BACKGROUND OF THE INVENTION FIELD OF THE INVENTION [0001]
The present invention relates to a technique of performing feedback control of a real time using artificial intelligence such as a neural net or the like.
DESCRIPTION OF THE RELATED ART
[0002]
Plant control based on various control theories have heretofore been executed in various plants to obtain a desired control result by control thereof.
[0003]
In control of, for example, a rolling mill as an example of a plant, fuzzy control or neuro-fuzzy control has been applied as a control theory targeted for shape control to control a waved state of a plate as an example of control. The fuzzy control has been applied to shape control using a coolant, and the neuro-fuzzy control has been applied to shape control of a Sendzimir mill. Of these, the shape control to which the neuro-fuzzy control is applied has been conducted by as shown in Patent Document 1, determining a similarity rate between the difference between an actual shape pattern detected by a shape detector and a target shape pattern, and a reference

shape pattern set in advance and determining a control output quantity relative to an operation end by a control rule represented by a control operation end operation quantity relative to the reference shape pattern being also set in advance from the similarity rate. The shape control of the Sendzimir mill using the neuro-fuzzy control is assumed to be used hereinafter as a prior art.
[0004]
FIG. 1 illustrates the shape control of the Sendzimir mill described in FIG. 1 of Patent Document 1. The neuro-fuzzy control has been used in the shape control of the Sendzimir mill. In this example, a pattern recognizing mechanism 51 performs shape pattern recognition from an actual shape detected by a shape detector 52 and computes whether the actual shape is the closest to any of reference shape patterns set in advance. The control computing mechanism 53 performs control using a control rule configured of control operation end operation quantities relative to such preset shape patterns as shown in FIG. 2. Described more specifically in terms of FIG. 2, the pattern recognizing mechanism 51 computes whether a difference (As) between the shape actual result detected by the shape detector 52 and a target shape (sref) is the closet to any of shape patterns (s) of 1 to 8. The control computing mechanism 53 selects and executes any of control methods of 1 to 8.
[0005]

In the method of Patent Document 1, however, there is shown a case in which a shape change which is contrary to expectations occurs, although there is shown a case in which in order to verify the control rule, an operator is allowed to perform a manual operation during rolling, thereby carrying out verification of the control rule or the like. That is, there occurs a case where the control rule determined in the above-described manner is actually not in conformity. This is due to shortage in examinations of mechanical characteristics, an operation state of a rolling mill, and a change in mechanical condition. It is difficult to verify one by one whether the control rule set in advance is the best rule, because there are many conditions to be considered. Therefore, once the control rule is set, the control rule is often used as it is unless it is faulty.
[0006]
Since the control rule is fixed where the control rule is not in conformity in fact due to a change in operation condition or the like, it becomes difficult to improve control accuracy to some extent or more. Further, since the operator does not perform the manual operation
(disturbance occurs for control) once the shape control is operated, it is also difficult to find a new control rule by any manual intervention of the operator. Further, even when a rolling material having a new standard is rolled, it is difficult to set the control rule correspondingly to its

material.
[0007]
In the conventional shape control as described above, since the control is performed using the control rule set in advance, a problem arises in that it is difficult to correct the control rule.
[0008]
In order to solve this problem, there have been realized the following (1) and (2) such as shown in Patent Document 2 by changing the control rule at random while performing the shape control, and learning a rule in which the shape becomes good:
(1) find a new control rule while executing the shape control during the rolling, and
(2) since a new control rule cannot be predicted preliminarily, and a completely unpredictable control rule may be made optimum, the control operation end is operated at random to find a control result relative to it while looking at it.
1: Japanese Patent No. 2804161 2: Japanese Patent No. 4003733
CITATION LIST [0009]
Patent Document Patent Document
SUMMARY OF THE INVENTION
[0010]

In the above prior art, a typical shape is set as the reference shape pattern in advance, and the control is performed based on the control rule indicative of the relation of the control operation end operation quantity to the reference waveform pattern. The learning of the control rule also relates to the control operation end operation quantity relative to the reference waveform pattern. The typical reference shape pattern defined in advance is used as it is. Therefore, a problem arises in that the control becomes shape control which reacts with only a specific shape pattern. [0011]
The reference shape pattern is defined by the knowledge related to the rolling mill targeted for a human being in advance, and accumulated experiences of a shape actual result and a manual intervention operation. It is however difficult to cover all shapes generated in a rolling mill and a rolled material to be targeted. Therefore, when a shape different from the reference shape pattern occurs, the control based on the shape control is not executed, and a shape deviation remains without being suppressed. Alternatively, there is also a case where the reference shape pattern is erroneously recognized as a similar reference shape pattern, and an erroneous control operation is performed, so that the shape is deteriorated in reverse. [0012]

Therefore, the conventional shape control has a problem in that since the control rule is learnt using the preset reference shape pattern and the control rule relative thereto, and the control is performed, there is a limit to an improvement in control accuracy. [0013]
In order to solve it, it is considered that there is used a plant control apparatus recognizing combination patterns of actual result data of a control target plant to perform control on the control target plant, which includes a control method learning device which learns a combination of the actual result data of the control target plant and a control operation, and a control executing device which performs control of the control target plant according to the combination of the learnt actual result data and the control operation, characterized in that the control executing device includes a control rule executing section which applies a control output in accordance with the defined combination of the actual result data of the control target plant and the control operation, a control output determining section which determines permission/non-permission of the control output outputted from the control rule executing section and notifies the control method learning device of the actual result data and the control operation being erroneous, and a control output suppressing section which, when the control output determining section has determined the actual result data of the control target

plant to be deteriorated where the control output is output to the control target plant, prevents the control output from being output to the control target plant, and in that the control method learning device includes a control result good/defect determining section which, when the control executing device actually outputs the control output to the control target plant, makes a determination as to a non-defect/defect of a control result about whether the actual result data is improved or deteriorated compared with that before the corresponding control after a time delay until a control effect appears in the actual result data, a learning data creating section which obtains teacher data by using the non-defect/detect of the control result in the control result good/defect determining section and the control output, and a control rule learning section which learns the actual result data and the teacher data as learning data, and characterized in that with the learning by the control method learning device, a combination of separate actual result data and a control operation to a plurality of controlled targets is obtained according to the state of the control target plant, and the so-obtained combination of the actual result data and the control operation is used as a defined combination of actual result data of the control target plant and a control operation in the control rule executing section. [0014]
At this time, it is very important that an evaluation

function used for the good/defect determination of the control result is appropriate. However, upon determining the evaluation function, a designer of the control apparatus determines the evaluation function subjectively while performing hearing investigations to an operation engineer, an operator, etc. of the control target plant and confirming the operation of an actual plant. There are many cases where it is unknown whether the evaluation function is set truly and appropriately.
[0015]
Consider shape control of a rolling mill as an example. It is ideal that in the shape control of the rolling mill, a target shape and an actual shape coincide with each other over the whole in a plate width direction. There are however many cases in which they do not coincide with each other in fact. Therefore, it is general that at the actual work, a specific area of a plate is emphasized, and the actual shape and the target shape are controlled to coincide in the area. As an evaluation function to evaluate the shape of the plate, there is used an evaluation function in which each part in the plate width direction is weighted in correspondence with a shape deviation (=shape actual result - target shape) at each part in the plate width direction.
[0016]
In the rolling mill, a control operation end relative to the shape of an end (plate end) in the plate width

direction is separated from a control operation end relative to a part (center part) excepting that. There are however many cases where they influence each other. Further, the shape deviation often becomes large because the plate end is not restrained from both sides as in the center part. When control is applied to the plate end in the plate width direction, its influence is exerted on the center part, so that the shape of the center part is deteriorated or its reverse case occurs. Thus, it is difficult to control the shapes of the plate end and the center part to match with a target value simultaneously. In many cases, the operator gives precedence to either the plate end or the center part and executes manual control thereof. [0017]
When the evaluation function applied upon the good/defect determination as to the control result perofms evaluation different from the idea of the operator, the operator cancels the operation from the shape control by the control apparatus and performs a manual operation in accordance with his/her own idea. In that case, the shape control by the control apparatus and the manual operation done by the operator are in a competing state. As a result, it is also considered that the operator makes OFF the shape control from the control apparatus, which disturbs the manual operation by the operator. There is a fear that when it is repeated, the operator does not make

ON the shape control by the control apparatus from the beginning.
[0018]
If the evaluation function applied to the good/defect determination as to the control result is set to perform evaluation matched with the idea of the operator, it is possible not only to reduce competition of the control by the control apparatus with the manual operation of the operator but also to reduce execution of the manual operation by the operator. A load imposed on the operator is reduced, and an improvement in the accuracy of the shape control is also expected.
[0019]
An object of the present invention is to provide a technique which makes it possible to execute control based on an appropriate good/defect determination as to a control result.
[0020]
A control apparatus of the present disclosure is a control apparatus which controls a controlled target. The control apparatus includes a control executing device which applies a control output to the controlled target in accordance with a control rule given thereto, a control method learning device which evaluates the control output applied to the controlled target by using an indicated evaluation function, creates learning data by utilizing an evaluation result thereof, and learning the learning data

to thereby construct the control rule, and applies the control rule to the control executing device, and an evaluation function setting section which holds a plurality of evaluation functions in advance, selects any of the plural evaluation functions, based on a control state to the controlled target, and specifies the selected evaluation function to the control method learning device. [0021]
According to the present disclosure, it is expected that control based on an appropriate good/defect determination as to a control result can be executed.
BRIEF DESCRIPTION OF THE DRAWINGS [0022]
FIG. 1 is a diagram showing shape control of a Sendzimir mill described in FIG. 1 of Patent Document 1.
FIG. 2 is a diagram showing a control rule configured by a control operation end operation quantity for each shape pattern.
FIG. 3 is a diagram showing an outline of a plant control apparatus according to an embodiment.
FIG. 4 is a diagram showing a specific example of a control rule executing section 10 according to the embodiment.
FIG. 5 is a diagram showing a specific example of a control rule learning section 11 according to the embodiment.

