape control during controlling. Therefore, it is possible to improve the control accuracy, shorten the start-up period of a control unit, and cope with the aged deterioration advantageously.
BRIEF DESCRIPTION OF THE DRAWINGS
Fig. 1 is a view showing the outline of a plant controller according to one embodiment of the invention;
Fig. 2 is a view showing a concrete component example of a control rule executing unit 10 according to the embodiment of the invention;
Fig. 3 is a view showing a concrete component example of a control rule learning unit 11 according to the embodiment of the invention;
Fig. 4 is a view showing a neural network structure when the invention is used for a shape control of the Sendzimir mill;
Fig. 5 is a view showing the shape control of the Sendzimir mill described in Fig. 1 of Japanese Patent No. 2804161;
Fig. 6 is a view showing a control rule in the shape control of the Sendzimir mill described in Fig. 1 of Japanese Patent No. 2804161;
Fig. 7 is a view showing the outline of a control input data creating unit 2;

Fig. 8 is a view showing the outline of a control output calculating unit 3;
Fig. 9 is a view showing the outline of a control output determining unit 5;
Fig. 10 is a view showing about a shape deviation and a controlling method;
Fig. 11 is a view showing the outline of a control result quality determining unit 6;
Fig. 12 is a view properly showing a relation between each data and symbol in the control output calculating unit 3;
Fig. 13 is a view showing the processing steps and the processing contents in a learning data creating unit 7;
Fig. 14 is a view showing a data example stored in a learning data database DB2;
Fig. 15 is a view showing an example of a neural network control table TB; and
Fig. 16 is a view showing an example of the learning data database DB2.
DETAILED DESCRIPTION
Hereinafter, one embodiment of the invention will be described in detail using the drawings; previous to that, findings in the invention and the background arriving at the invention are described using a shape control of a rolling mill as an example.
At first, in order to solve the problem in the invention, it is necessary:

1) a reference shape pattern and its control operation are separately set, not the control operation methods but various combinations of shape pattern and control operation are learned, to perform the control operation using the above; and
2) since a new control rule is not a predictable one but the never predictable control rule as the optimum rule, a control operating end is operated at random and while watching its control results, a new control rule is found.
In order to realize the above, it is necessary to change the combination of shape pattern and control operation used for shape control and to modify the control operation to obtain a better control result. In order to do the above, it is necessary to form a neural network capable of learning the combinations of shape pattern and control operation and to change the output of the control operation of the neural network for the shape pattern generated in a rolling mill depending on the quality of the control results.
When the above is performed while performing the shape control on the operating rolling mill, a wrong control output may be occasionally supplied, hence to deteriorate the shape and generate an operation failure such as sheet rupture. Once the sheet rupture occurs, it takes much time to replace rolls used in the rolling mill or a material to be rolled during the rolling is wasted, which is big damage. Therefore, it is necessary to avoid issuing a wrong control output to the rolling mill as much as possible.
As mentioned above, in order to realize the above in the invention, the quality of the control operation output from the neural network

is examined, for example, by using a simple model of the rolling mill, and the output considered to be definitely deteriorated in shape is not to be output to the control operating end of the rolling mill, hence to avoid the shape deterioration. Here, the neural network learns that the control operation for that shape pattern is wrong.
Since there is a possibility that the examination method itself about the quality of the control operation is wrong, also an unexpected combination of shape pattern and control operation can be also learned by supplying the control operation output of the neural network judged to be wrong with a certain probability to the control operating end of the rolling mill. First Embodiment
Fig. 1 shows the outline of a plant controller according to one embodiment of the invention. The plant controller in Fig. 1 includes a control target plant 1, a control executing device 20 which receives performance data Si from the control target plant 1 and gives a control operation amount output SO defined according to a control rule (neural network) exemplified in Fig. 6 to the control target plant 1 to control the same plant, a controlling method learning device 21 which receives the performance data Si from the control target plant 1 to learn and reflects the learned control rule as the control rule of the control executing device 20, a plurality of databases DB (DB1 to DB3), and a control table TB of the database DB.
The control executing device 20 includes a control input data creating unit 2, a control rule executing unit 10, a control output calculating unit 3, a control output suppressing unit 4, a control output

- 13 -
determining unit 5, and a control operation disturbance generating unit 16 as main elements.
Of the above, in the control executing device 20, at first, the control input data creating unit 2 creates input data S1 of the control rule executing unit 10 according to the performance data Si of a rolling mill as the control target plant 1. The control rule executing unit 10 creates a control operating end operation command S2 from the performance data Si of the control target, using a neural network (control rule) representing a relation between the performance data Si of the control target and the control operating end operation command S2. The control output calculating unit 3 calculates a control operation amount S3 for the control operating end, based on the control operating end operation command S2. According to this, the control operation amount S3 is created depending on the performance data Si of the control target plant 1, by using the neural network.
Further, the control output determining unit 5 in the control executing device 20 determines control operation amount output permission/inhibition data S4 to the control operating end, using the performance data Si from the control target plant 1 and the control operation amount S3 from the control output calculating unit 3. The control output suppressing unit 4 determines the output
permission/inhibition of the control operation amount S3 to the control operating end, depending on the control operation amount output permission/inhibition data S4 and outputs the permissive control operation amount S3 as the control operation amount output SO given to the control target plant 1. According to this, the control operation

