DE102021214272A1 - Plant control system, plant control method and program - Google Patents

Abstract

In a plant control system, a control rule is effectively corrected while reducing a risk of affecting plant control. A plant control system includes a control method learning unit that learns a combination of the result data of the control object creation and the control operation, a control execution unit that executes control of the object creation according to the combination of the result data and the control operation learned by the control method learning unit, and a quality determination rule learning unit that learns a combination of the combination of the result data of the object creation and the control operation and the control result quality. The quality determination of the control output is performed according to a specific combination of the object creation result data, the control operation, and the control result quality, and a control rule is learned with a quality determination result, the result data, and teacher data as learning data.

Description

BACKGROUND

The present invention relates to a plant control system, a plant control method and a program.

In various types of plants, plant control based on various control theories has been performed in order to obtain more appropriate control results through the control.

To describe an example of the facilities: For example, control of a rolling mill is performed with fuzzy control or neural fuzzy control as a control theory for shape control for controlling a corrugation of a plate. The fuzzy control is for shape control using a coolant, while the neural fuzzy control is for shape control of a Sendzimir plant. Here determined, as in Japanese Patent No. 2804161 described, the shape controller using the fuzzy neural controller detects a difference between an actual shape pattern detected by a shape detector and a target shape pattern, and determines a similarity of the actual shape pattern with a predetermined reference shape pattern. The shape control is performed by obtaining a control output quantity from the resultant similarity for an operation end using a control rule expressed by a control operation end operation quantity for the reference shape pattern set in advance.

An existing method of shape control of a Sendzimir plant using fuzzy neural control is described below.

The neural fuzzy control is used for form control of the Sendzimir plant. As in 25 shown, in a Sendzimir plant 50, a pattern recognition part 51 performs shape pattern recognition from an actual shape detected by a shape detector 52 and determines by calculation which of predetermined reference shape patterns is most similar to the actual shape. The data of the actual shape detected by the shape detector 52 is subjected to a pre-processing for pattern recognition by a shape detection pre-processing part 54 .

A control calculation part 53 performs control using a control rule including a control operation end operation amount for a predetermined shape pattern.

In this case calculated as in 26 1 shows the pattern recognition part 51 which of the shapes of patterns 1 to 8 is most similar to a difference (Δε) between an actual shape pattern (ε) detected by the shape detector 52 and a target shape (ε _ref ). The control calculation part 53 then selects and executes one of the control methods corresponding to patterns 1 to 8 based on the calculation result.

SUMMARY

im in Japanese Patent No. 2804161 According to the existing method described above, a typical shape is set in advance as the reference shape pattern, and control is performed based on the control rule for specifying a relation of the control operation end operation amount to the reference shape pattern. The learning of the control rule also concerns the control operation end operation amount with respect to the reference shape pattern, and the predetermined typical reference shape pattern is used directly. This leads to the disadvantage of achieving shape control responsive only to a specific shape pattern.

Although the reference shape pattern is determined in advance by a human based on knowledge of an object rolling mill or a history of stored shape results and a manual intervention operation, it is difficult to cover all shapes occurring in an object rolling mill and a material to be rolled. Thus, when a shape different from the reference shape pattern occurs, control based on the shape control is not performed and a shape deviation remains without being suppressed. Alternatively, such a shape is mistakenly recognized as a similar reference shape pattern, and the shape may be degraded by an erroneous control operation.

In the existing shape control, therefore, the control regulation is learned using the preset reference shape pattern and the corresponding control rule, and then the control is performed, resulting in the disadvantage that the improvement in the control accuracy is restricted.

To solve such a disadvantage, for example, Japanese Unexamined Patent Application Publication No. 2018-180799 ( JP-A-2018-180799 ) provided a method. JP-A-2018 - 180799 describes a process for generating a perturbation during control to allow a neural network to gradually become more intelligent through learning. In the process of generating a control disturbance as in JP-A-2018-180799 However, a control object facility is operated and the control disturbance is generated during operation, which disturbs the actual operation of the control object facility and is not preferable in operation. Furthermore, since a neural network is not appropriate until the control object plant is partially operated, appropriate control may not be performed for a while after the start of operation.

An object of the invention is to provide a plant control system, a plant control method, and a program, each of which can reduce a risk of deteriorating the control of a plant and effectively correct a control rule.

For example, any of the configurations described in the claims can be used to achieve the foregoing object.

The present application includes a variety of approaches to accomplish the task. In an example, a plant control system is applied to a control object plant to recognize a combination pattern of result data of the control object plant and perform control.

The plant control system includes a control method learning unit that learns a combination of the result data of the control object plant and the control operation, a control execution unit that executes control of the control object plant according to the combination of the result data and the control operation learned by the control method learning unit, and a quality determination rule learning unit, which learns a combination of the combination of the result data of the control object creation and the control operation and the control result quality.

• einen Steuerregel-Ausführungsteil, der eine Steuerausgabe entsprechend einer bestimmten Kombination der Ergebnisdaten der Steuerobjektanlage und der Steueroperation liefert,
• einen Steuerausgabequalität-Bestimmungsregel-Ausführungsteil, der eine Qualitätsbestimmung der Steuerausgabe entsprechend der bestimmten Kombination der Ergebnisdaten der Steuerobjektanlage, der Steueroperation und der Steuerergebnisqualität durchführt,
• einen Neusuchen-Operationsmengen-Berechnungsteil, der eine Neuoperation-Suchoperationsmenge auf der Basis der Qualitätsbestimmung durch den Steuerausgabequalität-Bestimmungsregel-Ausführungsteil berechnet, und
• einen Steuerausgaben-Unterdrückungsteil, der die Qualitätsbestimmung durch den Steuerausgabequalität-Bestimmungsregel-Ausführungsteil und Simulationsdaten unter Verwendung eines Steuersimulators verwendet, um zu verhindern, dass die Steuerausgabe an die Steuerobjektanlage gesendet wird, wenn bestimmt wird, dass sich die Ergebnisdaten der Steuerobjektanlage verschlechtert haben, wenn die Steuerausgabe an die Steuerobjektanlage gesendet wird.

The control execution unit includes

• a control rule execution part that provides a control output according to a specified combination of the result data of the control object creation and the control operation,
• a control output quality determination rule execution part that performs quality determination of the control output according to the determined combination of the control object creation result data, the control operation, and the control result quality,
• a re-search operation amount calculation part that calculates a re-operation search operation amount based on the quality determination by the control output quality determination rule execution part, and
• a control output suppression part that uses the quality determination by the control output quality determination rule execution part and simulation data using a control simulator to prevent the control output from being sent to the control object installation when it is determined that the result data of the control object installation has deteriorated the control output is sent to the control object facility.

• einen Steuerergebnisqualität-Bestimmungsteil, der die Qualität eines Steuerergebnisses nach einer Verzögerung bestimmt, bevor eine Steuerwirkung in den Ergebnisdaten erscheint, wenn die Steuerausführungseinheit die Steuerausgabe an die Steuerobjektanlage sendet, und
• einen Qualitätsbestimmungsregel-Lernteil, der als Lerndaten die Qualität des durch den Steuerergebnisqualität-Bestimmungsteil bestimmten Steuerergebnisses, der Ergebnisdaten und der Steuerausgabe lernt.

The quality determination rule learning unit includes

• a control result quality determination part that determines the quality of a control result after a delay before a control effect appears in the result data when the control execution unit sends the control output to the control object facility, and
• a quality determination rule learning part that learns, as learning data, the quality of the control result determined by the control result quality determination part, the result data, and the control output.

• einen Lerndaten-Erzeugungsteil, der die Qualitätsbestimmung der Steuerausgabe durch den Steuerausgabequalität-Bestimmungsregel-Ausführungsteil und die Steuerausgabe zum Erhalten von Lehrerdaten verwendet, und
• einen Steuerregel-Lernteil, der die Ergebnisdaten und die Lehrerdaten als Lerndaten lernt.

The control method lesson includes

• a learning data generation part that uses the quality determination of the control output by the control output quality determination rule execution part and the control output to obtain teacher data, and
• a control rule learning part that learns the result data and the teacher data as learning data.

According to the invention, a control rule of a mold pattern and an operation method can be effectively automatically corrected while reducing a hazard to a facility to be optimal in response to the age change in the environment of the facility. Therefore, according to the invention, it is possible to improve control accuracy, reduce a starting period of a control part, and respond to deterioration due to age.

Furthermore, according to the invention, the performance of the control rule is pre-evaluated, which advantageously reduces a risk to the plant due to the use of a new control rule and improves the control performance due to the selection of an optimal control rule.

Other problems, configurations and effects will be apparent from the following description of the embodiment.

character list

• 1 12 is a block diagram showing an overview of a plant control system according to an exemplary embodiment of the invention.
• 2 12 shows a specific configuration example of a control rule execution part of the exemplary embodiment.
• 3 Fig. 12 is a block diagram showing an example of a control output quality determination rule execution part of the exemplary embodiment.
• 4 12 shows a specific configuration example of a control rule learning part of the exemplary embodiment.
• 5 Fig. 12 is a block diagram showing an example of a quality determination rule learning part of the exemplary embodiment.
• 6 shows an example of the control result quality determination for a control method in the form control of a Sendzimir plant.
• 7 12 shows the configuration of a neural network used for the shape control of the Sendzimir plant according to the exemplary embodiment.
• 8A and 8B 14 each show an explanatory drawing of the shape deviation and a control method of the exemplary embodiment.
• 9 Fig. 12 is a block diagram showing an example of a control input data generation part of the exemplary embodiment.
• 10 Fig. 14 is a block diagram showing an example of a control output calculation part of the exemplary embodiment.
• 11 shows the configuration of a neural network used for the quality determination of the Sendzimir work. is, according to the exemplary embodiment.
• 12 12 shows an operation amount calculation method of a re-search operation amount calculation part of the exemplary embodiment.
• 13 Fig. 14 is a block diagram showing an example of a control output determination part of the exemplary embodiment.
• 14 Fig. 14 is a block diagram showing an example of a control output calculation part of the exemplary embodiment.
• 15 12 shows processing steps and processing contents of a learning data generation part of the example. embodiment.
• 16 12 shows an example of data stored in a learning data database of the exemplary embodiment.
• 17 Fig. 12 shows an example of a neural network management table of the exemplary embodiment.
• 18 Figure 12 shows a skeletal frame of an example of a learning data database of the exemplary embodiment.
• 19 12 shows a block diagram of an example of a control quality determination part of the exemplary embodiment.
• 20 Fig. 12 shows a data sample stored in a learning data database of the exemplary embodiment.
• 21 Fig. 12 shows an example of a neural network management table of the exemplary embodiment.
• 22 Figure 12 shows an example of a learning data database of the exemplary embodiment.
• 23 12 is a block diagram showing an example of a plant control system including a control rule evaluation unit of the exemplary embodiment.
• 24 12 is a block diagram showing an example hardware configuration of the plant control system of the example embodiment.
• 25 shows a configuration diagram of a. example of the Sendzimir work.
• 26 shows an example of a list of control rules in the form control of the Sendzimir factory.

