CN108687137A - Complete equipment control device, rolling mill control apparatus, control method and storage medium - Google Patents

Abstract

Learn complete equipment control device, rolling mill control apparatus, complete equipment control method, milling train control method and the storage medium having program stored therein of the best operating method for real data the present invention relates to the state deteriorating that will not make control object complete equipment.The present invention identifies the pattern of the combination of the real data of control object complete equipment for control object complete equipment to implement control, wherein having:Control method learning device learns the combination of the real data and control operation of control object complete equipment;And control executive device, according to the combination of the real data and control operation that learn, implement the control of control object complete equipment, wherein control executive device has the regular enforcement division of control, control output determination unit and control output suppressing portion, determination unit, learning data preparing department and control rule learning portion whether control method learning device has control result well.

Description

Complete equipment control device, rolling mill control apparatus, control method and storage medium

Technical field

The present invention relates in the real-time feedback control for using the artificial intelligence technologys such as neural network to carry out, complete equipments Control device and its control method, rolling mill control apparatus and its control method and the storage medium for storing program.

Background technology

In the past, it was implemented based on each in various complete equipments in order to obtain desired control result by its control The complete equipment control of kind control theory.

As an example of complete equipment, such as in milling train control, as an example of control, as with to plate Undulating state controlled shape control be object control theory, apply fuzzy control, nerve fuzzy control.It is fuzzy Control is applied to control using the shape of coolant, and in addition nerve fuzzy control is applied to the shape control of sendzimir mill System.Wherein, as shown in Patent Document 1, find out the true form pattern detected by shape detector and target shape pattern it It is poor with the similar ratio of preset reference figure pattern, using similar to ratio that it is also preset by being directed to according to this The control rule of the control operating side operating quantity performance of reference figure pattern, finds out the control output quantity for operating side, to Apply the shape control of nerve fuzzy control.Hereinafter, as the prior art, the Senghi using nerve fuzzy control is used The shape of Mir's milling train controls.

Fig. 5 shows the shape control of the sendzimir mill described in Fig. 1 of patent document 1.In sendzimir mill Shape control in use nerve fuzzy control.In this example embodiment, using pattern-recognition mechanism 51, according to by shape detector 52 The true form that detects carries out the pattern-recognition of shape, and operation true form is closest in preset reference figure pattern Which.In controlling arithmetical organ 53, the control operating side as shown in FIG. 6 by for preset shape model is used The control rule that operating quantity is constituted implements control.If more specifically describing Fig. 6, in pattern-recognition mechanism 51 operation by Shape mould of the difference (Δ ε) of shape actual result and target shape (ε ref) that shape detector 52 detects closest to 1 to 8 Which of formula (ε) selects some in 1 to 8 control method and is executed in controlling arithmetical organ 53.

But in the method for patent document 1, for access control rule, operating personnel is allowed to carry out in rolling sometimes It is manually operated to carry out the verification etc. of control rule, the change in shape opposite with anticipation is presented sometimes.That is, occurring as described above Determining control rule does not meet the situation of reality.Its reason is the research deficiency of mechanical property, the job state of milling train, machine The variation of tool condition, but if verifying whether preset control rule is best rule one by one, then being considered as Condition it is more and be difficult.Therefore, if once set at control rule, as long as no failure, it will usually keep intact.

When variation due to operating condition etc. controls rule and does not meet reality, due to control rule be it is fixed, So being difficult to export the control accuracy above to a certain degree.Once in addition, shape control acted, operating personnel just not into Row manual operation (being interference for control), so being also difficult to find new control by the manual intervention of operating personnel Rule.In turn, in the case where the rolling stock to new standard rolls, it is also difficult to accordingly set control with the material Rule.

As previously discussed, it in the control of existing shape, is controlled using preset control rule, so depositing In the problem for being difficult to Correction and Control rule.

In order to solve this problem, by as shown in patent document 2 on one side carry out shape control while make control rule with Change to machine and learning shape becomes good rule to carry out, to realize:

1) find that new control is regular while implementing shape control in rolling.

2) because new control rule and it is non-predetermined can envision, perfect forecast sometimes less than control rule be best , so control operating side is randomly made to act and be found while observing the control result to it.