- 14 -
amount S3 determined abnormal is not to be output to the control target plant 1. The control operation disturbance generating unit 16 is to create a disturbance in order to examine the plant controller and to give the above to the control target plant 1.
The control executing device 20 constituted as mentioned above executes the processing, referring to the control rule database DB1 and the output determination database DB3 as described later. The control rule database DB1 is coupled to both the control rule executing unit 10 within the control executing device 20 and a control rule learning unit 11 within the controlling method learning device 21 described later in an accessible way. The control rules (neural networks) as the learning results in the control rule learning unit 11 are stored in the control rule database DB1 and the control rule executing unit 10 refers to the control rule stored in the control rule database DB1. The output determination database DB3 is coupled to the control output determining unit 5 within the control executing device 20 in an accessible way.
Fig. 2 shows a concrete component example of the control rule executing unit 10 according to the embodiment of the invention. The control rule executing unit 10 receives the input data S1 created by the control input data creating unit 2 and gives the control operating end operation command S2 to the control output calculating unit 3. The control rule executing unit 10 includes a neural network 101 and in the neural network 101, a control operating end operation command S2 is defined basically according to the method in Japanese Patent No. 2804161 as exemplified in Fig. 6. In the invention, the control rule

- 15 -
executing unit 10 further includes a neural network selecting unit 102, to select the optimum control rule as the control rule in the neural network 101, by referring to the control rules stored in the control rule database DB1 and execute the above. As mentioned above, the control rule executing unit 10 in Fig. 2 selects a necessary neural network from a plurality of neural networks divided by the operator team and the control purpose and uses the above. The control rule database DB1 may also contain the performance data (data of the operating team) Si that can select a neural network and a quality determination reference, as the data from the control target plant 1. Here, considering such a relation that when a neural network is executed, it becomes a control rule, the neural network and the control rule are not distinguished but used synonymously in the specification.
Returning to Fig. 1, the controlling method learning device 21 learns about the neural network 101 used in the control executing device 20. When the control executing device 20 outputs the control operation amount output SO to the control target plant 1, it takes a long time by the time the control effect actually appears as a change of the performance data Si. Therefore, learning is performed with the data delayed by the time. In Fig. 1, a reference symbol Z-1 represents a proper time lag function as for each data piece.
The controlling method learning device 21 includes a control result quality determining unit 6, the learning data creating unit 7, the control rule learning unit 11, and a quality determination database DB4 as the main elements.
Of the above, the control result quality determining unit 6

- 16 -
determines whether the performance data Si has changed in a direction of becoming better or worse, by using the performance data Si and the previous performance data Si0 from the control target plant 1 and quality determination data S5 stored in the quality determination database DB4 and outputs control result quality data S6.
The learning data creating unit 7 within the controlling method learning device 21 creates new teacher data S7a used for learning the neural network, using the data obtained by delaying the input data such as the control operating end operation command S2, the control operation amount S3, and the control operation amount output
permission/inhibition data S4 created by the control executing device 20 by the same time and control result quality data S6 from the control result quality determining unit 6 and gives the above to the control rule learning unit 11. Here, the teacher data S7a corresponds to the control operating end operation command S2 output by the control rule executing unit 10 and it can be the data required by the learning data creating unit 7 which estimates the control operating end operation command S2 output by the control rule executing unit 10, using the control result quality data S6 from the control result quality determining unit 6, as the new teacher data S7a.
Fig. 3 shows a concrete component example of the control rule learning unit 11 according to the embodiment of the invention. The control rule learning unit 11 includes an input data creating unit 114, a teacher data creating unit 115, a neural network processing unit 110, and a neural network selecting unit 113 as the main components. Further, the control rule learning unit 11 receives data S8a which is obtained

- 17 -
by delaying the input data S1 from the control input data creating unit 2 as the input from the outside, receives the new teacher data S7a from the learning data creating unit 7, and refers to the data accumulated in the control rule database DB1 and the learning data database DB2.
In the control rule learning unit 11, after a proper delay compensation, the input data S1 is taken in by the neural network processing unit 110 through the input data creating unit 114.
In the control rule learning unit 11, the teacher data creating unit 115 gives the new teacher data S7a from the learning data creating unit 7 to the neural network processing unit 110, as the total teacher data S7c including the teacher data S7b in the past stored in the learning data database DB2. These teacher data S7a and S7b are properly stored in the learning data database DB2 and used.
Similarly, the input data creating unit 114 gives the input data S8a from the control input data creating unit 2 to the neural network processing unit 110, as the total input data S8c including the input data S8b in the past stored in the learning data database DB2. The input data S8a and S8b are properly stored in the learning data database DB2 and used.
The neural network processing unit 110 includes the neural network 111 and a neural network learning controlling unit 112, and the neural network 111 takes in the input data S8c from the input data creating device 114, the teacher data S7c from the teacher data creating unit 115, and the control rule (neural network) selected by the neural network selecting unit 113 and stores the neural network finally determined in the control rule database DB1.