FIG. 6 is a block diagram showing an internal configuration of an evaluation function setting section 17.
FIG. 7 is a diagram showing a configuration of a neural network when used for shape control of the Sendzimir mill.
FIG. 8 is a diagram for describing a shape deviation and a control method.
FIG. 9 is a diagram showing an outline of a control input data creating section 2.
FIG. 10 is a diagram showing an outline of a control output computing section 3.
FIG. 11 is a diagram showing an example of transition of a rolling speed in the rolling mill.
FIG. 12 is a diagram showing an example of an evaluation function DB DB5.
FIG. 13 is a diagram for describing an operation outline of an evaluation function selection method learning unit 173.
FIG. 14 is a diagram for describing an operation outline of an evaluation function learning unit 174.
FIG. 15 is a diagram for describing an outline configuration of the evaluation function learning unit 174.
FIG. 16 is a diagram for describing an outline of a control output determining section 5.
FIG. 17 is a diagram for describing an operation outline of a control result good/defect determining section 6.

FIG. 18 is a diagram for describing an operation outline of a learning data generating section 7.
FIG. 19 is a diagram showing a processing stage and a processing content in the learning data generating section 7.
FIG. 20 is a diagram showing a data example stored in a learning data database DB2.
FIG. 21 is a diagram showing an example of a neutral net management table TB.
FIG. 22 is a diagram showing an example of the learning data database DB2.
DESCRIPTION OF THE EMBODIMENT [0023]
The findings in the present invention, and the processes leading to the present invention will first be described by taking for example shape control of a rolling mill. [0024]
First, the following three are required to solve the above problems. [0025]
(1) A reference shape pattern and a control operation corresponding thereto are separately set in advance, the combination of the shape pattern and the control operation is learnt without learning a control operation method, and the control operation is performed using the same.

[0026]
(2) Since a new control rule cannot be predicted
preliminarily, and a completely unpredictable control rule
may be made optimum, the control operation end is operated
at random to find a control result relative to it while
looking at it.
[0027]
(3) An evaluation function is selected according to
the state of the rolling mill in terms of non-defect/defect
of the control result, and the selection of a suitable
control rule is made possible.
[0028]
In order to realize these three, the control operation may preferably be changed such that the control result is improved while changing the combination of the shape pattern and the control operation, which is used for the shape control. To this end, the combination of the shape pattern and the control operation suitable for the shape pattern is learnt by artificial intelligence such as a neural network or the like, and the output of the control operation relative to the shape pattern generated in the rolling mill may preferably be changed.
[0029]
When the control operation is changed while executing the shape control on the rolling mill during the operation, an erroneous control output is outputted so that the shape of the plate is deteriorated. Thus, an operation

abnormality such as plate fracture may occur. When the plate fracture occurs, it takes time to replace rolls used in the rolling mill, and a material to be rolled during rolling is wasted, thereby causing large damage. It is therefore necessary to prevent the erroneous control output from being output to the rolling mill as much as possible. Therefore, the evaluation function for determining the non-defect/defect of the shape may be changed according to the rolling state.
[0030]
The rolling state refers to a state related to rolling under which the rolling mill to be controlled is placed. If the controlled target is not limited to the rolling mill, the rolling state is generalized and can be referred to a control state. The rolling state can be discriminated by various parameters such as the control operation applied to the rolling mill, the state of the rolling mill, the state of rolling by the rolling mill, etc. In the present embodiment, as an example, the rolling state can be discriminated by a rolling speed.
[0031]
From the above, in the present embodiment, in order to realize the above, the non-defect/defect of the control operation output from the neural net is verified using, for example, a simplified model of the rolling mill. An output in which the shape is considered to be deteriorated clearly is avoided from being output to the control operation end

of the rolling mill, thereby preventing shape deterioration. At this time, in terms of the neural net, learning is executed assuming that the control operation relative to its shape pattern is incorrect. [0032]
Since there is a possibility that the verifying method itself of the non-defect/defect of the control operation is incorrect, a control operation output of a neural net determined to be erroneous with certain probability is also output to the control operation end of the rolling mill, thereby making it possible to learn even the combination of an unanticipated shape pattern and a control operation. [0033]
Embodiments of the present invention will hereinafter be described in detail using the drawings. [0034]
FIG. 3 shows an outline of a plant control apparatus according to an embodiment. The plant control apparatus of FIG. 3 is constituted of a control target plant 1, a control executing device 20 which inputs actual result data Si from the control target plant 1 therein and provides a control operation quantity output SO defined according to a control rule (neural net) such as illustrated in FIG. 2 to the control target plant 1 to control the control target plant 1, a control method learning device 21 which inputs the actual result data Si or the like from the control target plant 1 therein to perform learning and reflects the

learned control rule to the control rule in the control executing device 20, a plurality of databases DB (DB1 to DB3), and a management table TB for the databases DB.
[0035]
The control executing device 20 is comprised of as main elements, a control input data creating section 2, a control rule executing section 10, a control output computing section 3, a control output suppressing section 4, a control output determining section 5, and a control operation disturbance generating section 16.
[0036]
In the control executing device 20, the control input data creating section 2 is first used to create input data SI of the control rule executing section 10 from the actual result data Si of a rolling mill being the control target plant 1. The control rule executing section 10 creates a control operation end operation command S2 from the actual result data Si of the target to be controlled, by using the neural net (control rule) representing the relation between the actual result data Si of the target to be controlled and the control operation end operation command S2. The control output computing section 3 computes a control operation quantity S3 to the control operation end, based on the control operation end operation command S2. Thus, the control operation quantity S3 is created using the neural net according to the actual result data Si of the control target plant 1.

[0037]
Further, the control output determining section 5 in the control executing device 20 determines control operation quantity output permission/non-permission data S4 to the control operation end by using the actual result data Si from the control target plant 1 and the control operation quantity S3 from the control output computing section 3. The control output suppressing section 4 determines output permission/non-permission of the control operation quantity S3 to the control operation end according to the control operation quantity output permission/non-permission data S4 and outputs the permitted control operation quantity S3 as the control operation quantity output SO to be provided to the control target plant 1. Thus, the control operation quantity S3 determined to be abnormal is not output to the control target plant 1. Incidentally, the control operation disturbance generating section 16 generates disturbance and provides the same to the control target plant 1 for the purpose of verifying the plant control apparatus. [0038]
The control executing device 20 configured as described above further refers to a control rule database DB1 and an output determination database DB3 in order to execute its processing as will be described later. The control rule database DB1 is accessibly connected to both the control rule executing section 10 in the control

executing device 20 and the control rule learning section 11 in the control method learning device 21. A control rule (neural net) as a learning result in the control rule learning section 11 is stored in the control rule database DBl. The control rule executing section 10 refers to the control rule stored in the control rule database DBl. The output determination database DB3 is accessibly connected to the control output determining section 5 in the control executing device 20. [0039]
FIG. 4 shows a specific example of the control rule executing section 10 according to the present embodiment. The control rule executing section 10 inputs the input data SI created in the control input data creating section 2 therein and provides the control operation end operation command S2 to the control output computing section 3. The control rule executing section 10 is provided with a neural net 101. In the neutral net 101, the control operation end operation command S2 is basically defined by the method of Patent Document 1 such as illustrated in FIG. 2. In the present invention, the control rule executing section 10 is further provided with a neural net selecting unit 102, which selects, as the control rule in the neutral net 101, the optimum control rule by referring to the control rules stored in the control rule database DBl and causes the same to be executed. Thus, the control rule executing section 10 of FIG. 4 selects a necessary neural net from a

plurality of neutral nets divided for purposes of an operator group and control and uses it. The control rule database DBl may preferably include even such actual result data (data of operation group or the like) Si that a neural net and a good/defect determination reference can be selected, as data from the control target plant 1. Incidentally, since the neural net has a relation that it serves as the control rule when it is executed, in the present specification, the neural net and the control rule are not distinguished from each other and are used in the same sense. [0040]
Referring back to FIG. 3, the control method learning device 21 executes learning of the neural net 101 used in the control executing device 20. When the control executing device 20 outputs the control operation quantity output SO to the control target plant 1, it takes time for a control effect to actually appear as a change in the actual result data Si. Therefore, the learning is executed using data time-delayed by that time. In FIG. 3, Z_1 indicates an appropriate time delay function for each data. [0041]
The control method learning device 21 is comprised of as main elements, a control result good/defect determining section 6, a learning data creating section 7, a control rule learning section 11, and an evaluation function setting section 17.