DETAILED DESCRIPTION

A plant control system of an exemplary embodiment (hereinafter referred to as “this example”) of the invention will be described below with reference to the accompanying drawings.

First, before the description of this example, details and summary of the invention are described with an example case where the plant control system is applied to a mold controller of a rolling mill.

First, the following requirements (1), (2), (3), and (4) are required to achieve a plant control system that can effectively correct a control rule with less danger to a plant.
Requirement (1): To improve the control rule, when a control operation to provide a good control result has not been learned, the control operation is greatly changed, and when a good control result is obtained, the changed control operation is adopted as a new control operation method. When a control operation has been learned to provide good control results, the control operation is not changed or only slightly changed; and when a good control result is thereby obtained, this control operation is adopted as a new control operation method.

Requirement (2): A combination of a shape pattern, a control operation, and the quality of a control result is learned based on actual equipment data, thereby creating a model capable of estimating the quality of a control result more accurately than a simulator using a machine model , is determined, which can be continuously adapted to the last system status through regular automatic learning.

Requirement (3): A model with which the quality determination of a control result is estimated is used to improve the reliability of a function for suppressing the control output to a plant, in contrast to existing methods in which such a function was carried out only with a simple machine model .

Requirement (4): While a control rule learning data generation function is performed by a quality determination of a control result in the existing methods, a model for estimating the quality determination of a control result is used, thereby suppressing the influence of noise contained in plant data, enabling fine adjustment , which provides a small effect that is also included in learning data objects, and thereby erroneous determination of a control effect is prevented and thus variation in learning data is suppressed, resulting in stable control performance.

To meet such requirements (1) to (4), it is preferable to configure a neural network that can learn a combination of a shape pattern used for shape control, a control operation, and the quality of a control result in a controller. The controller requires estimating the quality of a control result by outputting a control operation on a mill-occurring shape pattern with a value obtained by sending the mill-occurring shape pattern and outputting the control operation to the neural network. Further, the controller uses an estimate of the quality of a control result to select a calculation method of the control operation set to search for a new control operation.

The controller checks the output using, for example, a simple model of a rolling mill, and expects an output for shape deterioration, and controls such an output not to send to a control operation end of the rolling mill, thereby preventing shape deterioration. At this time, the controller uses the estimation of the quality of a control result to determine the output suppression, and thus improves protection reliability and optimizes a suppression range, making it possible to expand an applicable range of the control function.

In an initial application stage with insufficient estimation accuracy of the quality of a control result, control operation output data must also be sent to a plant while being estimated to be poor, so that a learning range is expanded to a combination of an unlearned shape pattern, a control operation, and the quality of a control result.

At a stage where the estimation accuracy of the quality of a control result is sufficiently high since the quality of a control result can be estimated without sending the operation amount to the plant, the controller can generate learning data of the control rule.

The controller can estimate the quality of a control result using a neural network capable of estimating the quality of the control result to reduce the influence of noise in plant data, and can thus determine the quality of fine adjustment data that provides little effect. The control unit can thus generate learning data. Further, the controller can prevent erroneous quality determination caused by noise to improve the accuracy of learning data.

When the estimation accuracy of the quality determination of a control result is lowered due to environmental change of a plant due to aging deterioration, for example, the controller may perform relearning using the latest result data of the plant to estimate the quality determination of a control result according to the latest plant state.

In order to confirm the estimation accuracy of the quality determination of the control result, test data for the accuracy check is prepared separately from the data used for learning the neural network. The control device can estimate the estimation accuracy of the quality determination based on an error in the quality determination between an output value of the neural network and a value contained in the test data. Received shape pattern for accuracy check and output of the control operation, and confirm the control result included in the test data.

1 shows a configuration of a plant control system of this example. The plant control system of 1 includes a control execution unit 20, a control method learning unit 21, a quality determination rule learning unit 22, a plurality of databases DB (DB1 to DB8), and a management table TB of the databases DB.

The control execution unit 20 receives result data Si from a control object facility 1 and supplies a control operation set output determined according to a control rule (neural network) to the control object facility 1 and thus controls the control object facility 1. In this case, it is assumed that the control object facility 1 has the configuration shown in Fig. 25 Sendzimir movement 50 is as previously described. As before with 26 described, the control rule is a rule for calculating which of the shapes of a plurality of prepared shapes.a detected actual shape pattern A (ε) in a difference a (Δε) between the actual shape pattern A (ε) and a target shape (ε _ref ) is most similar. The control execution unit 20 selects and executes a control method from one of the patterns based on a calculation result corresponding to the control rule.

The control method learning unit 21 receives control input data S1 generated by the control execution unit 20 to perform learning and reflects a learned control rule to a control rule in the control execution unit 20. The quality determination rule learning unit 22 receives result data before and after controlling the control object facility 1 to perform learning and reflects the learned quality determination rule to a quality determination rule in the control execution unit 20.

The control execution unit 20 includes a control input data generation part 2, a control rule execution part 10, a control output calculation part 3, a control output suppression part 4, a control output determination part 5, a control output quality determination rule. execution part 17, a re-search operation amount calculation part 33, and a control output operation method selection part 18.

The control execution unit 20 allows the control input data generation part 2 to generate input data S<b>1 for the control rule execution part 10 from result data Si of a rolling mill that is the control object plant 1 .

The control rule execution part 10 performs control rule execution processing to generate a control operation end operation command S2 from the result data Si of the control object using a neural network (control rule) for expressing a relationship between the result data Si of the control object and the control operation end operation command S2 . The control expense calculation part 3 calculates a control operation amount S3 for a control operation end based on the control operation end operation command S2. Thus, the control execution unit 20 generates the control operation set S3 using the neural network depending on the result data Si of the control object installation 1.

The control output quality determination rule execution part 17 performs control output quality determination rule execution processing to calculate a control output quality determination estimate S9 on the result data Si of the control object and the control operation set S3 using a neural network (quality determination rule) for expressing a relationship between the result data Si of the control object, of the control operation set S and the control result quality data S6 on the control result. Further, the control output quality determination rule execution part 17 generates a control result quality determination estimate S11 from the result data Si of the control object and the control selective operation set S8 as described below.

The reseek operation amount calculation part 33 performs reseek operation amount calculation processing for calculating a reseek control operation amount S12 on the basis of the control expense quality determination estimate S9.

The control output operation method selection part 18 generates the selective control operation set S8 and a control method selection flag S14 on the basis of the control operation set S3 or the retrieval control operation set S12,

The control expense determination part 5 in the control execution unit 20 determines control operation amount output correctness data S4 for the control operation end using the result data Si from the control object facility 1 and the control operation amount S3 on the control expense calculation part 3 . The control output suppressing part 4 determines the output correctness of the control operation end selective control operation set S8 on the basis of the control operation set output correctness data S4 and the control result. Quality determination estimation S11 and outputs the allowable selective control operation set S8 as the control operation set output S0 for delivery to the control object facility 1. Thus, the selective control operation set S<b>8 determined to be abnormal is not sent from the control execution unit 20 to the control object facility 1 .

The control execution unit 20 configured in this way refers to a control rule database DB1, an output determination database DB3, and a quality determination rule database DB5 for executing the foregoing processing.

The control rule database DB1 is accessible to both the control rule execution part 10 in the control execution unit 20 and a control rule learning part 802 in the control method learning unit 21 as described below.

The control rule database DB1 stores a control rule (neural network) as a learning result in the control rule learning part 802. The control rule execution part 10 refers to the control rule stored in the control rule database DB1.

A learning data database DB2 stores learning data obtained by the control rule learning part 802 .

An output determination database DB3 is accessible to the control output determination part 5 in the control execution unit 20 and stores output determination results.

A quality determination database DB4 stores data for quality determination.

The quality determination rule database DB5 stores a quality determination rule (neural network) as a learning result in a quality determination rule learning part 31. The quality determination rule database DB5 is accessible with both the control output quality determination rule execution part 17 in the control execution unit 20 and the quality determination rule learning part 31 in of the quality determination rule learning unit 22 as described below. The control output quality determination rule execution part 17 refers to a quality determination rule stored in the quality determination rule database DB5.

A learning data database DB6 stores learning data learned by the control method learning unit 21 .

An inspection data database DB7 stores inspection data required for quality determination.

2 12 shows a specific configuration example of the control rule execution part 10 of this example. The control rule execution part 10 receives input data S<b>1 generated by the control input data generation part 2 . The control rule execution part 10 processes the input data S1 and supplies the control operation end operation command S2 to the control output calculation part 3. The control rule execution part 10 includes a neural network 101 that outputs the control operation end operation command S2 according to a shape control rule as in FIG 26 outputs shown.

The control rule execution part 10 further includes a neural network selection part 102 and refers to the control rules stored in the control rule database DB1 to select a control rule optimum as a control rule for the neural network 101 and allows the neural network 101 to execute the control rule .