- 18 -
The neural network learning controlling unit 112 controls the input data creating device 114, the teacher data creating unit 115, and the neural network selecting unit 113 at a proper timing, receives the input from the neural network 111, and stores the processing results in the control rule database DB1.
Here, both the neural network 101 in the control executing device 20 in Fig. 2 and the neural network 111 in the controlling method learning device 21 in Fig. 3 are of the same concepts; however, a difference in the basic concept for use can be described as follows. At first, the neural network 101 in the control executing device 20 is a neural network having the previously defined contents to require the control operating end operation command S2 as the output corresponding to the input data S1 given, in other words, a neural network used for the processing in one direction. On the contrary, the neural network 111 in the controlling method learning device 21 is to require a neural network that satisfies the input and output relation through learning when the input data S8c about the input data S1 as well as the control operating end operation command S2 and the teacher data S7c are set as the learning data.
The basic processing concept in the controlling method learning device 21 constituted as mentioned above is as follows. At first, when the contents of the control operation amount output
permission/inhibition data S4 are “permission”, the control operation amount output SO is output to the control target plant 1; when the contents of the control result quality data S6 are “good” (the performance data Si changes in a direction of becoming better), the

- 19 -
control operating end operation command S2 output by the control rule executing unit 10 is determined to be right, and the learning data is created so that the output of the neural network may be the control operating end operation command S2.
On the other hand, when the contents of the control operation amount output permission/inhibition data S4 are “inhibition”, or when the control operation amount output SO is output to the control target plant 1 and the contents of the control result quality data S6 are “bad” (the performance data Si changes in a direction of becoming worse), the control operating end operation command S2 output by the control rule executing unit 10 is determined to be wrong, and the learning data is crated so as not to issue the output of the neural network. Here, the neural network is designed to issue the two kinds of control outputs in the + direction and the - direction to the same control operating end and creates the learning data not to output the control operating end operation command S2 on the output side.
In the control rule learning unit 11 exemplified in Fig. 3, the data processing by the neural network learning controlling unit 112 is as follows. Here, the neural network 101 used in the control rule executing unit 10 is made to learn by using the learning data as the combination of the data S8c obtained by delaying the input data S1 to the control executing device 20 and the teacher data S7c created by the teacher data creating unit 115. Actually, the control rule learning unit 11 is provided with the same neural network 111 as the neural network 101 of the control rule executing unit 10 and performs operation tests under various conditions to learn the responses, and obtains the control

- 20 -
rule which is confirmed to generate the better result as the learning result. Since learning has to be performed by using several pieces of learning data, the several pieces of the learning data in the past are taken from the learning data database DB2 where the learning data created in the past are accumulated, to do the learning processing and the current learning data is stored in the learning data database DB2. Further, the neural network having finished learning is stored in the control rule database DB1, to be used for the control rule executing unit 10.
The learning of the neural network may be performed using together the learning data in the past, every time new learning data is created, or after a lump of learning data (for example, one hundred data pieces) are accumulated.
The control result quality determining unit 6 determines the quality based on the quality determination reference from the quality determination database DB4. The quality determination of the control results depends on the control purposes; therefore, a plurality of neural networks corresponding to the control purposes are created, and respective teacher data pieces are created corresponding to the control purposes even if the input data is the same, hence to learn the neural networks with the teacher data. A plurality of teacher data pieces are created for one input data piece and used for the learning of the neural networks corresponding to the respective teacher data pieces, which makes it possible to do the learning of the respective neural networks corresponding to the respective control purposes at once. Here, the plural control purposes in the case of the shape control

- 21 -
include, for example, which portion (sheet end portion, center portion, asymmetric portion, etc.) should be preferentially controlled in a sheet width direction, which of several control target items (for example, sheet thickness and tension, rolling load, etc.), and the like.
In the case of the above structure, once the neural network 101 used in the control rule executing unit 10 finished learning, a new control operation is not performed. Therefore, new operation methods are properly generated at random number by the control operation disturbance generating unit 16, and added to the control operation amount S3, to execute the control operation, thereby learning a new control method.
Hereinafter, the details of the plant controlling method intended for shape control in the Sendzimir mill as shown in Japanese Patent No. 2804161 will be described. The shape control will be described assuming that the following specifications A and B are adopted.
The specification A is a specification about the priority, including the information about the priority in a sheet width direction. For example, as for the shape control, to control a sheet to a target value is often hard in the whole sheet width direction because of the characteristics of machine. Therefore, specifications A1 and A2 are provided about the following two types of priorities in the width direction. The specification A1 is about “the priority given to the sheet end portion” and the specification A2 is about “the priority given to the center portion”; thus, control according to the two priorities A1 and A2 is performed. Control is performed considering which priority of the specification A1 or A2.