[0042]
Of these, the control result good/defect determining section 6 determines using the actual result data Si and actual result data previous value SiO from the control target plant 1, and the evaluation function set by the evaluation function setting section 17 whether the actual result data Si changes in a direction to be improved or in a direction to be deteriorated, and outputs a control result good/defect data S6.
[0043]
The learning data creating section 7 in the control method learning device 21 creates novel teacher data S7a used for the learning of the neural net by using the data obtained by time-delaying the input data such as the control operation end operation command S2, the control operation quantity S3, the control operation quality output permission/non-permission data S4, etc. created in the control executing device 20 by the same time respectively, and the control result good/defect data S6 from the control result good/defect determining section 6, and provides the same to the control rule learning section 11. Incidentally, the teacher data S7a corresponds to the control operation end operation command S2 output from the control rule executing section 10. It can be said that the learning data creating section 7 has determined, as the novel teacher data S7a, the data obtained by estimating the control operation end operation command S2 output from the

control rule executing section 10 by using the control result good/defect data S6 given from the control result good/defect determining section 6. [0044]
FIG. 5 illustrates a specific example of the control rule learning section 11 according to the present embodiment. The control rule learning section 11 is comprised of as main constituent elements, an input data creating unit 114, a teacher data creating unit 115, a neural net processing unit 110, and a neural net selecting unit 113. Further, the control rule learning section 11 obtains, as external inputs, data S8a obtained by time-delaying the input data SI from the input data creating section 2 and novel teacher data S7a from the learning data creating section 7, and refers to the data stored in the control rule database DB1 and the learning data database DB3. [0045]
In the control rule learning section 11, the input data SI is subjected to the appropriate time-delay compensation and then taken in the neural net processing unit 110 via the input data creating unit 114. [0046]
Further, in the control rule learning section 11, the novel teacher data S7a from the learning data creating section 7 is provided via the teacher data creating unit 115 to the neural net processing unit 110 as total teacher

data S7c including even past teacher data S7b stored in the learning data database DB2 . These teacher data S7a and S7b are stored in the learning data database DB2 and utilized as appropriate. [0047]
Likewise, the input data S8a from the control input data creating section 2 is provided via the input data creating unit 114 to the neural net processing unit 110 as total input data S8c including even past input data S8b stored in the learning data database DB2. These input data S8a and S8b are stored in the learning data database DB2 and utilized as appropriate. [0048]
The neural net processing unit 110 is comprised of a neural net 111 and a neural net learning control part 112. The neural net 111 takes in the input data S8c from the input data creating unit 114, the teacher data S7c from the teacher data creating unit 115, and the control rule (neural net) selected by the neural net selecting unit 113 and stores the finally-determined neural net in the control rule database DB1. [0049]
The neural net learning control part 112 controls the input data creating unit 114, the teacher data creating unit 115, and the neural net selecting unit 113 at appropriate timings and controls them to obtain the input of the neural net 111 and store the processing result in

the control rule database DB1.
[0050]
Here, the neural net 101 in the control executing device 20 of FIG. 4, and the neural net 111 in the control method learning device 21 of FIG. 5 are both neural nets having the same concept. A description will hereinafter be made as to the difference in basic concept where they are utilized.
[0051]
The neural net 101 in the control executing device 20 is a neural net of a predetermined content and is for obtaining the control operation end operation command S2 as the corresponding output when the input data SI is provided, a so-called neural net used for processing in one direction. On the other hand, the neural net 111 in the control method learning device 21 is for, when the input data S8c and teacher data S7c about the input data SI and the control operation end operation command S2 are set as the learning data, determining a neural net satisfying this input/output relation by learning.
[0052]
The basic idea of processing in the control method learning device 21 configured as described above is as follows: First, when the content of the control operation quantity output permission/non-permission data S4 is "permitted", the control operation quantity output SO is output to the control target plant 1. When the content of

the control result good/defect data S6 is "good" (the actual result data Si changes in the direction to be improved, the control operation end operation command S2 output from the control rule executing section 10 is determined to be correct, and the learning data is created such that the output of the neural net reaches the control operation end operation command S2. [0053]
On the other hand, when the content of the control operation quantity output permission/non-permission data S4 is "non-permitted", or the control operation quantity output SO is output to the control target plant 1 and the content of the control result good/defect data S6 is "defective" (the actual result data Si changes in the direction to be deteriorated), the control operation end operation command S2 output by the control rule executing section 10 is determined to be erroneous, and the learning data is created such that the output of the neural net is not provided. At this time, the neural net output is configured in such a manner that two types of outputs in + and - directions are output to the same control operation end as control outputs, and the learning data is created such that the control operation end operation command S2 on its output side is not output. [0054]
Further, in the control rule learning section 11 illustrated in FIG. 5, the processing is performed in the

following manner as a result of data processing by the neural net learning control part 112. Here, the learning of the neural net 101 used in the control rule executing section 10 is first carried out by using learning data being the combination of S8c obtained by time-delaying the input data SI to the control executing device 20, and the teacher data S7c created in the teacher data creating unit 115. Actually, the same neural net 111 as the neural net 101 of the control rule executing section 10 is provided in the control rule learning section 11 and is thereby adapted for performing an operation test on various conditions to learn a response at that time and then obtain a control rule under which it is confirmed that a better result occurs as its learning result. Since it is necessary to allow the learning to be conducted using a plurality of learning data, a plurality of past learning data are taken out from the learning data database DB2 having stored the learning data created in the past and learned to execute their processing. Along with it, the present learning data is stored in the learning data database DB2. Further, the learned neural net is stored in the control rule database DB1 to use it in the control rule executing section 10. [0055]
The neural net may be learned using the past learning data together each time the new learning data is created. After the learning data is stored to some extent (100 pieces, for example), the neural net may be learned using

the past learning data together. [0056]
Further, the control result good/defect determining section 6 performs a good/defect determination by using the evaluation function set from the evaluation function setting section 17. The good/defect determination of the control result differs in determination result according to the evaluation function to be used. Therefore, neural nets corresponding to a plurality of evaluation functions are created respectively. As for the same input data, teacher data are created by the evaluation functions respectively and then learned. Thus, a plurality of teacher data are created with respect to one input data and used for learning of the neural nets corresponding to the respective teacher data. At the same time, it is possible to learn the neural nets corresponding to the plural evaluation functions. Here, the plural evaluation functions are evaluation functions used for respective policies such as in the case of shape control, for example, which part (plate end, center part, asymmetrical part, etc.) is desired to be preferentially controlled in a plate width direction or which of a plurality of items to be controlled (e.g., plate thickness and tension, rolling load, etc.) is desired to be preferentially controlled, etc. [0057]
When the present embodiment is applied, it is considered that once the neutral net 101 used in the

control rule executing section 10 is learned, a new control operation is not executed. Therefore, a new operation method is appropriately generated at random by the control operation disturbance generating section 16 to execute a control operation in addition to the control operation quantity S3, whereby a new control method is learned.
[0058]
Hereinafter, as one example, the details of the present plant control method will be described with the shape control in such a Sendzimir mill as shown in Patent Document 1 as a target. Incidentally, the shape control will be described with the following specifications A and B being adopted.
[0059]
The specification A is a specification about an evaluation function and is assumed to have information about the priority in the plate width direction. For example, in the shape control, there are many cases in which it is difficult to control the plate thickness or the like to a target value over the whole region in the plate width direction in terms of mechanical characteristics. Therefore, there are provided in the plate width direction, evaluation functions Al through AN (N is the maximum number of evaluation functions to be set) corresponding to the following plural policies.
[0060]
The evaluation function is defined such that its value

becomes small as the evaluation becomes good. There are for example, the square mean of a control deviation, the maximum value - minimum value, etc. [0061]
Here, six types of policies and evaluation functions Al to A6 to be illustrated below are assumed to be used by way of example.

n
JAI(E(.0) = — / (WC(0 " £(0)2 edge part wc(V) = 3.0, center part wc(V) = 1.0
i=l

n
/42(£(0) = — / (wc(0 " £(0)2 edge part wc(V) = 1.0, center part wc(i) = 3.0
i=l

n
i=l Edge part s(i) = s(i): if (s(i) < 0),0: £/(e(£)>0)

n
JA4(e(Q) = ^Oc(i) ■ e(i))2, wc(i) = 1.0
i=l
Edge part s(i) = s(i):if(s(i)>O),0:if(s(i)

n
hMO) = ^X(wc(0' £«)2.wc(i) = 1.0
i=l
Edge part s(i) = 0: if(UL > s(i) > LL), elses(i) = s(i)

/46(£(t)) = max(e(i)~) — min(s(i))
[0062]
FIG. 6 is a block diagram illustrating an internal configuration of the evaluation function setting section 17. The evaluation function setting section 17 has an evaluation function manual setting unit 171, an evaluation function selecting unit 172, an evaluation function selection method learning unit 173, and an evaluation function learning unit 174. The evaluation function setting section 17 executes the following processing about the evaluation function in conjunction with the evaluation function DB DB5. [0063]
The evaluation function manual setting unit 171 sets an evaluation function. This is processing for mathematizing and setting in advance ways of thinking of shapes of an operation engineer and an operator.
The evaluation function selecting unit 172 selects an evaluation function used in the control executing device 20 according to a rolling state.