The control rule executing part 10 thus selects a required neural network from a plurality of neural networks sorted according to operator group or control purpose and uses the selected neural network. The control rule database DB1 preferably contains result data (such as operation group data) Si from which a neural network and a quality determination standard can be selected, as data from the control object facility 1.

Since the neural network becomes a control rule when executed, the neural network and the control rule are used interchangeably in the present specification.

3 Fig. 12 shows a specific configuration of the control output quality determination rule execution part 17.

The control output quality determination rule execution part 17 receives the input data S<b>1 generated by the control input data generation part 2 and the control operation set S<b>3 generated by the control output calculation part 3 . The control output quality determination rule execution part 17 generates the control output quality determination estimate S9 based on such data S1 and S3 and supplies the estimate S9 to the re-search operation amount calculation part 33.

Further, the control output quality determination rule execution part 17 receives the input data S<b>1 generated by the control input data generation part 2 and the selective control operation set S<b>8 generated by the control output operation method selection part 18 . The control output quality determination rule execution part 17 generates the control result quality determination estimate S11 on the basis of such data S1 and S8 and supplies the estimate S11 to the control output suppression part 4.

The control output quality determination rule execution part 17 includes a neural network 171 and a neural network selection part 172 .

The neural network 171 estimates a quality determination value of a control result based on previous control results when the control operation set S3 (control pattern) is output to the input data S1 (shape pattern).

The neural network selection part 172 refers to the quality determination rules stored in the quality determination rule database DB5 and thus selects the optimal quality determination rule for the neural network 171 .

The control output quality determination rule execution part 17 thus selects a required neural network from a plurality of neural networks sorted according to a difference in material property as a control object or a difference in quality determination standard.

The quality determination rule database DB5 preferably contains result data (such as operation group data) Si from which the material property can be selected as a control object and the quality determination standard, respectively, as data from the control object facility 1. Since the neural network becomes a control rule when executed, the neural network and the quality determination rule is used interchangeably in the present specification.

According to the description of 1 the control method learning unit 21 performs learning of the neural network 101 used in the control execution unit 20 .

The control method learning unit 21 includes a learning data generation part 801 and the control rule learning part 802.

The learning data generation part 801 in the control method learning unit 21 uses the selective control operation set S8 and the control method selection flag S14 generated in the control execution unit 20 and the control result quality determination estimate S11 generated by the control output quality determination rule execution part 17 to perform learning data generation processing to generate new teacher data S71 to be used for learning a neural network. The teacher data S7a generated by the learning data generating part 801 is supplied to the control rule learning part 802 .

The teacher data S71 corresponds to the control operation end operation command S2 issued from the control rule execution part 10.

The learning data generation part 801 uses the control result quality determination estimate S11 generated by the control output quality determination rule execution part 17 to obtain data as new teacher data S7a by estimating the control operation end operation command S2 output from the control rule execution part 10 .

4 shows a specific configuration example of the control rule learning part 802.

The control rule learning part 802 includes an input data generation part 114, a teacher data generation part 115, a neural network processing part 110, and a neural network selection part 113.

The control rule learning part 802 externally receives the input data S1 from the control input data generating part 2 and the new teacher data S7a from the learning data generating part 801. The control rule learning part 802 refers to the data stored in the control rule database DB1 and the learning data database Db2 .

In the control rule learning part 802 , the input data S<b>1 from the neural network processing part 110 is received via the input data generation part 114 .

In the control rule learning part 802, the new teacher data S7a from the learning data creating part 801 becomes in the teacher data creating part. 115 is combined with old teacher data S7b stored in the learning data database DB2, and the resultant total teacher data S7c is supplied to the neural network processing part 110. Such teacher data S7a and S7b are correspondingly stored and used in the learning data database DB2.

Also, the input data S8a from the control input data generating part 2 is combined with old input data S8b stored in the learning data database DB2 in the input data generating part 114, and the resulting total input data S8c is supplied to the neural network processing part 110. Such input data S8a and S8b are also stored and used accordingly in the learning data database DB2.

The neural network processing part 110 consists of a neural network 111 and a neural network learning control part 112 .

The neural network 111 receives the input data S8c from the input data generating part 114, the teacher data S7c from the teacher data generating part 115, and a control rule (a neural network) selected by the neural network selecting part 113, and stores a finally determined neural network in the control rule Database DB1.

The neural network learning control part 112 controls the input data generation part 114, the teacher data generation part 115 and the neural network selection part 113 at appropriate timings so that the neural network 111 receives inputs from the respective parts and stores them Processing results in the control rule database DB1.

Although the neural network 101 in the control rule execution part 10 of FIG 2 and the neural network 111 in the control method learning unit 21 of FIG 4 having the same concept, they differ from each other in the following point.

Specifically, the neural network 101 in the control rule execution part 10 has a predetermined content and obtains the control operation end operation command S2 as a corresponding output upon receiving the input data S1.

On the other hand, the neural network 111 in the control rule learning unit 21 learns an input/output relationship among the input data S1, the input data S8c to the control operation end operation command S2, and the teacher data S7c when such data is set as the learning data.

The basic idea of the processing of the control method learning unit 21 is as follows. First, when the control operation set output correctness data S4 indicates "Permissible", the control execution unit 20 sends the control operation set output S0 to the control object facility 1. Determined at this time when the control result quality determination estimate S11 indicates "Good" (the result data Si change for the better ), the learning data generating part 801 judges the selective control operation set S8 output from the control output operation method selecting part 18 as correct and generates learning data so that the output of the neural network agrees with the selective control operation set S8.

On the other hand, when the control operation amount output correctness data S4 indicates "Invalid" or the control result quality determination estimate S11 indicates "Invalid" (the result data Si changes for the worse) and if the control operation amount output S0 is sent to the control object facility 1, the learning data generation part 801 determines the selective control operation set S8 output from the control output operation method selection part 18 is false.

In such a case, the learning data generation part 801 checks whether the control-output operation method selection part 18 has selected the control operation set S3 based on the control method selection flag S14. When the control operation set S3 is determined to be selected, the learning data generating part 801 determines the control operation end operation command S2 output from the control rule executing part 10 as wrong and generates learning data so that no neural network output is performed. At this time, the output of the neural network is configured in advance so that one of two kinds of outputs, that is, outputs in the plus and minus directions, as the control output for the same control opera end of operation is provided, and the learning data is generated so that the control operation end operation command S2 is not issued to a side where the control output has been provided.

the inside 4 . The illustrated control rule learning part 802 performs the following processing as a result of the data processing by the neural network learning control part 112 .

First, the control rule learning part 802 uses learning data as a combination of the data S8c obtained from the input data S1 sent to the control execution unit 20 and the teacher data S7c generated by the teacher data generation part 115 to perform learning of the neural network used by the control rule execution part 10 101

The control rule learning part 802 in turn includes the neural network 111 identical to the neural network 101 in the control rule executing part 10, and performs operational tests on the neural network 111 under various conditions and learns responses to the tests, thus determining a control rule, which is confirmed to provide a better result as a result of learning.

Since such learning must be performed using a plurality of pieces of learning data, a plurality of pieces of old learning data for learning are extracted from the learning data database DB2 for storing learning data generated in the past, and processing is performed. The current learning data is stored in the learning data database DB2. The learned neural network is stored in the control rule database DB1 for use in the control rule execution part 10. FIG.

Learning of the neural network can be done collectively using old learning data every time new learning data is generated, or can be done collectively using old learning data after certain pieces (e.g., 100 pieces) of learning data have been stored.

According to such a configuration, the selection of a re-search operation set by the control-output operation method selection part 18 enables sending the re-search operation set to an object facility, generating learning data corresponding to the control result of the object facility, and learning a new control method.

According to the description of 1 the quality determination rule learning unit 22 performs the learning of the neural network 171 ( 3 ) for use in the control execution unit 20. When the control execution unit 20 sends the control operation set output S0 to the control object facility 1, a certain time is required until a control effect effectively occurs as a change in the result data Si. Thus, learning is performed using data delayed by the appropriate time. In 1 or others, a processing part DL, marked with "Z ^-1 ", indicates the presence of a corresponding delay during the transmission of the respective data.

The quality determination rule learning unit 22 includes a control result quality determination part 6, the quality determination rule learning part 31, and the quality determination database DB4.

The control result quality determination part 6 uses the result data Si and the result data previous value Si0 from the control object facility 1 and quality determination data S5 stored in the quality determination database DB4 to perform control result quality determination processing for determining whether the result data Si has changed for better or worse . The control result quality determination part 6 then outputs the control result quality data S6 indicating a determination result.

5 shows a specific configuration of the quality determination rule learning part 31.

The quality determination rule learning part 31 includes an input data generation part 314, a teacher data generation part 315, a neural network processing part 310, and a neural network selection part 313.

The quality determination rule learning part 31 externally receives data S12al as a delayed version of the input data S1 from the control input data generating part 2 and data S12a2 as a delayed version of the control output amount S0 from the control output suppressing part 4. The quality determination rule learning part 31 further receives the control result quality data S6 ( S13a) from the control result quality determination part 6.

In addition, the quality determination rule learning part 31 refers to data stored in the quality determination rule database Db5 and the learning data database DB6.

The input data S1 and the control output amount S0 are received from the neural network processing part 310 via the input data generating part 314 after appropriate delay compensation.

The control result quality data S6 (S13a) from the control result quality determination part 6 is combined with old teacher data S13b stored in the learning data database DB6 in the teacher data generation part 315, and the resulting total teacher data S13c is supplied to the neural network processing part 310. Such teacher data S13a and S13b are respectively stored and used in the learning data database DB6.

Also, the respective input data S12a1 and S12a2 from the control input data generating part 2 and control output suppressing part 4 are combined with old input data S12b stored in the learning data database DB6 in the input data generating part 314, and the resulting total input data S12c is sent to the neural network processing part 310 delivered. Such input data S12a1, S12a2 and S12b are respectively stored and used in the learning data database DB6.