Existing technical literature

Patent document 1:Japan Patent 2804161

Patent document 2:Japan Patent 4003733

Invention content

In the above prior art, representative shape is set as reference figure pattern in advance, according to expression and be directed to The control rule of the relationship of the control operating side operating quantity of reference waveform pattern is controlled.Study about control rule It is related with for the control operating side operating quantity of reference waveform pattern, directly use predetermined representative reference figure mould Formula.The problem of accordingly, there exist as only controlling the shape of specific shape model reaction.

Reference figure pattern be people previously according to knowledge related with the milling train as object, put aside shape actual result With the experience of manual interventional procedure and determination, it is difficult to enlist the services of the institute occurred in as the milling train of object and material to be rolled There is shape.Therefore, in the case where the shapes different from reference figure pattern have occurred, do not implement to control based on shape sometimes Control, form variations are not suppressed and remain, or are mistakenly identified as similar reference figure pattern, carry out the control behaviour of mistake Make, so that shape is deteriorated instead.

As previously discussed, in the control of existing shape, using preset reference figure pattern and to its control Rule and learn control rule and implement control, so there is a problem of that the raising of control accuracy is limited.

According to the above, in the present invention, providing a kind of " complete equipment control device, for the knowledge of control object complete equipment The pattern of the combination of the real data of other control object complete equipment implements control, which is characterized in that has:

Control method learning device learns the combination of the real data and control operation of control object complete equipment;And

Executive device is controlled, control object complete equipment is implemented according to the combination of the real data and control operation that learn Control,

Control executive device has:

The regular enforcement division of control, according to the determined combination of the real data and control operation of control object complete equipment To provide control output;

Control output determination unit, can the regular enforcement division of judgement control export control output, and learn to control method Device notifies that the real data and control operation are wrong;And

Control output suppressing portion is judged as being output to control object the whole series and setting will control in control output determination unit In the case that the real data of control object complete equipment deteriorates in the case of standby, prevent control being output to control Object complete equipment,

Control method learning device has:

Control result well whether determination unit, control executive device will control output be effectively outputed to control object the whole series In the case of equipment, after the time delay until control effect shows real data, judge about real data With control before compared to become well still become deterioration control result it is good whether;

Learning data preparing department, control result whether using control result well in determination unit it is good whether and control It exports and obtains teacher's data;And

Rule learning portion is controlled, real data and teacher's data are learnt as learning data,

Learnt by control method learning device, multiple control mesh are directed to according to the state of control object complete equipment It marks and obtains the combination of real data alone and control operation, the combination of obtained real data and control operation is used as control The determined combination of the real data and control operation of control object complete equipment in the regular enforcement division of system.".

In addition, the present invention is a kind of " rolling mill control apparatus, using there is complete equipment control device, which is characterized in that control Object complete equipment is milling train, and real data is the outlet side shape of milling train.".

In addition, the present invention is that " a kind of complete equipment control method, the complete equipment control method are applied to complete equipment control Device processed identifies the pattern of the combination of the real data of control object complete equipment for control object complete equipment to implement to control System, which is characterized in that the complete equipment control device has:

Control method learning device learns the combination of the real data and control operation of control object complete equipment;And

Executive device is controlled, control object complete equipment is implemented according to the combination of the real data and control operation that learn Control,

Executive device is controlled according to the determined combination of the real data of control object complete equipment and control operation to come Control output is provided, can judgement export control output, notify the real data and control to operate to control method learning device Be it is wrong, be judged as will control be output to control object complete equipment in the case of control object complete equipment In the case that real data deteriorates, prevent control being output to control object complete equipment,

Control output is effectively outputed to control object complete equipment by control method learning device in control executive device In the case of, after the time delay until control effect shows real data, judge about real data and the control System before compared to become well still become deteriorate control result it is good whether, using control result it is good whether and control It exports and obtains teacher's data, real data and teacher's data are learnt as learning data, to according to control object The state of complete equipment obtains the combination of real data and control operation alone, the reality that will be obtained for multiple control targes The combination of border data and control operation is used as the real data of the control object complete equipment in control executive device and control is grasped The determined combination made.".

In addition, the present invention is a kind of " milling train control method, using there is complete equipment control method, which is characterized in that control Object complete equipment is milling train, and real data is the outlet side shape of milling train.".