- 22 -
The specification B is a specification about response to the previously determined condition. Taking one example, since a relation between a shape pattern and a control method varies depending on various conditions, the specification B has to be divided as follows; for example, a specification B1 is about the sheet width and a specification B2 is about the steel type. Influence degree to the shape of a shape-operating end varies depending on the respective changes.
In this example, the control target plant 1 is the Sendzimir mill; therefore, the performance data is the actual shape. The Sendzimir mill is a rolling mill including a cluster roll for cold-rolling a hard material such as stainless. The Sendzimir mill uses a work roll of small diameter for the purpose of giving a high draft to a hard material. Therefore, it is hard to obtain a flat steel sheet. As a countermeasure, a cluster roll structure and various shape controlling units are adopted. The Sendzimir mill generally includes single tapered upper and lower first intermediate rolls which can be shifted, and further six divided rolls and two rolls called AS-U in upper and lower positions. In the example described below, detection data of a shape detector is used as the performance data Si of the shape and a shape deviation as a difference from the target shape is used as the input data S1. As the
control operation amount S3, the roll shift amounts of #1 to #n AS-U and the upper and lower first intermediate rolls are used.
Fig. 4 shows the neural network structure used for the shape control of the Sendzimir mill. Here, the neural network means the neural network 101 for the control rule executing unit 10 and the neural network 111 for the control rule learning unit 11 and they have the

- 23 -
same structure.
In the example of the shape control in the Sendzimir mill as shown in Fig. 4, the performance data Si from the control target plant 1 is the performance data of the Sendzimir mill including the data (here, the output of a shape deviation that is a difference between the actual shape and the target shape) of the shape detector, and the control input data creating unit 2 obtains a standardized shape deviation 201 and a shape deviation stage 202 as the input data S1. According to this, each input layer of the neural networks 101 and 111 is formed by the standardized shape deviation 201 and the shape deviation stage 202. In Fig. 4, although the shape deviation stage 202 is the input to the neural network input layer, the neural networks may be switched depending on the stages.
An output layer is formed by AS-U operation degree 301 and a first intermediate operation degree 302 in combination of the AS-U and the first intermediate roll as the shape control operating end in the Sendzimir mill. Each operation degree as for the respective AS-U includes an AS-U opening direction (a direction of expanding a roll gap (interval between the upper and lower work rolls of the rolling mill)) and an AS-U closing direction (a direction of narrowing the roll gap). Further, as for the first intermediate roll, each of the upper and lower first intermediate rolls has a first intermediate roll opening direction (a direction of the first intermediate roll operating from the rolling mill center to the outside) and a first intermediate roll closing direction (a direction of the first intermediate roll operating toward the rolling mill center). For example, when the shape detector

- 24 -
defines the shape deviation stage 202 as three stages (large, middle, small) in twenty zones, the input layer has twenty-three inputs. When with seven saddles of the AS-U, the upper and lower first intermediate rolls can be shifted in a sheet width direction, the output layer has eighteen outputs in total including fourteen AS-U operation degrees 301 and four first intermediate degrees. The number of the layers of the intermediate layer and the number of neurons in each layer are properly set. Although it will be described later with reference to Fig. 8, as for the shape control operating ends of the Sendzimir mill as the output layer, the neural network is designed to issue two types
of outputs in a + direction and a - direction to the individual control operating ends.
Fig. 10 shows the shape deviation and the control method. Here, the upper side of Fig. 10 shows the control method when the shape deviation is large and the lower side of Fig. 10 shows the control method when the shape deviation is small. A height direction indicates the size of the shape deviation, a horizontal direction indicates the sheet width direction, the both sides of the sheet width indicate the sheet end portions, and the center indicates the sheet center portion. As shown in the upper side of Fig. 10, when the shape deviation is large, modification of the whole shape takes priority over that of the local shape deviation in the sheet width direction. On the other hand, as shown in the lower side of Fig. 10, when the shape deviation is small, priority is given to the reduction of the local shape deviation.
Thus, since the control method has to be changed depending on the size of the shape deviation, the shape deviation stage 202 is

- 25 -
provided as shown in Fig. 4 and given to the neural networks 101 and 111, hence to determine the size of the shape deviation. The shape deviation can be determined by using, for example, standardized stages 0 to 1, regardless of the size of the shape deviation. This, however, is one example, and the shape deviation may be input to the input layer of the neural network without being standardized, or the neural network itself may be changed depending on the size of the shape deviation (for example, two neural networks are prepared and divided into a neural network used in the case of a large shape deviation and a neural network used in the case of a small shape deviation).
Learning of the operation methods of the shape patterns is performed in the above mentioned neural networks 101 and 111 as constituted in Fig. 4, and the shape control is performed using the neural network having learned. Even the neural networks having the same structure have various characteristics depending on the learning conditions and can be designed to issue various control outputs to the same shape pattern.
Therefore, by using a plurality of the neural networks in a divided way according to other conditions than the actual shape, the optimum control can be achieved on the various conditions. This corresponds to the specification B. The structure of Fig. 2 having been described above shows the concrete example in the case of performing the above specification. In the example of Fig. 2, as the neural network 101 used in the control rule executing unit 10, the separate neural networks are prepared according to the rolling performance, rolling mill operator name, steel type to be rolled, and sheet width, and registered