The neural network processing part 310 consists of a neural network 311 and a neural network learning control part 312 . The neural network 311 receives the input data S12c from the input data generating part 314, the teacher data S13c from the teacher data generating part 315, and a control rule (neural network) selected by the neural network selecting part 313. The neural network 311 stores a finally determined neural network in the quality determination rule database DB5.

The neural network learning control part 312 controls the input data generation part 314, the teacher data generation part 315 and the neural network selection part 313 at appropriate timings so that the respective data are sent to the neural network 311. Further, the neural network learning control part 312 stores processing results in the quality determination rule database DB5 via the neural network selection part 313 .

6 shows a specific example of the control result quality determination on a control method in the mold control of a Sendzimir plant. 6 shows a control result quality determination for each of the forms control methods as in 26 shown.

Although the neural network 171 in the control execution unit 20 as in FIG 3 shown and the neural network 311 in the quality determination rule learning unit 22 as in FIG 5 shown have the same concept, they differ from each other in the following point.

That is, the neural network 171 in the control execution unit 20 has a predetermined content. Specifically, the neural network 171 obtains the control output quality determination estimate S9 or S11 as a corresponding output upon receiving the input data S1 and the selective control operation set S8 or the control operation set S3. That is, the neural network 171 is a neural network used for one-way processing.

On the other hand, the neural network 311 in the quality determination rule learning unit 22 satisfies an input/output relationship among the input data S1, the input data S12c of the control output set S0, and the teacher data S13c when such data is set as learning data.

A specific example of a plant control method for mold control of a Sendzimir plant is described below. The shape control is described on the assumption that the following specifications A and B are used.

The specification A is related to the priority and has priority information in a disk width direction. For example, in the shape control, the shape is often difficult to control to have a target value over the entire area in the plate width direction in terms of mechanical properties. Therefore, the following two specifications A1 and A2 become priority in the Plat width direction. Of these, the priority specification A1 specifies that a disk end has priority. The priority specification A2 specifies that the center has priority. The control is carried out according to the two priority orders of the specifications A1 and A2. That is, when a plant control system performs control, one of the priority specifications A1 and A2 is taken into account.

Specification B relates to a response to a previously known condition. For example, since a relationship between the shape pattern and the control method varies depending on various conditions, the specification B needs to be divided into a classification of a plate width specification B1 and a steel grade specification B2. The variation of each specification varies the degree of influence on a shape of a shape operation end.

The control object plant 1 of this example is the Sendzimir plant and result data matches shape results. The Sendzimir mill features a cluster for cold rolling a hard material such as stainless steel. The Sendzimir mill uses a small diameter work roll to apply a high reduction to the hard material. Thus, the Sendzimir plant is difficult to control to produce a flat steel plate. As a measure for this, the Sendzimir plant uses a multi-roller structure or various mold control parts.

The Sendzimir mill typically has upper and lower first intermediate rolls, each single conical to be displaceable, six upper and lower nip rolls and two upper and lower rolls designated AS-U. In the following example, detection data from a shape detector is used as the result data Si on a shape, and a shape deviation that is a difference from a target shape is used as the input data S1. The control operation amount S3 is assumed to be the roll shift amount of each of No. 1 to No. n AS-Us or each of the upper and lower first intermediate rolls.

7 shows the configuration of a neural network used for the shape control of the Sendzimir factory. In 7 For example, the neural network represents the neural network 101 for the control rule execution part 10. Alternatively, the neural network represents the neural network 111 for the control rule learning part 802. The neural network 101 and the neural network 111 have the same architecture.

In the case of Sendzimir work shape control, the result data Si from the control object facility 1 is Sendzimir work result data including shape detector data. (Among other things, it is assumed that a shape deviation is output as a difference between an actual shape and a target shape.) The control input data generating part 2 receives a normalized shape deviation 201 and a shape deviation level 202 as the input data S1. As a result, an input layer of the neural network 101 or 111 consists of the normalized shape error 201 and the shape error level 202. Although the shape error level 202 is represented by the input layer of the neural network in 7 is received, the neural network may change depending on stages.

An output layer of the neural network 101 or 111 consists of an AS-U operation degree 301 and a first intermediate operation degree 302 corresponding to the AS-U and the first intermediate roller each of which is a shape control operation end of the Sendzimir plant. In the AS-U, each degree of operation has an AS-U opening direction (a direction in which a nip (gap between the upper and lower rolls of the rolling mill) increases) and an AS-U closing direction (a direction in which the nip decreases). ) for each AS-U.

In the first intermediate roll, each degree of operation has a first intermediate roll opening direction (a direction in which the first intermediate roll moves outward from the center of the rolling mill) and a first intermediate roll closing direction (a direction in which the first intermediate roll moves toward the center of the rolling mill). moved towards) for each of the upper and lower first intermediate rolls. For example, assuming that the shape detector has 20 zones and the shape error level 202 has three levels (large, medium, small), the input layer has 23 inputs. Assuming that the number of AS-U saddles is seven, and assuming that the upper and lower first intermediate rollers are each slidable in the plate width direction, the output layer 14 has AS-U operation grades 301 and four first intermediate operation grades , i.e. 18 degrees of operation in total. The number of intermediate layers and the number of neurons of each layer are fixed in time. In the shape control operation end of the Sendzimir work, which is the output layer, the output of the neural network is arranged so that two kinds of outputs, that is, outputs in plus and minus directions, are made to a single control operation end.

8A and 8B each show the shape deviation and a control method in this example.

8A shows a control method for a large shape deviation while 8B shows a control method for a small shape deviation. In 8A and 8B the height direction (vertical axis direction) indicates the magnitude of the shape deviation, the horizontal axis direction indicates a board width direction, both ends of the board width indicate board ends, and the center of the board width indicates the board center.

As in 8A when the shape deviation is large, the correction of the overall shape is performed before the correction of the local shape deviation in the plate width direction.

In the case of a small deviation in shape, however, as in 8B shown, the local shape deviation is preferably reduced.

Since the control method must thus be changed according to the magnitude of the shape deviation, the shape deviation stage 202 is provided for the neural network 101 or 111 as in FIG 7 plotted to determine the magnitude of the shape deviation. A shape error normalized to, for example, 0 to 1 is preferably used regardless of the magnitude of the shape error. This is exemplary only and the shape mismatch can be sent directly to the input layer of the neural network or the neural network itself can be modified depending on the magnitude of the shape mismatch. For example, two neural networks between a large shape deviation case and a small shape deviation case may be prepared and used separately.

In the plant controller of this example, the neural network 101 or 111 having the configuration described above is used in 7 is made to learn an operation method for the shape pattern, and shape control is performed using the neural network that has learned the operation method. Neural networks with the same configuration can have different properties depending on learning conditions and can provide different control outputs for the same shape pattern.

Therefore, a plurality of neural networks are used selectively depending on other conditions of shape results, enabling the optimal control to be determined for each of various conditions. This satisfies specification B. The configuration of previously described 2 shows a specific example in performing such a specification.

In particular, in the configuration example of 2 the neural network 101 used in the control rule executing part 10 is registered in the control rule database DB1, while various neural networks are prepared depending on rolling results, mill operator names, steel grades of material to be rolled, slab widths, and the like. The neural network selecting part 102 selects a neural network that satisfies a condition at each time and sets the neural network to the neural network 101 in the control rule executing part 10 .

The neural network selecting part 102 preferably uses, as the condition at each time, slab width data from the result data Si of the control object installation.1 and selects a neural network accordingly. The neural networks used here can have a different number of intermediate layers and/or a different number of units of each layer if the input layer and the output layer are as in 7 available.

9 Fig. 12 shows a configuration of the control input data generation part 2 for generating the data S1 (normalized shape deviation 201 and shape deviation level 202) for sending to the input layer of the neural network 101 or 111.

The control input data generation part 2 receives, as the result data Si, shape detection data from the shape detector that detects a plate shape during rolling by the Sendzimir mill constituting the control object plant 1 . In addition, the control input data generation part 2 obtains a shape deviation peak-to-peak (PP) value SPP, which is a difference between the maximum and the minimum of detection results of each shape detection zone, by a shape deviation PP value calculation part 210.

A shape deviation level calculation part 211 classifies the shape deviation into three levels of large, medium and small with the shape deviation PP value SPP. The shape corresponds to the distribution of the strain rate of the material to be rolled in a plate width direction, and I-UNIT for expressing the strain rate in 10-5 unit is used as a unit for the shape. For example, the shape deviation is classified as expressed by the following numerical formulas.

Here, the shape deviation is classified so that the shape deviation level is “Large=1, Medium=0, Small=0=” to construct numerical formula 1, “Large=0, Medium=1, Small=0” to construct numerical formula 2 and "Large = 0, Medium = 0, Small = 1" to create the numerical formula is 3. The shape deviation of each zone is normalized using SPM, assuming that SPM = SPP is true.

$$S_{PP}\ge 50I-uNIT$$

$$50I-uNIT>S_{PP}\ge 10l-uNIT$$

$$10l-uNIT>S_{PP}$$

In this way, the control input data generating part 2 generates the normalized shape deviation 201 and the shape deviation level 202 as the input data for the neural network 101. The normalized shape deviation 201 and the shape deviation level 202 collectively correspond to the input data S1 for the control rule executing part 10.

10 shows a configuration of the tax expense calculation part 3.

The control output calculation part 3 generates the control operation set S3, which is an operation command for each shape control operation end, from the control operation end operation command S2 as an output from the neural network 101 in the control rule execution part 10. In a shape control of the Sendzimir plant, the control operation end operation command corresponds to S2 the AS-U operation level 301 and the first intermediate operation level 302.

10 12 shows, while there are a plurality of AS-U operation degrees 301 and a plurality of first intermediate operation degrees 302, a data sample for each of the degrees 301 and 302, and each data consists of a pair of data, i.e., an opening direction degree and a closing direction degree, respectively.