The evaluation function selection method learning unit 173 performs learning in such a manner that an evaluation function corresponding to the rolling state is selected according to the rolling state and a manual operation actual result of the operator.

Since the evaluation set in advance by manual operation may not necessarily be correct, the evaluation function learning unit 174 learns the evaluation function itself. Here, the evaluation function to be learnt is referred to as a learning evaluation function. As the learning progresses to some extent, the evaluation is made possible using the learning evaluation function. In that case, the learning evaluation function may be used for evaluation as the evaluation function. [0064]
The specification B is a specification about correspondence to a previously known condition. To take an example, it is considered that since the relation between the shape pattern and the control method changes depending on various conditions, there is a need to divide the specification Bl and the specification B2 into the plate width and steel grade according to their categories, for example. With the above respective changes, the degree of influence on the shape of the shape operation end changes.

[0065]
In this example, the control target plant 1 is a Sendzimir mill, and the actual result data becomes a shape actual result. Incidentally, the Sendzimir mill is a rolling mill having a cluster roll for cold-rolling a hard material such as stainless steel or the like. In the Sendzimir mill, a small-sized work roll is used for the purpose of applying high draft rolling to the hard material. It is therefore difficult to obtain flat stainless plate. As countermeasures against it, there have been adopted the structure of the cluster roll and various shape control sections. The Sendzimir mill generally includes upper and lower first intermediate rolls each having one-sided taper and capable of shifting and is additionally provided vertically with six division rolls and two rolls called AS-U. In examples to be described below, detection data of a shape detector is used as the actual result data Si of the shape, and further a shape deviation being a difference from a target shape is used as the input data SI. Further, the control operation quantity S3 is assumed to be AS-U of #1 to #n and the amounts of roll shift of the upper and lower first intermediate rolls.
[0066]
FIG. 7 shows a configuration of a neural network when used for shape control of the Sendzimir mill. The neural network may be abbreviated as a neural net. Here, the neural net indicates the neural net 101 for the control

rule executing section 10 and indicates the neural net shown by the neural net 111 for the control rule learning section 11, but both are the same in structure.
[0067]
In the example of the shape control of the Sendzimir mill shown in the present embodiment, the actual result data Si from the control target plant 1 is actual result data of the Sendzimir mill including the data (here, the shape deviation being the difference between the actual result shape and the target shape is assumed to be output) of the shape detector. The control input data creating section 2 obtains as the input data SI, a standardized shape deviation 201 and a shape deviation stage 202. Thus, each of input layers of the neural nets 101 and 111 is comprised of the standardized shape deviation 201 and the shape deviation stage 202. Incidentally, though the shape deviation stage 202 is set as the input to the neural net input layer in FIG. 7, the neural net may be switched according to the stage.
[0068]
Further, an output layer is comprised of an AS-U operation degree 301 and a first intermediate operation degree 302 in accordance with AS-U and the first intermediate roll each being the shape control operation end of the Sendzimir mill. In terms of AS-U, its operation degree has an AS-U open direction (direction in which a roll gap (interval between vertical work rolls of the mill)

is opened), and an AS-U close direction (direction in which the roll gap is closed) for each AS-U. Further, in terms of the first intermediate roll, its operation degree has a first intermediate roll open direction (direction in which the first intermediate roll is operated from the center side of the mill to the outside), and a first intermediate roll close direction (direction in which the first intermediate roll is operated toward the center side of the mill) for each vertical first intermediate roll. For example, when the shape detector includes 20 zones, and the shape deviation stage 202 is set to three stages (large, medium and small), the input layer becomes 23 inputs. Further, when the AS-U has seven saddles, and the vertical first intermediate rolls are assumed to be capable of shifting in the plate width direction, the output layer includes fourteen AS-U operation degrees 301 and four first intermediate operation degrees, which become 18 in total. The layer number of intermediate layers and the neuron number of each layer are set as appropriate. Incidentally, although being described later with reference to FIG. 10, in terms of the shape control operation end of the Sendzimir mill being the output layer, the neural net output is configured such that two types of outputs in + and - directions are output to each individual control operation end. [0069]
FIG. 8 shows the shape deviation and the control

method. Here, there is shown in the upper part of FIG. 8, a control method where the shape deviation is large. There is shown in the lower part of FIG. 8, a control method where the shape deviation is small. Incidentally, the vertical direction indicates the magnitude of the shape deviation, the horizontal axis direction indicates the plate width direction, both sides of the plate width indicate a plate end, and the center indicates a plate central part. As indicated in the upper part of FIG. 8, when the shape deviation is large, the whole shape is modified more preferentially than the local shape deviation in the plate width direction. On the other hand, as indicated in the lower part of FIG. 8, when the shape deviation is small, the local shape deviation is preferentially reduced. [0070]
Thus, since it is necessary to change the control method according to the magnitude of the shape deviation, as shown in FIG. 7, the shape deviation stage 202 is provided and applied to the neural nets 101 and 111 to determine the magnitude of the shape deviation. In terms of the shape deviation, for example, those standardized into 0 to 1 may be used regardless of the magnitude of the shape deviation. This is an example and is also considered to be input to the input layer of the neural net as it is without standardizing the shape deviation. It is also considered that the neural net itself is changed according

to the magnitude of the shape deviation (e.g., two neural nets are provided and divided into a neural net used where the shape deviation is large, and a neural net used where the shape deviation is small).
[0071]
The neural nets 101 and 111 each having such a configuration as shown in FIG. 7 described above are allowed to learn the operation method for the shape pattern, and the shape control is executed using the neutral nets subjected to the learning. Even the neural nets having the same configuration differ in characteristic depending on a learning condition, and is capable of outputting control outputs different relative to the same shape pattern.
[0072]
Therefore, the optimum control can be configured relative to various conditions by properly using the plural neural nets depending on the other conditions of the shape actual results. This corresponds to the specification B. The previously-described configuration of FIG. 4 indicates a specific example where such a specification is made. In the example of the configuration in FIG. 4, as the neural net 101 used in the control rule executing section 10, another neural net is provided according to rolling actual results, a rolling-mill operator name, a steel grade of a material to be rolled, a plate width thereof, etc., and is registered in the control rule database DB1. The neural

net selecting unit 102 selects a neural net matched to the condition at that time and sets the same to the neural net 101 of the control rule executing section 10. Incidentally, as the condition at that time in the neural net selecting unit 102, data about the plate width is taken from within the actual result data Si in the control target plant 1, and a neural net is preferably selected according to the data. Further, if the plural neural nets used herein have such input and output layers as illustrated in FIG. 7, the layer number of intermediate layers and the unit number of each layer may differ. [0073]
FIG. 9 shows an outline of the control input data creating section 2 which creates the data SI (standardized shape deviation 201 and shape deviation stage 202) to be input to the input layer of each of the neural nets 101 and 111. Here, with shape detector data of the shape detector which detects a plate shape at rolling in the Sendzimir mill being the control target plant 1 as the actual result data Si being taken as the input, a shape deviation PP value (Peak To Peak value) Spp being the difference between the maximum and minimum values of the detection result of each shape detector zone is first determined by a shape deviation PP value computing unit 210. A shape deviation stage computing unit 211 classifies the shape deviation into three stages of large, medium and small according to the shape deviation PP value Spp. The shape is a

distribution in the plate width direction, of an extending ratio of the rolled material, and I-UNIT indicating the extending ratio in 10-5 unit is used as the unit. For example, the shape deviation is classified like the following equations. [0074]
Here, the shape deviation is classified in such a manner that the shape deviation stage is taken to be (large = 1, medium = 0, and small = 0) by the establishment of the (1) equation, the shape deviation stage is taken to be (large = 0, medium = 1, and small = 0) by the establishment of the (2) equation, and the shape deviation stage is taken to be (large = 0, medium = 0, and small = 1) by the establishment of the (3) equation. Incidentally, here, the shape deviation of each zone is standardized using SPM with SPM=SPP .
SPP> 501-UNIT ••■ (1)
50/ -UNIT >SPP> 10/ -UNIT ••• (2)
10I-UNIT>SPP ••• (3)
[0075]
The standardized shape deviation 201 and the shape deviation stage 202 being the input data to the neural net 101 are created in the above-described manner. The standardized shape deviation 201 and the shape deviation

stage 202 are the input data SI of the control rule
executing section 10.
[0076]
FIG. 10 illustrates an outline of the control output computing section 3. The control output computing section 3 creates a control operation quantity S3 being an operation command to each shape control operation end according to the control operation 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) in the control rule executing section 10 being the output from the neural net 101. Incidentally, here, each one data example is shown for the AS-U operation degree 301 and the first intermediate operation degree 302 which exist in plural form, and each data is comprised of a pair of data of an open direction degree and a close direction degree. [0077]
Since the input AS-U operation degree 301 has outputs in the respective AS-U open and close directions, the control output computing section 3 multiplies the difference therebetween by a conversion gain GASU therein to thereby output an operation command to each AS-U. Since the control output to each AS-U becomes an AS-U position change quantity (whose unit is the length), the conversion gain GASU becomes a conversion gain from the degree to the position change quantity.