- 26 -
in the control rule database DB1. The neural network selecting unit 102 selects a neural network which agrees with the condition at that time and sets the above as the neural network 101 in the control rule executing unit 10. As the condition at that time, in the neural network selecting unit 102, sheet width data is taken in from the performance data Si in the control target plant 1, and the neural network according to this should be selected. The plural neural networks used here may have a different number of the layers in the intermediate layer and a difference number of units in each layer as far as the neural network includes the input layer and the output layer as shown in Fig. 4.
Fig. 7 shows the outline of the control input data creating unit 2 which creates the data S1 (standardized shape deviation 201, shape deviation stage 202) to be supplied to the input layers of the neural networks 101 and 111. Here, as the performance data Si, the shape detector data of the shape detector which detects a sheet shape at rolling in the Sendzimir mill as the control target plant 1 is received and at first a shape deviation PP value (Peak To Peak value) SPP that is a difference between the maximum value and the minimum value of the detection results in the respective shape detector zones, is required in a shape deviation PP value calculating unit 210. The shape deviation is classified into three stages of large, medium, and small, according to the shape deviation PP value SPP, in a shape deviation stage calculating unit 211. The shape is estimated by the sheet width direction distribution of the extension coefficient of a rolled material and I-UNIT to represent the extension coefficient as 10 to 5 unit is used by the unit. For example, it is classified as the

- 27 -
following expressions.
Here, the shape deviation stage is defined as (large = 1, medium = 0, and small = 0) when the expression (1) is satisfied; the shape deviation stage is defined as (large = 0, medium = 1, and small = 0) when the expression (2) is satisfied; and the shape deviation stage is defined as (large = 0, medium = 0, and small = 1 when the expression
(3) is satisfied. Here, the shape deviation in each zone is standardized by using SPM, with SPM = SPP.
[Expression 1]
SPP> 50I-UNIT
[Expression 2]
50I - UNIT >SPP>10I- UNIT
[Expression 3] 10I -UNIT >SPP
As mentioned above, the standardized shape deviation 201 and the shape deviation stage 202 are created as the input data to the neural network 101. The standardized shape deviation 201 and the shape deviation stage 202 are the input data S1 of the control rule executing unit 10.
Fig. 8 shows the outline of the control output calculating unit 3. The control output calculating unit 3 creates the control operation amount S3 as the operation command to each shape control operating end, according to the control operating end operation command S2 (corresponding to the AS-U operation degree 301 and the first intermediate operation degree 302 in the example of the shape control in the Sendzimir mill) that is the output from the neural network 101

- 28 -
in the control rule executing unit 10. Here, only one example of the AS-U control degree 301 and the first intermediate operation degree 302, each formed by a pair of data including the opening direction degree and the closing direction degree, is shown, of several degrees 301 and 302.
Since the input AS-U control degree 301 has the outputs of the AS-U opening direction and closing direction, the control output calculating unit 3 multiplies the difference therebetween by the conversion gain GASU and outputs the operation command to each AS-U. The conversion gain GASU becomes a conversion gain from the degree to the position change amount, because the control output to each AS-U becomes the AS-U position change amount (the unit is length).
Similarly, since the input first intermediate operation degree 302 has the outputs of the first intermediate outside and inside, the above unit 3 multiplies the difference therebetween by the conversion gain G1ST and outputs the operation command to each first intermediate roll shift. The conversion gain G1ST becomes a conversion gain from the degree to the position change amount because the control output to each first intermediate roll becomes a first intermediate roll shift position change amount (the unit is length).
As mentioned above, the control operation amount S3 can be calculated. The control operation amount S3 includes #1 to #n AS-U position change amounts (n is 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. Fig. 8 shows a system of adding the disturbance data from the control operation disturbance

- 29 -
generating unit 16 to the control operating end operation command S2.
Fig. 9 shows the outline of the control output determining unit 5. The control output determining unit 5 includes a rolling phenomenon model 501 and a shape modification permission/inhibition determining unit 502, and upon receipt of the performance data Si from the control target plant 1, the control operation amount S3 from the control output calculating unit 3, and the information of the output determination database DB3, the unit 5 gives the control operation amount output permission/inhibition data S4 to the control operating end. According to the structure, when the control output determining unit 5 predicts that the shape gets worse by inputting a change of the shape at a time of outputting the control operation amount S3 calculated in the control output calculating unit 3 to the rolling mill as the control target plant 1 to the known model of the control target plant 1 (in the case of the embodiment of Fig. 9, the rolling phenomenon model 501), the control operation amount output SO is suppressed, hence to prevent the shape from becoming worse.
More specifically, the control output determining unit 5 receives the control operation amount S3 in the rolling phenomenon model 501 to predict the shape change by the control operation amount S3 and calculates shape deviation modification amount prediction data 503. On the other hand, the above unit obtains shape deviation prediction data 505 by adding the shape deviation modification amount prediction data 503 to the shape detector data Si (the shape deviation performance data 504 at the current point) from the control target plant 1, and estimates the shape deviation prediction data 505, hence to be able

- 30 -
to predict how the shape changes when outputting the control operation amount S3 to the control target plant 1. The shape modification permission/inhibition determining unit 502 determines whether the shape changes in a direction of becoming better or worse, according to the current shape deviation performance data 504 and the shape deviation prediction data 505 and obtains the control operation amount output permission/inhibition data S4.
The shape modification permission/inhibition determining unit 502 determines permission or inhibition of the shape modification in the following way. At first, as shown in the specifications A1 and A2 about the priority of the shape control, in order to consider the control priority in the sheet width direction, the weight coefficient w(i) in the sheet width direction is set for each of the specifications A1 and A2, in the output determination database DB3. By using the above, for example, the permission or inhibition of the shape change is determined by an evaluation function J as the following expression (4). In the expression (4), the symbol w(i) is the weight coefficient, sfb(i) is a shape deviation result 504, sest(i) is a shape deviation prediction 505, i is a shape detector zone, and rand is a random number item. [Expression 4]
J = 1JT (w(i)■ sfb(i))2 -1 JT(w(i)-sest(i))2 + rand
n i=1 n i=1
When using the evaluation function J of the expression (4), the evaluation function J becomes positive when the shape gets better, while the evaluation function J becomes negative when the shape gets worse. Further, the rand is the random number item to change the evaluation