In the control output calculation part 3, the received AS-U operation degree 301 has an AS-U opening direction output and an AS-U closing direction output, and a difference between the outputs is calculated by a subtracter 303. A multiplier 304 multiplies the output of the subtracter 303 by a conversion gain G _ASU , thereby generating and outputting an operation command for each AS-U. Since the control output for the AS-U corresponds to the AS-U position change amount (in length), the conversion gain is. G _ASU a gain in the conversion from degrees to the amount of position change.

Likewise, the received first intermediate operation degree 302 has an intermediate outside first output and an intermediate inside first output, and a difference between the outputs is calculated by a subtractor 305 . A multiplier 306 multiplies the output of the subtracter 305 by a conversion gain G _ACT , thereby generating and outputting an operation command for each first intermediate roll shift. Since the control output for the first intermediate roll shift corresponds to the first intermediate roll shift position change amount (in length), the conversion gain G _IST is a gain in the conversion from degrees to the position change amount.

In this way, the control expense calculation part 3 can calculate the control operation amount S3. The control operation amount S3 includes the AS-U position change amount from #1 to #n (n is the number of saddles of the AS-U roll), the upper inter-shift first position change amount, and the lower inter-shift first position change amount.

11 Fig. 12 shows a configuration of the neural network used in the quality determination of a shape control result of the Sendzimir work when used in the control output quality determination re gel executing part 17 or quality determination rule learning part 31 is used. In this case, although the neural network represents the neural network 171 used for the control output quality determination rule execution part 17 or the neural network 311 used for the quality determination rule learning part 31, such neural networks have the same architecture.

It will use the same signals as the respective inputs for the 7 is used as the input data for the normalized shape error 201 and the shape error level 202. Further, the control operation set S3 or the selective control operation set S8 described below is used as an input for the input layer. The control operation set S3 or the selective control operation set S8 consists of a position change set of each control operation device.

The output layer outputs a value of an estimate of the quality determination of a control result when the control operation set S3 or the selective control operation set S8 is output in response to the input data S1. The number of intermediate layers and the number of neurons of each layer are fixed in time.

12 shows an operation amount calculation method of the re-seek operation amount calculation part 33. The re-seek operation amount calculation part 33 uses the control-output quality determination estimate S9 output from the control-output quality determination rule execution part 17 to calculate the re-seek control operation amount S12 in the following line.
In particular, in the case of a large value of the control output quality determination estimate S9, the quality determination of the control operation is well estimated, and thus the reseek operation amount calculation part 33 finely adjusts the reseek operation amount.
In the case of a small value of the control output quality determination estimate S9, the quality determination of the control operation is poorly estimated, and thus the re-search operation amount calculating part 33 largely changes the control operation to re-search an appropriate operation method.

$$\text{If}\left(\text{th >}=\text{S9}\right)\text{THEN Crand}=\text{Cref}+\mathrm{\gamma }\text{* th2 * G}$$

In den Formeln stellen β und γ jeweils einen zwischen -1 und 1 erzeugten Zufallswert dar. Zusätzlich stellt th1 einen Grad der Feinanpassung dar. Beispielsweise wird im Falle einer Feinanpassung in einem Bereich von ±10 % eines ursprünglichen Befehls th1 auf 0,1 festgelegt.According to the foregoing line, formulas for finding a re-search operation set Crand are set as follows. $$\text{If}\left(\text{S9 > th}\right)\text{THEN Crand}=\text{cref *}\left(1+\mathrm{\beta }\text{*th1}\right)$$

$$\text{If}\left(\text{th >}=\text{S9}\right)\text{THEN Crand}=\text{cref}+\mathrm{g}\text{*th2* G}$$

In the formulas, β and γ each represents a random value generated between -1 and 1. In addition, th1 represents a degree of fine adjustment. For example, in the case of fine adjustment in a range of ±10% of an original command, th1 is set to 0.1.

Further, th2 represents a fixed degree of change when a surgical procedure changes greatly. For example, if th2 is set to 0.1, an offset of 10% is added to an original command, and thus the operational polarity may change or a command may be issued to a device that is essentially not operating.

For each of the values of β and γ, a different value is used for each operation device, and the operation amount of each device is changed independently. G represents a maximum operating position control command from each control operation device and is multiplied by the previous command percentage so that the percentage is converted into an operating position control command.

$$\text{ELSE C''ref}=\text{Crand}\text{,}\mathrm{\alpha =}\text{0}$$

The control-issuance operation procedure selection part 18 selects the control operation set S3 or the re-search control operation set S12 for output as the selective control operation set S8. Whether the control operation set S3 or the re-search control operation set S12 is selected is stochastically determined, and a probability Prand of using the re-search control operation set S12 is made user-settable in a range of 0-1. The selection is determined by the following formula using a value 5 which randomly takes on a value from 0 to 1. $$\text{IF}\left(\mathrm{\delta >}\text{Prand}\right)\text{THEN C''ref}=\text{cref}\text{,}\mathrm{\alpha =}\text{1}$$

$$\text{ELSE C''ref}=\text{crand}\text{,}\mathrm{\alpha =}\text{0}$$

In the formula, C"ref represents the selective control operation amount S8 for sending to a subsequent calculation part by the control output operation method selection part 18. A common value for calculations of operation amounts from all devices is used as δ, and the operation amounts from all devices are on α represents the control method selection flag S14 and is 1 in the case of selecting the control operation set S3 and 0 in the case of selecting the re-search control operation set S12. The control method selection flag S14 is sent to the following calculation part together with the selective control operation set S8 sent In setting Prand, Prand is set to 0 when it is desirable to have no risk of random operation for a plant in control of the equipment concerned, and is set to a rate other than 0 when it is desirable to have a retrieval operation set to improve e to be output in a control rule.

13 shows a configuration of the control expense determination part 5.

The control expense determination part 5 consists of a rolling appearance determination part 501 and a shape correction quality determination part 502. The control expense determination part 5 obtains the result data Si from the control object facility 1, the control operation amount S3 from the control expense calculation part 3, and data in the expense determination database DB3 and supplies the control operation amount output correctness data S4 to the control operation end.

When the selective control operation set S8 calculated by the control output operation method selection part 18 is sent to a rolling mill that is the control object plant 1, the control output determination part 5 with the foregoing configuration says a shape change by sending the selective control operation set S8 to a known model of the control object plant 1 ahead. The known model of the control object plant 1 is the rolling appearance model 501 mentioned here. In this prediction, when the shape is expected to deteriorate, the control output determination part 5 suppresses the control operation amount output S0 to prevent the shape from being greatly deteriorated.

Described in detail, the control output determination part 5 inputs the selective control operation amount S8 to the rolling appearance model 501, predicts the shape change caused by the selective control operation amount S8, and calculates shape deviation correction amount prediction data 503.

The control expense determination part 5 in turn adds the shape deviation correction amount prediction data 503 to shape detection data Si from the control object facility 1, and thus obtains shape deviation prediction data 505 and evaluates the shape deviation prediction data 505. Thus, the control expense determination part 5 can predict how the Shape changes when the control operation set S3 is sent to the control object facility 1. At this time, the shape detection data Si corresponds to current shape deviation result data 504.

In the control output determination part 5, the shape correction quality determination part 502 determines whether the shape changes for better or worse based on the current shape deviation result data 504 and the shape deviation prediction data 505, and obtains the control operation amount output correctness data S4.

Specifically, the shape correction quality determination part 502 performs quality determination of the shape correction as follows. First, as described with the shape control priority specifications A and B, a weight coefficient w(i) in the plate width direction is set in advance for each of the specifications A1 and A2 in the output determination database DB3 for considering the control priority in the plate width direction. Such a weight coefficient is used to determine the quality of the shape change using an evaluation function J such as the following numerical formula 4. In the numerical formula 4, w(i) represents the weight coefficient, εfb(i) represents the shape deviation result data 504, εest(i) represents the shape deviation prediction data 505, i represents a shape detector zone, and rand represents a random number expression.

$$J=\frac{1}{n}{\displaystyle \sum _{}^n{\left(\dot{w}\left(i\right)\cdot efb\left(i\right)\right)}^2-\frac{1}{n}{\displaystyle \sum _{i=1}^n{\left(w\left(i\right)\cdot e_{est}\left(i\right)\right)}^2+\mathrm{right}an\mathrm{i.e}}}$$

In the case of using the evaluation function J of the numerical formula 4, the evaluation function J is positive when the form is getting better, while it is negative when the form is getting worse. In addition, rand is the random number expression and randomly changes an evaluation result of the evaluation function J. As a result, even if the shape deteriorates since the evaluation function J is positive in some cases, a relationship between the shape pattern and the control method can also be learned in the case , in which the rolling appearance model 501 is incorrect.

The random number expression rand is timely changed so that the maximum is increased when the model of the control object plant 1 is uncertain, such as at the start of test operation, while it is 0 when the control method has been learned to some extent and stable control is desirably performed becomes.

The shape correction quality determination part 502 calculates the evaluation function J and outputs the control operation amount output correctness data S4, while setting the control operation amount output correctness data S4 to 1 (allowable) when J≥0 and setting the control operation amount output correctness data S4 to 0 (incorrect) when J<0 be determined.

As previously described, the control output 17 receives the normalized shape error 201, the shape error level 202 and the selective control operation set S8, and outputs the control result quality determination estimate S11. The control result quality determination estimate S11 is 1 in the case of an estimate better than a control result, and is 0 in other cases.

The control-output suppression part 4 determines whether to send the control-operation-amount output S0 to the control-object facility 1 depending on the control-operation-amount output correctness data S4 which is a determination result of the control-output determination part 5 and the control-result quality determination estimate S11 The control-operation-amount output correctness data S4 includes AS-U position change amount output #1 to #n, upper first inter-shift position change amount output, and lower first inter-shift position change amount output, and is determined according to the following condition.