In addition, the present invention is that " a kind of storage medium storing program, the program are to realize the whole series by computer system Program when plant control unit, the complete equipment control device are a whole set of for control object complete equipment identification control object The pattern of the combination of the real data of equipment implements control, and the storage medium is characterized in that,

Computer system has:

Control method learning device learns the combination of the real data and control operation of control object complete equipment;And

Executive device is controlled, control object complete equipment is implemented according to the combination of the real data and control operation that learn Control,

Described program is for realizing the processing for controlling executive device, control rule executes program, control exports Decision procedure and control output inhibit program, wherein control rule executes actual number of the program according to control object complete equipment Control output is provided according to the determined combination with control operation;Control output decision procedure judges that the control rule executes journey Can sequence export control output, and notify that the real data and control operation are wrong to control method learning device;Control System output inhibits program to be judged as being output to control object complete equipment will control in control output decision procedure In the case of control object complete equipment the real data deteriorate in the case of, prevent will control be output to described in Control object complete equipment,

Described program be the processing for realizing the control method learning device, control result well whether judge journey Sequence, learning data production process and control rule learning program, wherein control result well whether decision procedure for realizing Control result well whether the processing that judges, wherein control output is effectively outputed to control object the whole series in control executive device Judge about real data and the control after the time delay for showing real data up to control effect in the case of equipment System before compared to become well still become deteriorate control result it is good whether;Learning data production process is tied using the control Fruit well whether decision procedure in control result it is good whether with control output obtain teacher's data;Control rule learning Program learns real data and teacher's data as learning data,

Learnt by control method learning device, multiple control mesh are directed to according to the state of control object complete equipment It marks and obtains the combination of real data alone and control operation, the combination of obtained real data and control operation is used as control System rule executes the determined combination of the real data and control operation of the control object complete equipment in program.".

By using the present invention, the shape model used in shape control and operation can be automatically corrected in control The control rule of method is simultaneously set as best.Therefore, have improve control accuracy, shorten control unit startup during, Neng Gouying To changing year in year out and other effects.

Description of the drawings

Fig. 1 is the figure of the summary for the complete equipment control device for showing the embodiment of the present invention.

Fig. 2 is the figure of the concrete structure example for the control rule enforcement division 10 for showing the embodiment of the present invention.

Fig. 3 is the figure of the concrete structure example in the control rule learning portion 11 for showing the embodiment of the present invention.

Fig. 4 is the neural network structure for showing the present invention being used in the case of the shape control of sendzimir mill Figure.

Fig. 5 is the figure of the shape control of the sendzimir mill described in the Fig. 1 for show patent document 1.

Fig. 6 is the control rule in the shape control of the sendzimir mill described in the Fig. 1 for show patent document 1 Figure.

Fig. 7 is the figure for the summary for showing input data preparing department 2.

Fig. 8 is the figure for the summary for showing control output operational part 3.

Fig. 9 is the figure for the summary for showing control output determination unit 5.

Figure 10 is to show form variations and the figure of control method.

Figure 11 is the figure for showing the summary of determination unit 6 whether control is good.

Figure 12 is the figure for arranging each section data, the relationship of mark in control output operational part 3 and showing.

Figure 13 is the figure for showing processing step and process content in learning data preparing department 7.

Figure 14 is the figure for the data example for showing to preserve in learning data database D B2.

Figure 15 is the figure for the example for showing neural network management table TB.

Figure 16 is the figure for the example for showing learning data database D B2.

(symbol description)

1:Control object complete equipment;2:Control input data preparing department;3:Control output operational part;4:Control output suppression Portion processed;5:Control output determination unit;6:Control result well whether determination unit;7:Learning data preparing department;10:Control rule is held Row portion;11:Control rule learning portion;20:Control executive device;21:Control method learning device;DB1:Control regular data Library;DB2:Output judgement database;DB3:Learning data database;Si:Real data;SO:Control operating quantity output;S1:It is defeated Enter data;S2:Control operating side operational order;S3:Control operating quantity;S4:Controlling operating quantity output could data;S5:Well Whether judge data;S6:Control result well whether data;S7a,S7b,S7c:Teacher's data;S8a,S8b,S8c:Input number According to (control rule learning portion use).

Specific implementation mode

Hereinafter, using the attached drawing embodiment that the present invention will be described in detail, before this, controlled as example with the shape of milling train Illustrate the idea in the present invention and realizes the process of the present invention.