- 31 -
results of the evaluation function J in a way of random number. According to this, even when the shape gets worse, there occurs the case where the evaluation function J gets positive; therefore, even in the case where the rolling phenomenon model 501 is not right, the relation between the shape pattern and the control method can be learned. Here, the rand is properly changed; for example, its maximum value is made larger when the model of the control target plant 1 is uncertain, and it is 0 when the control method is learned to some degree and a stable control is wanted, similarly at a time of the initial trial operation.
The shape modification permission/inhibition determining unit 502 calculates the evaluation function J and outputs the control operation amount output permission/inhibition data S4 as follows: when J > 0, the control operation amount output permission/inhibition data S4 = 1 (permission); and when J < 0, the control operation amount output permission/inhibition data S4 = 0 (inhibition).
The control output suppressing unit 4 determines whether or not the control operation amount output SO is supplied to the control target plant 1, according to the control operation amount output permission/inhibition data S4 as the determination result of the control output determining unit 5. The control operation amount output permission/inhibition data S4 includes the #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 according to the following formula: IF (control operation amount output permission/inhibition data S4 =

- 32 -
0) THEN
#1 to #n AS-U position change amount output = 0
Upper first intermediate shift position change amount output = 0
Lower first intermediate shift position change amount output = 0
ELSE
#1~#n AS-U position change amount output = #1 to #n AS-U position change
amount
Upper first intermediate shift position change amount output = upper
first intermediate shift position change amount
Lower first intermediate shift position change amount output = lower
first intermediate shift position change amount
ENDIF.
The control executing device 20 executes the above calculation according to the performance data Si from the control target plant 1 (rolling mill) and outputs the control operation amount output SO to the control target plant 1 (rolling mill), hence to perform the shape control.
Next, the operation outline of the controlling method learning device 21 will be described. The controlling method learning device 21 uses the time lag data of the data used in the control executing device 20. The time lag Z-1 means e-TS, to be delayed by a predetermined time T. Because of having a time response, the control target plant 1 has the time lag until the control operation amount output SO changes the performance data. Therefore, learning is performed by using the performance data at the elapsed point of the delay time T, after execution of the control operation. In the shape control, since the

- 33 -
shape detector takes some seconds until detecting the shape change, after the operation command output to the AS-U and the first intermediate roll, T = 2 to 3 seconds is better (since the time lag changes depending on the type of the shape detector and the rolling speed, the optimum time taken until the change of the control operating end becomes the shape change may be set as T).
Fig. 11 shows the operation outline of the control quality determining unit 6. The shape change quality determining unit 602 uses the quality determination evaluation function JC as the following expression. [Expression 5]
JC = 1 JT(wC(i). £fb(i))2 _1 JT(wC(i) .aast (i))2
In the expression (5), the symbol sfb(i) is the shape deviation performance data included in the performance data Si, the slast(i) is the previous value of the shape deviation performance data, and the wC(i) is the sheet width direction weight coefficient for determining the quality. Here, the weight coefficient wC(i) for the quality determination is set depending on the specifications A1 and A2 about the priority of control, according to the quality determination database DB4. Quality of the control result is determined according to the quality determination evaluation function Jc. Also when the control operation amount output permission/inhibition data S4 as the determination result of the control output determining unit 5 is 0 (control output inhibition), with the control operation amount output to the control target plant 1 actually = 0, the shape is determined

- 34 -
to be worse.
Here, assume that when the control operation amount output
permission/inhibition data S4 = 9, the control result quality data S6 = -1. Further, the threshold upper limit LCU and the threshold lower
limit LCL are previously set under the threshold condition (LCU > 0 > LCL). When the comparison result with the quality determination evaluation function Jc is Jc > LCU, it is determined as the control result quality data S6 = -1 (the shape has become worse); when LCU > Jc > 0, it is determined as the control result quality data S6 = 0 (the shape changes in a direction of becoming worse); when 0 > Jc > LCL, it is determined as the control result quality data S6 = 1 (the shape changes in a direction of becoming better); and when Jc < LCL, it is determined as the control result quality data S6 = 0 (the shape has become better).
Here, the control result quality data S6 = -1 means that since the shape has become worse, the supplied control output is controlled; the control result quality data S6 = 0 means no shape change or that since the shape has become better, the supplied control output is kept; and the control result quality data S6 = 1 means that since the shape has changed in a direction of becoming better and still has a possibility of becoming better further, the supplied control amount is increased.
As mentioned above, since the weight coefficient wC(i) in the sheet width direction changes depending on the specifications A1 and A2 about the priority of the control, the quality determination evaluation function Jc varies. Therefore, the determination results of the control result quality data S6 may be various. In the controlling