       IF (control method selection flag = 1) THEN IF (control operation set output correctness data S4 = 0 OR control result quality determination estimate S11 <= thprot) THEN AS-U position change set output #1 to #n = 0 upper first inter-shift position change set output = 0 lower first inter-shift position change amount output = 0 ELSE AS-U position change amount output #1 to #n = AS-U position change amount #1 to #n upper first inter-shift position change amount output = upper first inter-shift -position change amount lower first inter-shift position change amount output = lower first inter-shift position change amount END IF ELSE IF (control operation amount output correctness data S4 = 0 OR control result quality determination estimate S11 <= thprot) AND (PTRIAL < η) ) THEN AS-U position change amount output No. 1 to No. n = 0 upper first inter-shift position change change quantity output = 0 lower first inter-shift position change quantity output = 0 ELSE AS-U position change quantity output #1 to #n = AS-U- Position change amount #1 to #n inter-shift upper first position change amount output=inter-shift upper first position change amount lower inter-shift first position change amount output=inter-shift first position change amount lower END IF END IF
    

A reference value for output suppression is set as thprot based on an estimate of the control result quality determination. In particular, at the first start with insufficient operation data of a plant, the estimation accuracy of the quality determination is considered to be low, and thus the reference value is set low in order to avoid a serious suppression of expenditure based on the quality determination estimation.
On the other hand, when the quality determination estimate has improved by storing operation result data sufficiently, the reference value is set high to increase an effect of the expense suppression based on the quality determination estimate of a control result. With the quality determination estimation, a quality determination rule accuracy S15 is received from a quality determination rule accuracy checking part 34 in the quality determination rule learning unit, whereby the quality determination accuracy is determined based on a check result of the estimation accuracy of a currently used quality determination rule.

In addition, η is a variable that can take a random value in a range from 0 to 1, and PTRIAL indicates a probability of releasing the output suppression and sending a retrieval operation to a facility. Since a case of the control method selection flag S14 of 0 includes a case of checking an effect of a control method in an unknown area, the re-seeking operation is sent to the facility at a certain probability regardless of the output suppression at the facility.

The control execution unit 20 performs the above-described calculation based on the result data Si from the control object facility 1 (rolling mill) and sends the control operation amount output S0 to the control object facility i (rolling mill), thus performing shape control. The control method learning unit 21 uses the data used in the control execution unit 20.

The operation performed by the learning data generating part 801 is described below.

As in 1 shown, the learning data generating part 801 generates the teacher data S7a for the neural network 111 used in the control rule learning part 802 from the control operation end operation command S2, the selective control operation set S8, the control method selection flag S14, and a determination result of the control output suppressing part (control operation set output correctness data S4 ) based on the control result quality determination estimate S11 from the control output quality determination rule execution part 17.

In this case, the teacher data S7a corresponds to the AS-U operation degree 301 or the first intermediate operation degree 302 output from the neural network output layer 111 as in FIG 7 shown. A learning data generating part 7 uses the control operation end operation command S2 (AS-U operation degree 301 or first intermediate operation degree 302) output from the neural network 101, with the AS-U position change amount output No. 1 to No. n the selective control operation set S8, the upper first inter-shift position change quantity output and the lower first inter-shift position change quantity output for generating the teacher data S7a for the neural network 111 used in the control rule learning part 802.

For describing the operation of the learning data generating part 801, FIG 14 a mutual relation of data from different parts and/or characters in the tax expense calculation part 3 as in 10 shown. In 14 the AS-u operation degree 301 is typically represented as the control operation end operation command. S2 output from the neural network 101 on the assumption that the data on the operation degree plus side is OPref, the data on the operation degree minus side is OMref, a conversion gain is G, and the control operation amount S3 is Cref.

A difference between operation degree plus data OPref and operation degree minus data OMref is used by a subtracter 701, and the difference is multiplied by the conversion gain G by a multiplier 702 to obtain the control operation amount output Cref. The control operation amount output Cref is supplied to the control output operation method selection part 18 to obtain a selected operation command value C"ref.

For simplification, as an output from the output layer of the neural network 101 in the control rule execution part 10, operation degrees of operation degree plus side, operation degree minus side, and a control operation disturbance generation part 16 occurring at random are assumed here as operation degree random numbers. The control operation amount output S0 for the control operation end is adopted as an operation command value.

15 shows processing steps and processing contents of the learning data generating part 7.

In a first processing step 71, the selection control operation set S8, which is an output value of the control output operation method selection part 18, is referred to as the operation command value C'' ref.

In a second processing step 72, an operation command C" ref is corrected to C'ref according to the control result quality determination estimate S11 and the control operation amount output correctness data S4. Specifically, the corrected value C'ref of the operation command value C" ref is determined by the numerical formula 7 when the control result quality determination estimate S11=0 or the control operation amount output correctness data S4=0 is obtained, and is determined by the numerical formula 8 when the control result quality determination estimate S11=1 and the control operation amount output correctness data S4=1 are obtained.

$$\begin{array}{l}\text{IF C''ref > 0 THEN C'ref =C''ref}-\Delta \text{cref}\\ \text{IF C'' ref < 0 THEN C'ref =C'' ref}+\Delta \text{cref}\end{array}$$

$$\begin{array}{l}\text{IF C''ref > 0 THEN C'ref =C''ref}+\Delta \text{cref}\\ \text{IF C'' ref < 0 THEN C'ref =C'' ref}-\Delta \text{cref}\end{array}$$

In a processing step 73, the corrected operation degree quantity ΔOref is found by the numerical formula 9 and the numerical formula 10 with the corrected operation command value C'ref.

$$\text{C}\text{'}\mathrm{right}ef=G\cdot \left(\left(\text{OP}\mathrm{right}ef+\Delta \text{O}\mathrm{right}ef\right)-\left(\text{OM}\mathrm{right}ef-\Delta \text{O}\mathrm{right}ef\right)\right)$$

$$\Delta \text{O}\mathrm{right}ef=\frac{1}{2}\left(\frac{\text{C'}\mathrm{right}ef}{G}-\left(\text{OP}\mathrm{right}ef-\text{OM}\mathrm{right}ef\right)\right)$$

In a processing step 74, teacher data OP'ref and teacher data OM'ref for the neural network 111 are determined by the numerical formula 11.

$$\begin{array}{l}\text{OP'}\mathrm{right}ef=\text{OP}\mathrm{right}ef+\Delta \text{O}\mathrm{right}ef\\ \text{OM'}\mathrm{right}ef=\text{OM}\mathrm{right}ef-\Delta \text{O}\mathrm{right}ef\end{array}\}$$

Thus, the learning data generating part 7 calculates the operation command value correction value C'ref from the operation command value C"ref each sent to the control object facility 1 as in FIG 14 shown depending on the control result quality determination estimate S11 from the control output quality determination rule execution part 17 and the control operation amount output correctness data S4 for the control output suppressing part 4.
Specifically, in the case of the control result quality determination estimate S11=1 and the control operation amount output correctness data S4=1, the operation is determined good, and thus the learning data generating part 7 increases the operation command value by ΔCref in the same direction.

On the other hand, in the case of the control result quality determination estimate S11=0 or control operation amount output correctness data S4=0, the operation is determined bad and thus the learning data generating part 7 decreases the operation command value by ΔCref in an opposite direction. Since the conversion gain G is predetermined and known, when a value on the positive side of the degree of operation and a value on the negative side of the degree of operation are known, the corrected quantity can be obtained ΔCref can be determined. A corresponding value is determined in advance, for example by simulation, and defined as ΔCref. According to the above method, the teacher data OP'ref and OM'ref used in the control rule learning part 802 can be found by the numerical formula 11.

Although 14 represents a simple example, all the AS-U operation degrees 301 for the AS-Us #1 to #n and all the first intermediate operation degrees 302 for the upper first intermediate-roller shift and the lower first intermediate-roller shift are subjected to an operation so that they are stored in teacher data ( AS-U operation degree teacher data and first intermediate operation degree teacher data) for the neural network 111 used in the control rule learning part 802 flow in.

16 shows an example of the data stored in a learning data database DB2.
For learning the neural network 111, a combination of many input data S8a and many teacher data S7a is required. Therefore, the teacher data S7a generated by the learning data generating part 7 is combined in the control execution unit 20 with the input data S1 (S8a) received from the control rule executing part 10 and stored in the learning data database DB2 in a form of a set of learning data. At this time, the teacher data S7a includes the AS-U operation level teacher data and first intermediate operation level teacher data. The input data S1 (S8a) includes the normalized shape error 201 and the shape error level.

The in the plant control system of 1 used various databases DB1, DB2, DB3 and DB4 are managed and operated in cooperation with each other by a neural network management table TB.

17 Fig. 12 shows a configuration of the neural network management table TB.
The neural network management table TB is divided into A1 and A2 depending on specifications of (B1) plate width, (B2) steel grade, and control priority. For example, four divisions of three feet wide, one meter wide, four feet wide and five feet wide are used as (B1) plate width, while ten divisions of steel grades (1) to (10) are used as the steel grade. The control priority specification A includes two specifications Ai and A2. in this case, there are 80 partitions, and 80 neural networks are used separately depending on rolling conditions.

The neural network learning part 112 stores learning data as a combination of the input data and the teacher data as in FIG 16 shown in the learning data database DB2 as in 18 shown, while a reference of the learning data to a corresponding number of the neural network according to the in 17 illustrated table TB for managing a neural network.

When the control execution unit 20 performs the shape control for the control object installation 1, two sets of learning data are generated. The reason for this is as follows: since the control result quality determination is made on the same input data or the same control output using two criteria of specifications A1 and A2 of control priority, two kinds of teacher data are generated. When certain sets (e.g., 200 sets) of teacher data are stored or newly stored in the learning data database DB2, the neural network learning control part 112 instructs the neural network 111 to learn.

The control rule database DB1 stores a plurality of neural networks corresponding to the management table TB as in FIG 17 shown. The neural network learning control part 112 specifies the neural network number required for learning, and the neural network selecting part 113 extracts a corresponding neural network from the control rule database DB1 and sets this neural network as the neural network 111 Celebration.