[0078]
Further, since the first intermediate operation degree 302 input in the same manner has outputs on the first intermediate outside and inside, the difference between the outputs is multiplied by a conversion gain GIST to thereby output an operation command to each first intermediate roll shift. Since the control output to each first intermediate roll becomes a first intermediate roll shift position change quantity (whose unit is the length), the conversion gain GIST becomes a conversion gain from the degree to the position change quantity. [0079]
Thus, the control operation quantity S3 can be computed. The control operation quantity S3 is comprised of #1 to #n AS-U position change quantities (n is based on the saddle number of AS-U roll), an upper first intermediate shift position change quantity, and a lower first intermediate shift position change quantity. Incidentally, FIG. 10 illustrates a system in which disturbance data from the control operation disturbance generating section 16 is added to the control operation end operation command S2. [0080]
An operation outline of the evaluation function setting section 17 will be described with reference to FIG. 6. The evaluation function is one in which the intention of the operator relative to the control of the shape in the

rolling mill is reflected. The intention of the operator changes according to the rolling state. Here, the rolling state is assumed to be distinguished by the rolling speed. As illustrated in FIG. 11, the rolling speed of the rolling mill changes in such a manner that the rolling speed is increased from its stop state to perform rolling at a constant speed, and the rolling speed is reduced to stop rolling. The rolling state also changes like 17-1, 17-2, 17-3 ... according to a change in the rolling speed. Then, the intention of the operator also changes like an intention 1, an intention 2, an intention 3, ... according to a change in the rolling state. The intention of the operator may include those shown below, for example.
[0081]
At the start of rolling at a low speed, a central part of a plate is given priority to ensure the stability of its passage.
When the rolling is accelerated, an emphasis is placed on a plate end to prevent meandering of the plate or the like.
When the rolling speed is constant, the quality of a material to be rolled is taken into consideration, and a shape deviation of the plate end in its extending direction is allowed and the shape of the central part is given priority in such a manner that plate fracture is not generated.
[0082]

When the above respective intentions are associated with evaluation functions Al to AN, the following are brought about:
The evaluation function A2 corresponds to the intention 1. The evaluation function Al corresponds to the intention 2. The evaluation function A3 corresponds to the intention 3.
[0083]
The corresponding relation between the above-described intentions of operator and the evaluation functions is stored in the evaluation function DB DB5. An example of the evaluation function DB DB5 is illustrated in FIG. 12. It is defined which one of evaluation functions Al to A6
(evaluation function NO) should be used for each intention of the operator corresponding to the above rolling state.
[0084]
Since the rolling state to which the intentions 1, 2 and 3 is applied can be distinguished by the rolling speed, it is possible to select whether to use any of the evaluation functions Al to AN according to the rolling speed. The operator or the operation engineer or the like manually sets the correspondence between the rolling speed and the evaluation speeds Al to AN to the evaluation function DB DB5 by using the evaluation function manual setting unit 171. The evaluation function selecting unit 172 selects the evaluation function corresponding to the intention of the operator corresponding to the rolling state defined by the rolling actual result Si (including

the actual result of the rolling speed) in accordance with its manual setting, and sets the same to the control output determining section 5 and the control result good/defect determining section 6.
[0085]
The manual setting for the selection of the evaluation function by the operator or the operation engineer may differ in practice due to the cases where the determination of an actual operator is not set properly, and the operator finds and uses a new determination criterion. In order to evaluate a good/defect determination as to the manual setting, the evaluation function selection method learning unit 173 determines based on the actual result data obtained by actual rolling operation whether a method for selecting the evaluation function is good or defective. Further, when it is determined that the selecting method is defective, the evaluation function selection method learning unit 173 changes the setting of the selecting method of the evaluation function in the evaluation function database DB5.
[0086]
FIG. 13 is a diagram for describing an operation outline of the evaluation function selection method learning unit 173. When the shape of the plate is determined to be defective during the rolling operation, the operator starts a manual operation and continues the manual operation until it is determined that the shape

becomes good. Thus, the intention of the operator is reflected when the operator starts the manual operation and finishes the manual operation. The evaluation function selection method learning unit 173 calculates shape evaluation results at the respective evaluation functions Al to AN based on data at that time. Comparing those shape evaluation results with each other, the evaluation function selection method learning unit 173 is capable of determining the relative quality of the evaluation function, i.e., which evaluation function is close to the intention of the operator. [0087]
Assuming it is shown that the shape is satisfactory as a shape evaluation value becomes small in value, the evaluation function in which the shape evaluation value at the time of the start of the manual operation is large and the shape evaluation value at the time of completion of the manual operation is small, can be determined to be an evaluation function suitable in its rolling state (rolling speed). [0088]
In the present embodiment, since the calculation method differs for each evaluation function as in the evaluation using the square average, the evaluation function using the maximum value or the minimum value, etc., a common index is required to be used as and then compared with an index (evaluation function good/defect

determination index) for evaluating whether the evaluation function is good or defective. Here, as an example, the evaluation function selection method learning unit 173 compares the evaluation functions by using a ratio Xi expressed in the following equation:
Ratio Xi=(a-b)/b
where a is a shape evaluation value when the manual operation is started, b is a shape evaluation value when the manual operation is ended. The evaluation function selection method learning unit 173 determines the evaluation function Ai of the evaluation functions Al to AN, at which the ratio Xi being the evaluation function good/defect determination index becomes the largest value, to be an evaluation function at which an evaluation best matched with the intention of the operator in the rolling state at that time is obtained, and selects the same as the best evaluation function. [0089]
The rolling state when the manual operation is started or ended, and the intention of the operator at that time can be determined by rolling actual results. If the evaluation function associated with the corresponding intention of the operator in the evaluation function DB DB5 is different from the best evaluation function selected herein, the evaluation function selection method learning unit 173 updates the evaluation function corresponding to the corresponding intention of the operator to the

evaluation function Ai. Then, the evaluation function selection method learning unit 173 sets the evaluation function Ai to the control output determining section 5 and the control result good/defect determining section 6 in accordance with settings after the change from the next time on.
[0090]
A graph of FIG. 13 shows the temporal transition of the shape evaluation values of the two evaluation functions AI and A2. The shape evaluation value of the evaluation function AI at the time of the start of the manual operation is A1S, and the shape evaluation value at the time of the end of the manual operation is A1E. The shape evaluation value of the evaluation function A2 at the time of the start of the manual operation is A2S, and the shape evaluation value at the time of the end of the manual operation is A2E.
[0091]
As is apparent from FIG. 13, a ratio X2=(A2S-A2E)/A2E of the evaluation function A2 is larger than a ratio Xl=(A1S-A1E)/A1E of the evaluation function AI.
[0092]
Further, the evaluation function learning unit 174 performs the learning of the evaluation function in consideration of even the possibility that the manually-set evaluation function itself will not be appropriate.
[0093]

FIG. 14 is a diagram for describing an operation outline of the evaluation function learning unit 174. The evaluation function learning unit 174 takes as inputs, a shape actual result being an actual result value of the shape of the plate, which is obtained by rolling, and a rolling actual result being a parameter value for a control operation in the rolling, installs a neural network (neural net for evaluation function) for the evaluation function, which outputs a shape evaluation value, and performs the learning of the neural net for the evaluation function by using actual result data. Incidentally, as the rolling actual result taken as the input to the neural net for the evaluation function, such a rolling actual result (e.g., rolling speed) as to influence the evaluation function may be selected. The learnt neural network can be used as the evaluation function. [0094]
As described previously, in terms of the evaluation of the shape that the operator intends, it can be interpreted that the shape of the plate is defective when the operator starts the manual operation, and the shape thereof is good when the operator finishes the manual operation. Thus, in the process of creating the plate by the rolling mill, the evaluation function learning unit 174 makes the shape evaluation value at the time of the start of the manual operation by the operator to be equal to 1 (1 indicates that the shape is defective) and makes the shape evaluation

value at the time of completion of the manual operation to be equal to 0 (0 indicates that the shape is good), and stores the same as teacher data together with a shape actual result and a rolling actual result at that time. Then, the evaluation function learning unit 174 performs supervised learning of the neural net by using the stored teacher data. Thus, since the shape evaluation value is output when the rolling actual result and the shape actual result are input, the learned neural net can be used as the evaluation function. [0095]
An outline configuration of the evaluation function learning unit 174 is shown in FIG. 15. The control output determining section 5 and the control result good/defect determining section 6 initially uses the evaluation function set manually by the operator. The evaluation function learning unit 174 adds teacher data to be described later to the rolling actual result data Si including the shape actual result and the rolling actual result and learns its result as learning data to thereby construct a neural net for the evaluation function, which provides an evaluation function instead of the initial evaluation function. [0096]
The evaluation function learning unit 174 has an evaluation executing part and a learning execution part. [0097]

The evaluation executing part includes a neural net 1740 for the evaluation function, which is used in the control output determining section 5 and the control result good/defect determining section 6, and performs the evaluation by using the neural net 1740 for the evaluation function. [0098]
The learning executing part has a neural net 1741 for an evaluation function equivalent to the neural net 1740 for the evaluation function and performs learning by using the neural net 1741 for the evaluation function. Here, as illustrated in FIG. 14, the neural net 1741 for the evaluation function is a neural net which takes as inputs, a shape actual result and a rolling actual result and outputs a shape evaluation value. Upon the learning of the neural net 1741 for the evaluation function, rolling actual result data Si including the shape actual result and the rolling actual result is taken as input data, a shape evaluation value to be described later is taken as teacher data, and a combination of those is taken as learning data. Thus, the combination of the shape actual result and the rolling actual result, and the teacher data is stored in the evaluation function learning data database 1743 as learning data. The learning executing part may execute learning of the neural net in a stage in which learning data has been accumulated to some extent. [0099]