- 35 -
method learning device 21, the control result quality data S6 is determined as for the two types of the specifications A1 and A2 about the priority of the control.
Next, the outline of the learning data creating unit 7 will be described. As shown in Fig. 1, the learning data creating unit 7 creates teacher data S7a for the neural network 111 used in the control rule learning unit 11, based on the determination result (control result quality data S6) from the control result quality determining unit 6, according to the control operating end operation command S2, the control operation amount S3, and the determination results (control operation amount output permission/inhibition data S4) of the control output suppressing unit.
The teacher data S7a in this case is the AS-U control degree 301 and the first intermediate operation degree 302 as the output from the output layer of the neural network 111, as shown in Fig. 4. The learning data creating unit 7 creates the teacher data S7a for the neural network 111 used in the control rule learning unit 11, by using the control operating end operation command S2 (the AS-U control degree 301 and the first intermediate operation degree 302) as the output of the neural network 101 and the #1 to #n AS-U position change amount outputs, upper first intermediate shift position change output, and lower first intermediate shift position change output as the control operation amount output SO.
For the sake of describing the operation outline of the learning data creating unit 7, a relation between each unit data and each symbol in the control output calculating unit 3 of Fig. 8 is illustrated in

- 36 -
Fig. 12. Here, the AS-U control degree 301 about the control operating
end operation command S2 as the output of the neural network 101 is
typically shown, with the data on the operation degree positive side
as OPref, the data on the operation degree negative side as OMref, the
operation degree occurring at random number from the control operation
disturbance generating unit 16 as operation degree random number Oref,
the conversion gain as G, and the control operation amount output SO
as Cref. According to this, for the sake of simplicity, as the output
from the output layer from the neural network 101 of the control rule
executing unit 10, there are shown the outputs of the operation degree
position side, the operation degree negative side, and the operation
degree random number which is the operation degree occurring at random
number from the control operation disturbance generating unit 16.
Further, the control operation amount output SO to the control operating
end is defined as an operation command value.
Fig. 13 shows the processing steps and processing contents in
the learning data creating unit 7. According to the rule of the symbols
defined in Fig. 12, the operation command value Cref is required by
the expression (6), in the initial processing step 71.
[Expression 6]
Cref = G ■ (OPref - OMref + ORref)
In the next processing step 72, the operation command value Cref is modified to Cref, according to the control result quality data S6. Specifically, the modified value Cref of the operation command value Cref is obtained by using the expression (7) in the control result quality data S6 = -1, the expression (8) in the control result quality

- 37 -
data S6 = 0, and the expression (9) in the control result quality data S6 = 1. [Expression 7]
IF Cref > 0THEN C' ref = Cref - ACref IF Cref < 0THEN C' ref = Cref + ACref
[Expression 8] C'ref = Cref
[Expression 9]
IF Cref > 0THEN C' ref = Cref + ACref IF Cref < 0THEN C' ref = Cref - ACref
IN the processing step 73, the operation degree modified amount
AOref is required according to the expressions (10) and (11) using the
modified operation command value Cref.
[Expression 10]
C' ref = G- ((OPref + AOref)- (OMref - AOref))
[Expression 11]
AOref = 1 ( C' ref _ (OPref _ OMref)) 2 G
In the processing step 74, the teacher data OP´ref and OM´ref for the neural network 111 is required according to the expression (12).
[Expression 12]
OP'ref = OPref + AOref OMref = OMref - AOref
According to this, as shown in Fig. 12, the learning data creating unit 7 calculates the operation command value Cref actually output to the control target plant 1 and requires the operation command modified

- 38 -
value C´ref, according to the control result quality data S6 as the determination result in the control result quality determining unit 6. Specifically, when the control result quality data S6 = 1, it is determined that the control direction is OK but that the control output runs short, where the operation command value is increased in the same
direction by ACref. On the contrary, when the control result quality data S6 = -1, it is determined that the control direction is wrong, where the operation command value is decreased in the contrary direction by ACref. The conversion gain G is previously determined and known; therefore, it is possible to require the modified amount AOref if the values of the operation degree positive side and the operation degree negative side are found. Here, a proper value is previously required through a simulation and set as the ACref. According to the above procedure, the teacher data OP´ref and OM´ref used in the control rule learning unit 11 are required by the above expression

claimed

1.A plant controller of performing control on a control target plant, in recognition of a combination pattern of performance data of the control target plant, comprising a controlling method learning device that learns a combination of performance data and control operation of the control target plant, and a control executing device that performs control on the control target plant depending on the learned combination of performance data and control operation,
wherein the control executing device includes a control rule executing unit that gives a control output according to a predetermined combination of performance data and control operation of the control target plant, a control output determining unit that determines permission or inhibition of the control output supplied by the control rule executing unit and notifies the controlling method learning device that the above performance data and control operation are an error, and a control output suppressing unit that inhibits supplying a control output to the control target plant when the control output determining unit determines that in being supplied to the control target plant, the control output is possible to deteriorate the performance data of the control target plant,
the controlling method learning device includes a control result quality determining unit that, when the control executing device actually supplies a control output to the control target plant, determines quality of a control result whether the performance data has become better or worse than before the control at a delayed time