The neural network learning control part 112 instructs the input data generation part 114 and the teacher data generation part 115 to extract input data and teacher data, respectively, corresponding to the respective neural networks from the learning data database DB2, and executes the learning of the neural network 111 using such input data and teacher data. Although various approaches are provided for the method of learning a neural network, any approach can be used.

When the neural network 111 has been learned, the neural network learning control part 112 rewrites the neural network 111 as the learning result to a position of the corresponding one the neural network number of the control rule database DB1 so that the learning is completed.

Learning can be performed together on all defined neural networks as in 17 be performed at a fixed time interval (e.g. every day) or may be performed at an appropriate point on only one number of the neural network for which certain sets (e.g. 100 sets) of new learning data are stored.

The operation of the quality determination rule learning unit 22 is described below.

The quality determination rule learning unit 22 uses data as a delayed version of the data used in the control execution unit 20 . Such a delay Z ^-1 means e-TS to indicate that data is delayed by a predetermined time T .

The control object installation 1 has a timing that results in a time lag due to the control operation set output S0 until the result data is changed. Thus, the learning is performed using result data at a point after the delay T after the start of the control operation.

In the shape control, T is preferably set to about 2 to 3 s because several seconds are required before a shape measuring device detects a shape change after an operation instruction is sent to the AS-U or the first intermediate roll. Since the delay time also varies depending on a type of shape detector or the rolling speed, the optimum time for the shape change due to a change in a control operation end is preferably set as T.

19 12 shows the operation of the control result quality determination part 6. A shape change quality determination part 602 uses a quality determination evaluation function Jc expressed by the numerical formula 12.

$$Jc=\frac{1}{n}{\displaystyle \sum _{i=1}^n{\left(wc\left(i\right)\cdot e_{fb}\left(i\right)\right)}^2-\frac{1}{n}{\displaystyle \sum _{i=1}^n{\left(wc\left(i\right)\cdot e_{last}\left(i\right)\right)}^2}}$$

In the numerical formula 12, εfb(i) represents the shape deviation result data included in the result data Si, ε last (i) represents a shape deviation result data previous value, and wC(i). is a weight coefficient in a plate width direction for quality determination. The weight coefficient wC(i) for quality determination is determined by the quality determination database DB4 depending on the control priority specifications A1 and A2. The control result quality is determined with the quality determination evaluation function Jc.

A threshold upper limit LCU and a threshold lower limit LCL are set in advance under a threshold condition of LCU≧0≧LCL. In a result of the comparison with the quality determination evaluation function Jv, when Jc > LCU, it is assumed that the control result quality data S6 is 0 (the shape has deteriorated), and when Jc < LCL, it is assumed that the control result quality data S6 is 1. (The form has improved.)

As described above, since the weight coefficient wc(i) in the plate width direction varies depending on the control priority specifications A1 and A2, the quality determination evaluation function Jc also varies depending on the specifications. Thus, a determination result of the control result quality data S6 also presumably varies. Therefore, the quality determination rule learning unit 22 performs determination of the control result quality data S6 for each of the two kinds of the control priority specifications A1 and A2.

The control result quality data S6 is directly used as the teacher data S13a for the neural network 311 used in the quality determination rule learning part 31. FIG.

20 shows a data example stored in the learning data database DB6.

For learning the neural network 311, a combination of many input data S12a and many teacher data S13a is required. Therefore, the control result quality determination part 6 generated teacher data S13a (control result quality data) in the learning data database DB6 in a form of a set of learning data in combination with the data S12a as the delayed version of the input data S1 (normalized shape deviation 201 and shape deviation level) received from the control rule execution part 10 in the control execution unit 20 ) saved.

At this time, the learning data is stored in the verification data database DB7 at a constant ratio to be usable for quality determination rule verification by the quality determination rule accuracy verification part 34 .

The quality determination rule accuracy checking part 34 has a neural network that performs calculation only in one direction like the control output quality determination rule execution part 17 . The quality determination rule accuracy checking part 34 extracts test data from the checking data database DB7 and calculates an error between output data obtained by inputting the input data of the test data itself in the neural network and the output data of the test data itself. For example, the quality determination rule accuracy checking part 34 calculates the average of errors of all pieces of test data as the quality determination rule accuracy S15 for the quality determination rule.

While the plant control system of 1 uses DB5 and DB6 databases 21 a configuration of the neural network management table TB for managing and operating the databases DB5 and DB6 in cooperation with each other. That is, the management table TB includes a specification management table.

Specifically, the management table TB is as in 21 shown depending on specifications to (B1) plate width, (B2) steel grade and control priority A1 and A2 divided. For example, four divisions of three feet wide, one meter wide, four feet wide and five feet wide are used as (B1) plate width, while ten divisions of steel grades (1) to (10) are used as the steel grade. The control priority specification A includes two specifications A1 and A2. In this case, there are 80 partitions, and 80 neural networks are used separately depending on rolling conditions.

The neural network learning part 312 stores learning data as a combination of the input data and the teacher data as in FIG 20 is represented in the learning data database DB6 as shown in Fig. 22, while the learning data is referred to a corresponding neural network number according to the neural network management table TB in Fig. 21.

When the control execution unit 20 performs the shape control for the control object installation 1, two sets of learning data are generated. The reason for this is as follows: since the control result quality determination is made on the same input data or the same control output using two criteria of specifications A1 and A2 of control priority, two kinds of teacher data are generated. When certain sets (e.g., 200 sets) of teacher data are stored or newly stored in the learning data database DB6, the neural network learning control part 312 instructs the neural network 311 to learn.

The quality determination rule database DB5 stores a plurality of neural networks corresponding to the management table TB as in FIG 21 shown. The neural network learning control part 312 specifies the neural network number required for learning, and the neural network selecting part 313 extracts a corresponding neural network from the control rule database DB5 and sets this neural network as the neural network 311 fixed. The neural network learning control part 312 instructs the input data generation part 314 and the teacher data generation part 315 to extract input data and teacher data, respectively, corresponding to the respective neural networks from the learning data database DB6, and executes the learning of the neural network 311 using such input data and teacher data. Although various approaches are provided for the method of learning a neural network, any approach can be used.

When the neural network 311 has been learned, the neural network learning control part 312 rewrites the neural network 311 as the learning result to a position of the corresponding neural network number of the control rule database DB6 so that the learning is completed.

Learning can be shared at all through the in 21 illustrated management table TB defined neural networks are performed at a fixed time interval (for example, every day). Alternatively, the learning may be performed at an appropriate point on only a neural network of a neural network number for which certain sets (e.g., 100 sets) of new learning data are stored.

In addition, the input data of the quality determination rule includes rolling results, a steel grade, and a plate width, which makes it possible to perform neural network learning while including a difference in quality determination standard. This eliminates the need for changing the quality determination rule depending on rolling conditions while executing the quality determination rule.

As described above, in order to improve the control rule of the control object facility 1, when a control operation for providing a good control result has not been learned, the control operation is greatly changed. When a good control result is obtained, the changed control operation is used as a new control operation method. If a control operation to provide a good control result has not been learned, the control operation is not changed or only slightly changed. If a good control result is thereby obtained, this control operation is effectively used as a new control operation method.

A combination of a shape pattern, a control operation and the quality of a control result is learned based on actual equipment data, whereby a model capable of estimating the quality of a control result more accurately than a simulator using a machine model can be obtained can be continuously adapted to the last system situation through regular automatic learning.

A model that estimates the quality determination of a control result can be used to improve the reliability of a function for suppressing the control output to a plant, in contrast to existing methods where such a function was only performed with a simple machine model.

While generation of control rule learning data is performed by quality determination of a control result in the existing methods, a model for estimating the quality determination of a control result is used, thereby suppressing the influence of noise contained in plant data, enabling fine adjustment that has a small effect supplies, which is also contained in learning data objects. In addition, in this example, erroneous determination of a control effect is prevented, and thus variation in learning data is suppressed, resulting in stable control performance.

The control rule database DB1 stores a neural network for use in the control execution unit 20. When the neural network to be stored is subjected only to initialization with a random number, a certain time is required before learning the advances of the neural network until appropriate control is enabled becomes. Therefore, when a control part is provided for the control object plant 1, learning of a control rule is previously performed in a simulation based on a control model of the control object plant 1 known at that point. The neural network learned by a simulator is stored in a database in advance, whereby the performance can be controlled to some degree from the start of control object creation.

Alternatively, the quality determination rule is learned by the quality determination rule learning unit 22 based on result data of operation data of the equipment concerned, whereby the control rule can be learned without controlling the equipment concerned, and thus the control of the power can be performed to a certain extent before the control rule is applied to the control object attachment.

23 Fig. 12 shows the configuration of a plant control system of this example, which further includes a control rule evaluation unit 23 that performs evaluation processing of a control rule.

The control rule evaluation unit 23 includes a control rule quality determination data collection part 35, a control rule evaluation data calculation part 36, a control rule database update processing part 37, a control rule evaluation data database DB8 and a control rule evaluation value database DB9.

The control rule quality determination data collection part 35 receives the control output quality determination estimate S9 from the control output quality determination rule execution part 17 and receives a quality determination rule accuracy S15 from the quality determination rule accuracy checking part 34. The control rule quality determination data collection part 35 stores control rule quality determination data S16 together with a in the Control execution unit 20 used control rule number in control rule evaluation data database DB8. The control rule quality determination data S16 agrees with the control expenditure quality determination estimate S9. However, if the quality determination rule accuracy S15 is less than or equal to a certain accuracy, the control rule. Quality determination data S16 not stored in the control rule evaluation database DB8.

When the control execution unit 20 calculates a control output using a control rule, new data of the control rule quality determination data S16 is acquired and such acquired control rule quality determination data S16 is stored in the control rule evaluation data database DB8. In this case, since a large amount of data is stored for each control rule, it is determined that the control rule evaluation data database DB8 has an upper limit for data to be stored for each control rule, and when the stored data is limited to a certain amount or more grow, old data is erased to save new data.