In addition to the above-described neural net 1741 for the evaluation function, the learning executing part has a neural net learning control portion 1744 for the evaluation function, an input data creating portion 1745, and a teacher data creating portion 1746. [0100]
The teacher data creating portion 1746 creates teacher data of the shape evaluation value=l in timing at which the manual operation is started, by using a manual operation signal relative to the shape of the operator. Further, the teacher data creating portion 1746 notifies the timing at which the manual operation is started, to the input data creating portion 1745. The input data creating portion
1745 obtains a shape actual result and a rolling actual result in the timing at which the manual operation is started, and takes the same as input data. The input data created by the input data creating portion 1745 and the teacher data created by the teacher data creating portion
1746 are stored in the evaluation function learning data database 1743 as a pair of learning data.
[0101]
Similarly, the teacher data creating portion 1746 creates teacher data of the shape evaluation value=0 in timing at which the manual operation is ended, by using the manual operation signal relative to the shape of the operator. Further, the teacher data creating portion 1746 notifies the timing at which the manual operation is ended,

to the input data creating portion 1745. The input data creating portion 1745 obtains a shape actual result and a rolling actual result in the timing at which the manual operation is ended, and takes the same as input data. The input data created by the input data creating portion 1745 and the teacher data created by the teacher data creating portion 1746 are stored in the evaluation function learning data database 1743 as a pair of learning data.
[0102]
When the learning data is accumulated in the evaluation function learning data database 1743 to some extent (e.g., 1,000 sets), the neutral net learning control portion 1744 for the evaluation function reads the learning data from the evaluation function learning data database 1743, acquires input data and teacher data from the teacher data, and applies the same to the neutral net 1741 for the evaluation function to execute the learning of the neural net.
[0103]
When the learning of the neural net 1741 for the evaluation function is completed in the learning executing part, the neural net 1741 for the evaluation function is copied into the neural net 1740 for the evaluation function of the evaluation executing part. Thus, the neural net 1740 for the evaluation function is updated to a new one. As a result, the control output determining section 5 and the control result good/defect determining section 6 are

capable of performing evaluation by a new neural network for the evaluation function.
[0104]
In the present embodiment, if the conditions such as the plate width, the plate thickness, the steel grade of the material, etc. taken as the controlled targets differ, the suitable evaluation functions are considered to differ. Therefore, the learned neural nets separately having done the learning every their conditions may be stored in the evaluation function database DB5 as the evaluation functions and properly used according to the conditions. Further, by taking into consideration the plate width, the plate thickness, the steel grade, etc. as the rolling actual results, they can also be covered with one neural net.
[0105]
The numeric value of the evaluation function obtained from the neural net for the evaluation function is possibly incorrect during a period till the learning progresses to some extent. Therefore, the evaluation function selection method learning unit 173 may select and use the evaluation function in consideration of not only the setting evaluation functions Al to AN but also the rolling state.
[0106]
As described above, the evaluation function setting section 17 sets the optimum evaluation function corresponding to the rolling state to the control output

setting section 5 and the control result good/defect determining section 6.
[0107]
FIG. 16 is a diagram for describing an outline of the control output determining section 5. The control output determining section 5 is comprised of a rolling phenomenon model 501 and a shape correction good/defect determining part 502. The control output determining section 5 obtains the actual result data Si from the control target plant 1, the control operation quantity S3 from the control output computing section 3, and the information of the output determination database DB3 and applies the control operation quantity output permission/non-permission data S4 to the control operation end. With such a configuration, the control output determining section 5 predicts a change in the shape where the control operation quantity S3 computed in the control output computing section 3 is output to the rolling mill being the control target plant 1, by being input to the known model of control target plant 1 (rolling phenomenon model 501 in the case of the embodiment of FIG. 16), and suppresses the control operation quantity output SO where the shape is expected to be deteriorated, to thereby prevent the shape from being greatly deteriorated.
[0108]
Described more specifically, the control operation quantity S3 is input to the rolling phenomenon model 501 to

predict a change in the shape due to the control operation quantity S3 and thereby compute shape deviation correction quantity prediction data 503. On the other hand, the shape deviation correction quantity prediction data 503 is added to the shape detector data Si (shape deviation actual result data 504 at the present time) from the control target plant 1 to obtain shape deviation prediction data 505. The shape deviation prediction data 505 is evaluated to make it possible to predict how the shape changes when the control operation quantity S3 is output to the control target plant 1. Based on the current shape deviation actual result data 504 and shape deviation prediction data 505, the shape correction good/defect determining part 502 determines whether the shape changes in the direction to be improved or the direction to be deteriorated, to obtain the control operation quantity output permission/non-permission data S4. [0109]
The shape correction good/defect determining part 502 performs a good/defect determination as to the shape correction specifically in the following manner. First, in order to consider the priority of control in the plate width direction, the shape correction good/defect determining part 502 makes a good/defect determination as to a change in the shape using the evaluation function corresponding to the rolling state set from the evaluation function setting section 17. For example, the shape

correction good/defect determining part 502 performs a good/defect determination as to the shape change by using an evaluation function J expressed in the following equation. In the following equation, sfb (i) is a shape deviation actual result 504, sest (i) is shape deviation prediction 505, i is a shape detector zone, rand is a random number term, and JAI is an evaluation function set by the evaluation function setting section 17.
J = JAi(sfb(i)) - JAi(sest(i)) + rand
[0110]
When the above evaluation function J is used, the evaluation function J becomes positive when the shape becomes good. When the shape becomes defective, the evaluation function J becomes negative. Further, rand is the random number term and changes the evaluation result of the evaluation function J on the random number basis. Thus, since there occurs a case where the evaluation function J becomes positive even when the shape is deteriorated, it is possible to learn the relation between the shape pattern and the control method even in the case where the rolling phenomenon model 501 is not proper. Here, rand is changed as appropriate in such a manner that the maximum value is made large where the model of the control target plant 1 is uncertain, and 0 is taken where it is desired to learn the control method to some extent and then execute stable control.

[0111]
The shape correction good/defect determining part 502 computes the evaluation function J and outputs the control operation quantity output permission/non-permission data S4 in such a manner as to when J>0, make the control operation quantity output permission/non-permission data S4=l (permission) and when J<0, make the control operation quantity output permission/non-permission data S4=0 (non-permission) . [0112]
The control output suppressing section 4 determines the presence or absence of the output of the control operation quantity output SO to the control target plant 1 according to the control operation quantity output permission/non-permission data S4 being the result of determination by the control output determining section 5. The control operation quantity output permission/non-permission data S4 corresponds to #1 to #nAS-U position change quantity outputs, an upper first intermediate shift position change quantity output, and a lower first intermediate shift position change quantity output, and is determined by:
IF (control operation quantity output permission/non-permission data S4=0) THEN
#1 to #nAS-U position change quantity outputs=0
Upper first intermediate shift position change quantity output=0

Lower first intermediate shift position change quantity output=0
ELSE
#1 to #nAS - U position change quantity outputs=#l to #nAS - U position change quantities
Upper first intermediate shift position change quantity output=upper first intermediate shift position change quantity
Lower first intermediate shift position change quantity output=lower first intermediate shift position change quantity and
ENDIF. [0113]
The control executing device 20 executes the above computation according to the actual result data Si from the control target plant 1 (rolling mill) and outputs the control operation quantity output SO to the control target plant 1 (rolling mill) to thereby execute the shape control. [0114]
Next, a description will be made as to an operation outline of the control method learning device 21. The control method learning device 21 makes use of time-delayed data of the data used in the control executing device 20. A time delay Z_1 means e~TS and indicates that data is delayed by a time T set in advance. Since the control target plant 1 has a time response, a time delay exists due

to the control operation quantity output SO until the actual result data changes. Therefore, after execution of the control operation, the learning is executed using the actual result data at the lapse of only the delay time T. Since it takes a few seconds for a shapemeter to detect a change in shape after the operation command output to the AS-U or first intermediate roll upon the shape control, the time may be set to T=2 to 3 seconds or so (since the delay time changes even depending on the type of the shape detector or the rolling speed, the optimum time taken until the change of the control operation end reaches a shape change may be set as T). [0115]
FIG. 17 is a diagram for describing an operation outline of the control result good/defect determining section 6. Such a good/defect determination evaluation function Jc as expressed in the following equation is used in the shape change good/defect determining unit 602.
Jc = JAi&fbQ)) ~ JAi^lasS))
[0116]
Incidentally, in the above equation, sfb (i) is shape deviation actual result data included in the actual result data Si, slast (i) is a shape deviation actual result data previous value, and JAI is an evaluation function set by the evaluation function setting section. Here, as the

evaluation function JAI, the evaluation function JAI manually set in advance by the evaluation function setting section 17, or the evaluation function (learning evaluation function) learned by the evaluation function learning unit 174 is set. It is determined by the good/defect determination evaluation function Jc whether a control result is good or defective. Further, even in the case where the control operation quantity output permission/non-permission data S4 being the determination result of the control output determining section 5 is 0 (control output non-permission), the shape is determined to have been deteriorated though the control operation quantity output to the control target plant 1 = 0 in practice. [0117]
Here, when the control operation quantity output permission/non-permission data S4=0, the control result good/defect data S6 is assumed to be S6=-l. Further, a threshold upper limit LCU and a threshold lower LCL are set in advance under a threshold condition (LCU>0>LCL). At this time, if the result of comparison with the good/defect determination evaluation function Jc is Jc>LCU, the control result good/defect data S6 is assumed to be S6=-l (the shape is deteriorated). If LCU>Jc>0, the control result good/defect data S6 is assumed to be S6=0 (the shape changes in the direction to be deteriorated). If 0>Jc>LCL, the control result good/defect data S6 is assumed to be S6=l (the shape changes in the direction to be

indefective). If Jc OTHEN C ref = Cref - ACref IF Cref < OTHEN Cref = Cref + ACref
C ref =s Cref • • • ( 8 )
••• (9)
IF Cref > OTHEN Cref = Cref + ACref IF Cref < OTHEN Cref - Cref - ACref
[0125]
In the processing stage 73, the operation degree correction quantity ACref is determined by (10) and (11) equations according to the corrected operation command value Cref.
C ref = G • ({OVref + AOref) - (OMref - AOref)) ■■■ (10)