until the control effect appears in the performance data, a learning data creating unit that obtains teacher data, using the quality of the control result in the control result quality determining unit and the control output, and a control rule learning unit that learns the performance data and the teacher data as learning data, and
through learning by the controlling method learning device, other combinations of performance data and control operation are obtained on a plurality of control targets according to a state of the control target plant and the obtained combinations of performance data and control operation are used as predetermined combinations of performance data and control operation of the control target plant in the control rule executing unit.
2.The plant controller according to claim 1,
wherein since the combination of performance data and control operation is changed according to size of the performance data of the control target plant, the combinations of performance data and control operation are learned for controlling by using information about the size of the performance data and information including standardized performance data for easy pattern recognition.
3. The plant controller according to claim 1 or 2,
wherein the control rule executing unit holds the predetermined combinations of performance data and control operation of the control target plant as a first neural network, the control rule learning unit holds the combinations of performance data and control operation as

a second neural network, and the second neural network obtained as a result of learning in the controlling method learning device is used as the first neural network in the control rule executing unit.
4. The plant controller according to any one of claims 1 to 3,
wherein the control executing device further includes a control
operation disturbance generating unit that gives disturbance to the control output, and the controlling method learning device performs learning even when the disturbance is applied.
5. The plant controller according to any one of claims 1 to 4,
wherein the controlling method learning device obtains a
plurality of the combinations of performance data and control operation through learning in predetermined several specifications, and the control executing device selects one combination, according to an operational state of the control target plant, of performance data and control operation, hence to give the control output.
6. The plant controller according to claim 3,
wherein the neural network of learning the combination of performance data and operation method to use is changed according to the size of the performance data.
7. The plant controller according to any one of claims 1 to 6,
wherein based on a state of the control target plant or an
experience of an operator of the control target plant, each quality

determination reference of the control result is changed, each relation between performance data and operation method about the control target plant is required and stored in a database, hence to perform control in various controlling methods according to the state of the control target plant or the experience of the operator thereof.
8. The plant controller according to any one of claims 1 to 7,
wherein before performing control in the control target plant,
the combination of performance data and control operation is created through a simulation using a control model of the control target plant, hence to shorten a learning period of the combination of performance data and control operation in the control target plant.
9. A rolling mill controller with the plant controller according
to any one of claims 1 to 8 applied,
wherein the control target plant is a rolling mill and the performance data is of a shape on an output side of the rolling mill.
10 . A plant controlling method of performing control on a control target plant, in recognition of a combination pattern of performance data of the control target plant,
wherein a controlling method learning device that learns a combination of performance data and control operation of the control target plant, and a control executing device that performs control on the control target plant depending on the learned combination of performance data and control operation are provided,

the control executing device gives a control output according to a predetermined combination of performance data and control operation of the control target plant, determines permission or inhibition of the control output, notifies the controlling method learning device that the above performance data and control operation are an error, and inhibits supplying a control output to the control target plant when determining that in being supplied to the control target plant, the control output is possible to deteriorate the performance data of the control target plant,
when the control executing device actually supplies a control output to the control target plant, the controlling method learning device determines quality of a control result whether the performance data has become better or worse than before the control at a delayed time until the control effect appears in the performance data, obtains teacher data, using the quality of the control result and the control output, learns the performance data and the teacher data as learning data, hence to obtain other combinations of performance data and control operation on a plurality of control targets according to a state of the control target plant, and uses the obtained combinations of performance data and control operation as predetermined combinations of performance data and control operation of the control target plant in the control rule executing unit.
11. A rolling mill controlling method with the plant controlling method according to claim 10 applied,
wherein the control target plant is a rolling mill and the

performance data is of a shape on an output side of the rolling mill.
12. A program for realizing a plant controller of performing control on a control target plant, in recognition of a combination pattern of performance data of the control target plant, by a computer system,
wherein the computer system includes a controlling method learning device that learns a combination of performance data and control operation of the control target plant, and a control executing device that performs control on the control target plant according to the learned combination of performance data and control operation, the program comprising:
for a purpose of achieving processing of the control executing device, a control rule executing program for giving a control output according to a predetermined combination of performance data and control operation of the control target plant; a control output determining program for determining permission or inhibition of the control output supplied by the control rule executing program and notifying the controlling method learning device that the above performance data and control operation are an error; and a control output suppressing program for inhibiting supplying a control output to the control target plant when the control output determining program determines that in being supplied to the control target plant, the control output is possible to deteriorate the performance data of the control target plant; and
for a purpose of achieving processing of the controlling method

learning device, a control result quality determining program for, when the control executing device actually supplies the control output to the control target plant, achieving control result quality determination processing of determining quality of a control result whether the performance data has become better or worse than before the control at a delayed time until the control effect appears in the performance data; a learning data creating program for obtaining teacher data, using the quality of the control result in the control result quality determining program and the control output; and a control rule learning program for learning the performance data and the teacher data as learning data, and
through learning by the controlling method learning device, other combinations of performance data and control operation are obtained on a plurality of control targets according to a state of the control target plant and the obtained combinations of performance data and control operation are used as predetermined combinations of performance data and control operation of the control target plant in the control rule executing program.