The control rule evaluation data calculation part .36 extracts control rule quality determination data S17 stored collectively for each control rule from the control rule evaluation data database DB8 and averages the data S17 as control rule evaluation data S18 by calculation. The determined average corresponds to an evaluation value.
The control rule evaluation data S18 calculated by the control rule evaluation data calculation part 36 is stored in the control rule evaluation value database DB9. However, when the number of the control rule quality determination data is less than a certain number, the reliability of the evaluation value is low and thus no evaluation result is stored.

The neural network numbers (control rules) to be used depending on conditions are sequentially stored in the database management table TB. In turn, the control rule evaluation value database DB9 manages the evaluation values of a plurality of control rules. The control rule database update part 37 refers to the control rule evaluation value database DB9 for comparing a control rule evaluation value of the neural network number (control rule) registered in the database management table TB with control rule evaluation values of other control rules applicable to the relevant condition and updates the neural network number (control rule) in the database management table TB with a control rule having the highest evaluation value among the control rules.

Other sections of the in 23 illustrated plant control system are configured in the same way as that in 1 illustrated plant control system. In the case of the plant control system of 23 since the control rule evaluation unit 23 performs evaluation of the old result data of the control object facility 1 based on old results of the control object facility 1, the control execution unit 20 need not control the control object facility i. Specifically, the control-output suppressing part 4 does not need to supply the control-output amount S0 to the control-object facility 1 .

According to the plant control system of 23 a control rule to be evaluated is set in the control rule execution part 10, and old result data is supplied to the control rule execution part 10 as Si, whereby even if the control output is not sent to the control object facility 1, the control rule evaluation value database DB9 can be updated.

modifications

The invention is not limited to the exemplary embodiment described above and may include various modifications. For example, the exemplary embodiment has been described in detail in order to clearly explain the invention, and the invention is not necessarily limited to the embodiment having all of the described configurations.

For example, this is in 1 or 23 The plant control system illustrated is configured to include processing parts that perform various types of processing such as data generation, learning, and control. The control execution unit 20, the control method learning unit 21, the quality determination rule learning unit 22, and the control rule evaluating unit 23 as in FIG 1 or 23 each may consist of a program (software) by which a processor executes a function of each unit to enable a computer to execute the program. 24 shows a configuration example of the computer in such a case.

In particular, the computer forming the units 20 to 23 comprises as in 24 shown a Central Processing Unit (CPU) a, a Ready Only Memory (ROM) b and a Random Access Memory (PAM). c, each connected to a bus. The computer also includes a non-volatile memory d and a network interface e.

The CPU is an arithmetic processing part that reads a program code of a software for performing the processing of each of the units 20 to 23 from the ROM b and executes the program code. A variable, a parameter, or the like occurring during arithmetic processing is temporarily written in the RAM c. A large-capacity information storage such as a hard disk drive (HDD) or a solid state drive (SDD) is used as the non-volatile memory d for storing programs and data executed by the units 20 to 23 in the databases.
Although the respective units 20 to 23 may be formed by separate computers, the respective programs may be executed in a small number of computers, such as one computer, for execution at a time.

For example, a network interface card (NIC) is used as the network interface e and performs data transmission and reception to/from another unit or to/from the control object facility 1 .

In this case, information of the program performing each processing function can be stored on a recording medium such as a memory, an IC card, an SD card or an optical disk in addition to the volatile memory d such as an HDD or SDD.

The function performed by each of the units 20-23 can be performed by hardware such as a Field Programmable Gate Array (FPGA) and an Application Specific Integrated Circuit (ASIC).

The block diagram of 1 or 23 shows only control lines and information lines considered necessary for the description and does not necessarily show all control lines and information lines for a product. Almost all configurations can be considered connected.

Although the invention is applied to a Sendzimir work as the control object plant 1 in the above embodiment, the invention can also be applied to the control of any other plant. As for the control rule in the case where the invention is applied to the Sendzimir work, the control rule is described only as an example, and the invention is not limited to the embodiment.

QUOTES INCLUDED IN DESCRIPTION

This list of documents cited by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent Literature Cited

• JP 2804161 [0003, 0008]
• JP 2018180799 A [0011]
• JP 2018 A [0011]
• JP 180799 [0011]

Claims (7)

Facility control system for a control object facility, the facility control system recognizing a combination pattern of result data of the control object facility and performing control, the facility control system comprising: a control method learning unit that learns a combination of the result data of the control object creation and the control operation; a control execution unit that executes control of the control object installation according to the combination of the result data and the control operation learned by the control method learning unit; and a quality determination rule learning unit that learns a combination of the combination of the result data of the control object creation and the control operation and the control result quality, wherein the control execution unit comprises a control rule execution part that provides a control output according to a specified combination of the result data of the control object creation and the control operation, a control output quality determination rule execution part that performs quality determination of the control output according to the determined combination of the result data of the control object creation, the control operation and the control result quality, a re-search operation amount calculation part that calculates a re-operation search operation amount based on the quality determination by the control output quality determination rule execution part, and a control output suppression part that uses the quality determination by the control output quality determination rule execution part and simulation data using a control simulator to prevent the control output from being sent to the control object installation when it is determined that the result data of the control object installation has deteriorated the control output is sent to the control object system, wherein the quality determination rule learning unit comprises a control result quality determination part that determines the quality of a control result after a delay before a control effect appears in the result data when the control execution unit sends the control output to the control object facility, and a quality determination rule learning part that learns, as learning data, the quality of the control result determined by the control result quality determination part, the result data, and the control output, wherein the control method learning unit comprises a learning data generation part that uses the quality determination of the control output by the control output quality determination rule execution part and the control output to obtain teacher data, and a control rule learning part that learns the result data and the teacher data as learning data. plant control system claim 1 wherein the control method learning unit performs learning whereby combinations of result data and control operations are obtained separately for a plurality of control targets depending on states of the control object creation, and the obtained combinations of result data and control operations respectively as the specific combination of the result data of the control object creation and the control operation used in the control rule executive. plant control system claim 1 , wherein the control output quality determination rule execution part stores as a first neural network the determined combination of the result data of the control object creation, the control operation and the control result quality, the quality determination rule learning part stores as a second neural network the combination of the result data, the control operation and the control result quality, and the second neural network obtained as a result of learning by the quality determination rule learning part is used as the first neural network in the quality determination rule execution part. plant control system claim 1 wherein the quality determination rule learning unit has a quality determination rule accuracy check part, and uses a quality determination rule accuracy generated by the quality determination rule accuracy check part to change a standard of the output suppression using the control result quality by the control output suppression part. Plant control system according to one of Claims 1 until 4 , further comprising a control rule evaluation unit comprising a control rule quality determination data collecting part stored in a database Quality determination data by the quality determination rule learning part of the control execution unit and stores an accuracy check result of a quality determination rule obtained by the quality determination rule learning unit, and a control rule evaluation data calculation part that calculates control rule evaluation data on the basis of the quality determination data and the accuracy check result of the quality determination rule stored in the database, wherein the control rule evaluation unit performs an evaluation of a control rule used in the control execution unit without sending the control rule to the control object facility. A facility control method for a control object facility, the facility control method being a method of recognizing a combination pattern of result data of the control object facility and performing control of the control object facility by a computer, wherein the computer performs multiple pieces of processing comprising a control method learning processing of learning a combination of the result data of the control object installation and the control operation, control execution processing of performing control of the control object installation according to the combination of the result data and the control operation learned by the control method learning processing, and a quality determination rule learning processing of learning a combination of the combination of the result data of the control object creation and the control operation and the control result quality, wherein the control execution processing comprises a control rule execution processing of providing a control output corresponding to a specified combination of the result data of the control object creation and the control operation, control output quality determination rule execution processing of performing quality determination of the control output according to the determined combination of the control object creation result data, the control operation and the control result quality, a re-search operation amount calculation processing of calculating a re-operation search operation amount based on the quality determination by the control output quality determination rule execution processing, and a control output suppression processing of using the quality determination by the control output quality determination rule execution processing and simulation data using a control simulator to prevent the control output from being sent to the control object facility when it is determined that the result data of the control object facility has deteriorated when the control output is sent to the control object system, wherein the quality determination rule learning processing comprises control result quality determination processing of determining the quality of a control result after a delay before a control effect appears in the result data when the control output is sent to the control object facility by the control execution processing, and a quality determination rule learning processing of learning as. learning data of the quality of the control result determined by the control result quality determination processing, the result data and the control output, wherein the control method learning processing includes a learning data generation processing of using the quality determination of the control output by the control output quality determination rule execution processing and the control output to obtain teacher data, and a control rule learning processing of learning the result data and the teacher data as the learning data. A program for a control object installation, the program being a program for recognizing a combination pattern of result data of the control object installation and allowing a computer to perform control of the control object installation, the control including methods of a control method learning method of learning a combination of the result data of the control object installation and the control operation , a control execution method of performing control of the control object creation according to the combination of the result data and the control operation learned by the control method learning method, and a quality determination rule learning method of learning a combination of the combination of the result data of the control object creation and the control operation and the control result quality, the control execution method comprising a control rule execution method of providing a control output according to a certain combination of the result data of the St your object plant and the control operation, a control output quality determination rule execution method of performing a quality determination of the control output according to the determined combination of the result data of the control object creation, the control operation and the control result quality, a re-search operation amount calculation method of calculating a re-operation search operation amount on the basis of the quality determination by the Control output quality determination rule execution method, and a control output suppression method of using the quality determination by the control output quality determination rule execution method and simulation data using a control simulator to prevent the control output from being sent to the control object plant when it is determined that the result data of the control object facility have deteriorated when the control output is sent to the control object facility, the quality determination rule ler n method includes a control result quality determination method of determining the quality of a control result after a delay before a control effect appears in the result data when the control output is sent to the control object facility through the control execution method, and a quality determination rule learning method of learning as learning data of the quality of the through the Control result quality determination method determined control result, the result data and the control output, wherein the control method learning method includes a learning data generation method of using the quality determination of the control output by the control output quality determination rule execution method and the control output to obtain teacher data, and a control rule learning method of learning the result data and the teacher data as learning data.

Images