AOref = -(^-^-(OFref-OMref)) •■• (11)
2 G
[0126]
In the processing stage 74, the teacher data OP'ref and OM'ref to the neural net 111 are determined by a (12) equation.
OP1 ref = OPref + AOref 1
>- •»* f i o OM ref = OMref - AOref J
[0127]
Thus, as shown in FIG. 18, the learning data creating section 7 actually computes the operation command value correction value C'ref according to the control result good/defect data S6 being the determination result in the control result good/defect determining section 6 from the operation command value Cref output to the control target plant 1. Specifically, in the case of the control result good/defect data S6=l, it corresponds to the case in which the control direction is OK, but it is determined that the control output is insufficient, and the operation command value is increased by ACref in the same direction. In the case of the control result good/defect data S6=-l in reverse, it corresponds to the case in which the control direction is determined to be incorrect, and the control command value is decreased by ACref in a reverse direction. Since the conversion gain G is set in advance and thereby

already known, the correction quantity AOref can be determined if the values on the operation degree positive and negative sides are known. Here, ACref is set by obtaining an appropriate value in advance through a simulation or the like. According to the above procedures, the teacher data OP'ref and OM'ref used in the control rule learning section 11 can be determined by the above (12) equation.
[0128]
Incidentally, although the above description has been made by the simple case examples in FIG. 19, the AS-U operation degree 301 relative to the #1 to #nAS-U, and the first intermediate operation degree 302 relative to the upper first intermediate roll shift and the lower first intermediate roll shift are all executed in practice and defined as the teacher data (AS-U operation degree teacher data and first intermediate operation degree teacher data) of the neural net 111 used in the control rule learning section 11.
[0129]
FIG. 20 shows a data example stored in the learning data database DB2. Combinations of a large number of input data S8a and teacher data S7a are required to learn the neural net 111. Thus, the teacher data S7a (AS-U operation degree teacher data, first intermediate operation degree) created in the learning data creating section 7 is combined with the time-delayed data S8a of the input data SI

(standardized shape deviation 201 and shape deviation stage) input to the control rule executing section 10 by the control executing device 20 and defined as a pair of learning data Sll, which in turn is stored in the learning data database DB2.
[0130]
Incidentally, while the various databases DB1, DB2, DB3, DB4, and DB5 have been used in the plant control apparatus of FIG. 3, there is shown in FIG. 20, a configuration of a neural net management table TB for managing and operating the respective databases DB1, DB2, DB3, and DB4 on a linkage basis. The management table TB includes a specification management table. Specifically, the management table TB is divided according to (Bl) the plate width, (B2) the steel grade, and the evaluation functions Al to AN about the priority of control in terms of the specification. As (Bl) the plate width, for example, four divisions of a 3 feet width, a meter width, a 4 feet width, and a 5 feet width are used, and as the steel grade, 10 divisions of steel grades (1) to (10) or so are used. Further, N (N is the set number of evaluation functions, and N=6 in the present embodiment) types are taken in terms of the control's evaluation functions. In this case, the number of divisions becomes 80, and 240 neural nets are properly used according to the rolling conditions and utilized.
[0131]

The neural net learning control part 112 ties the learning data being the combination of the input data and the teacher data such as shown in FIG. 20 to the corresponding neural net No. in accordance with the neural net management table TB of FIG. 21 and is stored in such a learning data database DB2 as shown in FIG. 22.
[0132]
Each time the control executing device 20 performs the shape control on the control target plant 1, the learning data is created in two sets according to the evaluation function. This is because N types of teacher data are created since the control result good/defect determination is made for the same input data and control output by using the N evaluation functions about the priority of the control. When the teacher data is accumulated to some extent (e.g., 200 sets) or accumulated in the learning data database DB2, the neural net learning control part 112 instructs the learning of the neural net 111.
[0133]
A plurality of neural nets have been stored in the control rule database DB1 in accordance with such a management table TB as shown in FIG. 21. The neural net learning control part 112 specifies the neural net No. requiring learning, and the neural net selecting unit 113 takes out the corresponding neural net from the control rule database DB1 and sets the same to the neural net 111. The neural net learning control part 112 instructs the

input data creating unit 114 and the teacher data creating unit 115 to take out the input data and the teacher data corresponding to the corresponding neural net from the learning data database DB2 and executes the learning of the neural net 111 by using those. Incidentally, various methods have been proposed as the learning method of the neural net, and any of them may be utilized.
[0134]
When the learning of the neural net 111 is completed, the neural net learning control part 112 writes the neural net 111 being the learning result back to the position of the corresponding neural net No. of the control rule database DB1, where the learning is completed.
[0135]
The learning may be performed on all neural nets defined in FIG. 21 at a time at prescribed time intervals
(every day, for example), or only neural nets in the neural net No. in which new learning data is accumulated to some extent (100 sets, for example) may be learnt at that time.
[0136]
Thus, the following can be implemented without greatly disturbing the shape of the rolling mill being the control target plant 1:
1) A reference shape pattern and a control operation corresponding thereto are separately set in advance, the combination of the shape pattern and the control operation is learnt without learning a control operation method, and

the control operation is performed using the same,
2) since a new control rule cannot be predicted preliminarily, and a completely unpredictable control rule may be made optimum, the control operation end is operated at random to find a control result relative to it while looking at it, and
3) an evaluation function to determine the priority of control relative to a controlled target is set so as to match with the sense of the operator and match with a manual operation method of the operator according to the state of the controlled target.
[0137]
Incidentally, the neural net used in the control executing device 20 is stored in the control rule database DB1. However, if the stored neural net is merely subjected to the initial processing at a random number, the learning of the neural net progresses, and time is taken until appropriate control is made possible. Therefore, when the control section is constructed for the control target plant 1, the learning of the control rule is performed by simulation in advance, based on the control model of the control target plant 1 which has been known at that time, and the neural net in which learning in a simulator is completed is stored in the database, whereby it is possible to control the performance to some extent from the start-up of the control target plant.
[0138]

Further, since the learning of the neural net is carried out using the evaluation function matched with the manual operation method of the operator, it is not necessary for the operator to perform the manual operation on the change in the controlled target due to the control output. It is possible to reduce loads imposed on the operator and improve control accuracy and operational efficiency. [0039]
The embodiment described above includes items shown below. However, the items included in the embodiment are not limited to the items shown below. [0140]
The control apparatus of the present disclosure is a control apparatus which controls a controlled target. The control apparatus includes a control executing device which applies a control output to the controlled target in accordance with a control rule given thereto, a control method learning device which evaluates the control output applied to the controlled target by using an indicated evaluation function, creates learning data by utilizing an evaluation result thereof, and learning the learning data to thereby construct the control rule, and applies the control rule to the control executing device, and an evaluation function setting section which holds a plurality of evaluation functions in advance, selects any of the plural evaluation functions, based on a control state to

the controlled target, and specifies the selected evaluation function to the control method learning device. [0141]
According to this configuration, since the control output is applied to the controlled target in accordance with the control rule constructed by learning the learning data having used the evaluation result of the evaluation for the control output by the evaluation function selected based on the control state, it is expected that control based on an appropriate good/defect determination as to the control result can be executed. [0142]
Also, according to the present disclosure, the evaluation function setting section calculates an evaluation function good/defect determination index for each of the plural evaluation functions, based on the control state to the controlled target and a manual operation by an operator, and selects an evaluation function specified to the control method learning device, based on the evaluation function good/defect determination index. According to this configuration, the evaluation function in which control that the operator intends becomes high in evaluation becomes likely to be selected by using the relation between the manual operation by the operator and the control state to the controlled target. [0143]
Further, according to the present disclosure, the

evaluation function setting section calculates an evaluation value of the evaluation function at each of a timing at which the operator starts the manual operation and a timing at which the operator finishes the manual operation, and calculates the evaluation function good/defect determination index by using the evaluation value. According to this configuration, since the operator starts the manual operation when the shape of a plate is determined to be deteriorated during a rolling operation, and continues the manual operation until the shape is determined to be improved, it is possible to obtain the intention of the operator from the evaluation value at that time. [0144]
Furthermore, according to the present disclosure, the evaluation function setting section calculates an evaluation value a of the evaluation function in the timing at which the operator starts the manual operation, and an evaluation value b of the evaluation function in the timing at which the operator finishes the manual operation, and calculates the evaluation function good/defect determination index as (a-b)/b. According to this configuration, even when a calculation method differs every plural evaluation functions, it is possible to mutually compare evaluation function good/defect indices. [0145]
Still further, according to the present disclosure,

the evaluation function serves to output the evaluation result with the control output to the controlled target and actual result data of the controlled target on which a control result of the control output is reflected, as inputs. The evaluation function setting section learns learning data based on the manual operation by the operator, the control output to the controlled target, and the actual result data of the controlled target to thereby construct the evaluation function. According to this configuration, since the manual operation of the operator is used, it is possible to construct an evaluation function on which the intention of the operator is reflected. [0146]
Still further, according to the present disclosure, the evaluation function setting section learns learning data based on the control output to the controlled target and the actual result data of the controlled target in the timing at which the operator starts the manual operation and the timing at which the operator finishes the manual operation, to thereby construct the evaluation function. According to this configuration, since the operator starts the manual operation when the shape of the plate is evaluated to be deteriorated during the rolling operation, and continues the manual operation until the shape is evaluated to be improved, it is possible to construct an evaluation function which performs evaluation close to the evaluation of the operator with an evaluation value

reflecting the evaluation of the operator as learning data.
[0147]
Still further, according to the present disclosure, the evaluation function setting section generates learning data with an evaluation value in the timing at which the operator starts the manual operation as a predetermined value c, generates learning data with an evaluation value in the timing at which the operator finishes the manual operation as a predetermined value d, and learns the learning data to thereby construct the evaluation function. According to this configuration, since the operator starts the manual operation when the shape of the plate is determined to be deteriorated during the rolling operation, and continues the manual operation until the shape is determined to be improved, it is possible to obtain the intention of the operator from the evaluation value at that time.
[0148]
Still further, according to the present disclosure, the control executing device includes a control rule executing section which applies a control output to the controlled target in accordance with a combination of the actual result data of the controlled target and a control operation, a control output determining section which determines using the evaluation function whether the control output outputted from the control rule executing section is applicable, and notifies inappropriateness of

the combination of the actual result data and the control operation to the control method learning device when the control output is determined not to be applied, and a control output suppressing section which, when the control output determining section has determined the control output not to be applied, prevents the control output from being output to the controlled target. The control method learning device includes a control result good/defect determining section which, when the control executing device actually outputs the control output to the controlled target, makes a determination as to a non-defect/defect of a control result about whether the actual result data is improved or deteriorated by the control output, using the evaluation function set by the evaluation function setting section after a time delay until the control output is reflected on the actual result data of the controlled target, a learning data creating section which obtains teacher data by using the non-defect/detect of the control result determined by the control result good/defect determining section and the control output, and a control rule learning section which learns the actual result data and the teacher data as learning data. With the learning by the control method learning device, a combination of separate actual result data and control operations to a plurality of controlled targets is obtained according to the state of the control target plant. The so-obtained combination of the actual result data and the

control operations is used as a defined combination of actual result data and a control operation of the control target plant in the control rule executing section.
[0149]
Further, the plant control apparatus of the present disclosure can be actually realized as a computer system. In this case, however, a plurality of program groups are formed in the computer system.
[0150]
These program groups are, for example, programs for achieving processing of a control executing device: a control rule executing program which applies a control output in accordance with a defined combination of actual result data of a control target plant and a control operation, a control output determining program which determines whether the control output outputted by the control rule executing program is allowed, and notifies the control method learning device of the actual result data and the control operation being erroneous, and a control output suppressing program which, when the control output determining program has determined the actual result data of the control target plant to be deteriorated where the control output is output to the control target plant, prevents the control output from being output to the control target plant, and
programs for achieving processing of the control method learning device: a control result good/defect

determining program for, when the control executing device actually outputs the control output to the control target plant, achieving processing of a control result good/defect determination to make a determination as to a non-defect/defect of a control result about whether the actual result data is improved or deteriorated compared with that before the corresponding control after a time delay until a control effect appears in the actual result data, a learning data creating program which obtains teacher data by using the non-defect/detect of the control result in the control result good/defect determining program and the control output, and a control rule learning program which learns the actual result data and the teacher data as learning data. Further, with the learning by the control method learning device, a combination of separate actual result data and a control operation to a plurality of controlled targets is obtained according to the state of the control target plant, and the so-obtained combination of the actual result data and the control operation is used as a defined combination of actual result data and a control operation of the control target plant in the control rule executing program. [0151]
Incidentally, it is necessary to define an initial value of a neural net when the apparatus of the present invention is applied to an actual plant. As regards this point, however, it is desirable to create a combination of

actual result data and a control operation by simulation using a control model of a control target plant before executing control in the control target plant and shorten a period for learning the combination of the actual result data and the control operation in the control target plant. [Industrial Applicability] [0152]
The present invention relates to a control method and section of a rolling mill being one of rolling facilities, for example, and there is no particular problem when actually applied.

WE CLAIM

1.A control apparatus which controls a controlled
target, comprising:
a control executing device which applies a control output to the controlled target in accordance with a control rule given thereto;
a control method learning device which evaluates the control output applied to the controlled target by using a specified evaluation function, creates learning data by utilizing an evaluation result thereof, learns the learning data to thereby construct the control rule, and applies the control rule to the control executing device; and
an evaluation function setting section which holds a plurality of evaluation functions in advance, selects any of the plural evaluation functions, based on a control state to the controlled target, and specifies the selected evaluation function to the control method learning device.
2.The control apparatus according to claim 1,
wherein the evaluation function setting section
calculates an evaluation function good/defect determination index for each of the plural evaluation functions, based on the control state to the controlled target and a manual operation by an operator, and selects an evaluation function specified to the control method learning device, based on the evaluation function good/defect determination index.

3.The control apparatus according to claim 2,
wherein the evaluation function setting section
calculates an evaluation value of the evaluation function at each of a timing at which the operator starts the manual operation and a timing at which the operator finishes the manual operation, and calculates the evaluation function good/defect determination index by using the evaluation value.
4.The control apparatus according to claim 3,
wherein the evaluation function setting section
calculates an evaluation value a of the evaluation function in the timing at which the operator starts the manual operation, and an evaluation value b of the evaluation function in the timing at which the operator finishes the manual operation, and calculates the evaluation function good/defect determination index as (a-b)/b.
5.The control apparatus according to claim 1,
wherein the evaluation function serves to output the
evaluation result with the control output to the controlled target and actual result data of the controlled target on which a control result of the control output is reflected, as inputs, and
wherein the evaluation function setting section learns learning data based on the manual operation by the

operator, the control output to the controlled target, and the actual result data of the controlled target to thereby construct the evaluation function.
6. The control apparatus according to claim 5,
wherein the evaluation function setting section learns learning data based on the control output to the controlled target and the actual result data of the controlled target in the timing at which the operator starts the manual operation and the timing at which the operator finishes the manual operation, to thereby construct the evaluation function.
7. The control apparatus according to claim 6,
wherein the evaluation function setting section
generates learning data with an evaluation value in the timing at which the operator starts the manual operation as a predetermined value c, generates learning data with an evaluation value in the timing at which the operator finishes the manual operation as a predetermined value d, and learns the learning data to thereby construct the evaluation function.
8. The control apparatus according to claim 1,
wherein the control executing device includes a
control rule executing section which applies a control output to the controlled target in accordance with a

combination of the actual result data of the controlled target and a control operation, a control output determining section which determines using the evaluation function whether the control output outputted from the control rule executing section is applicable, and notifies impropriety of the combination of the actual result data and the control operation to the control method learning device when the control output is determined not to be applied, and a control output suppressing section which, when the control output determining section has determined the control output not to be applied, prevents the control output from being output to the controlled target, and
wherein the control method learning device includes a control result good/defect determining section which, when the control executing device actually outputs the control output to the controlled target, makes a determination as to a non-defect/defect of a control result about whether the actual result data is improved or deteriorated by the control output, using the evaluation function set by the evaluation function setting section after a time delay until the control output is reflected on the actual result data of the controlled target, a learning data creating section which obtains teacher data by using the non-defect/detect of the control result determined by the control result good/defect determining section and the control output, and a control rule learning section which learns the actual result data and the teacher data as

learning data, and, with the learning by the control method learning device, a combination of separate actual result data and a control operation to a plurality of controlled targets is obtained according to the state of the control target plant, and the so-obtained combination of the actual result data and the control operation is used as a defined combination of actual result data and a control operation of the control target plant in the control rule executing section.
9.A control method for controlling a controlled target, which is executed by a computer, comprising:
applying a control output to the controlled target in accordance with a given control rule;
evaluating the control output applied to the controlled target by using a specified evaluation function;
creating learning data by utilizing an evaluation result thereof;
learning the learning data to construct the control rule; and
selecting and specifying any of a plurality of
evaluation functions held in advance, based on a control
state to the controlled target.