First, it in order to solve the above subject in the present invention, needs:

1) and non-predetermined reference figure pattern is set separately and its control is operated to learn to control operating method, study Shape model and the group of control operation merge implements control operation using it.

2) because new control rule and it is non-predetermined can envision, perfect forecast sometimes less than control rule be best , so control operating side is randomly made to act and be found while observing the control result to it.

In order to be realized, need that the shape model used in shape controls and the combination of control operation is made to be become Change, change control operation is so that control result becomes good.For this reason, it may be necessary to be configured to study shape model and control operation The neural network of combination, according to control result it is good whether, change is for the nerve net of shape model occurred in milling train The output of the control operation of network.

When implementing above-mentioned while implementing shape control to the milling train in operation, the control of output error sometimes is defeated Go out, thus shape deteriorates sometimes and that the operations such as plate fracture occur is abnormal.When plate fracture occurs, the roller used in milling train is more Change that need the material to be rolled in time, rolling the losses such as to waste big.Therefore, it is necessary to be avoided as far as possible to milling train output error Control output.

According to the above, in the present invention, in order to be realized, such as using naive model of milling train etc. nerve net is verified Network output control operation it is good whether, avoid to milling train control operating side output can obviously think that shape deteriorates Output prevent the shape from deteriorating.At this point, about neural network, will be set as being mistake for the control of its shape model operation , and implement study.

Due to be possible to control operation it is good whether verification method itself generation mistake, so sentencing about with certain probability Break and operate output for the control of the neural network of mistake occurs, the control operating side of milling train is also output to, to learn The combination of shape model and control operation other than expectation.

[Embodiment 1]

Fig. 1 shows the summary of the complete equipment control device of the embodiment of the present invention.The complete equipment control device packet of Fig. 1 It includes:Control object complete equipment 1;Executive device 20 is controlled, the real data Si from control object complete equipment 1 is inputted, it will The control operating quantity output SO determined according to the control regular (neural network) as illustrated in Fig. 6 is supplied to control object complete equipment 1 and controlled;Control method learning device 21 inputs real data Si from control object complete equipment 1 etc. to carry out Study makes the control rule reflection that study goes out to the control rule in control executive device 20;(DB1 is extremely by multiple database D B DB3);And the management table TB of database D B.

Control executive device 20 is will to control input data preparing department 2, the regular enforcement division 10 of control, control output operational part 3, control output suppressing portion 4, control output determination unit 5, control operation interference generating unit 16 are constituted as main element.

Wherein, in controlling executive device 20, first according to the real data of the milling train as control object complete equipment 1 Si makes the input data S1 of control rule enforcement division 10 using control input data preparing department 2.Control rule enforcement division 10 makes With the neural network (control rule) of the relationship of the real data Si and control operating side operational order S2 of performance control object, root Control operating side operational order S2 is made according to the real data Si of control object.In control exports operational part 3, grasped according to control Make end operational order S2, control operating quantity S3 of the operation to control operating side.As a result, according to the reality of control object complete equipment 1 Border data Si makes control operating quantity S3 using neural network.

In addition, during the control in control executive device 20 exports determination unit 5, using from control object complete equipment 1 Real data Si and from control output operational part 3 control operating quantity S3, true oriented control operating side control operation Amount output could data S4.Control export suppressing portion 4 in, according to control operating quantity output could data S4 determinations could be to control Output control operating quantity S3 in operating side processed will be confirmed as exportable control operating quantity S3 conducts and be supplied to control object a whole set of The control operating quantity of equipment 1 exports SO and exports.It is judged as abnormal control operating quantity S3 as a result, and is not output to control pair As complete equipment 1.In addition, control operation interference generating unit 16 generates interference, is supplied to verify complete equipment control device Control object complete equipment 1.

The control executive device 20 constituted as previously discussed executes for its processing, and then as described later, with reference to control rule Then database D B1 and output judgement database DB3.Control rule database DB1 is connected to control executive device addressablely 11 this both sides of control rule learning portion in the regular enforcement division 10 of control and aftermentioned control method learning device 21 in 20.Make The control regular (neural network) of the learning outcome in rule learning portion 11 is stored in control rule database DB1 in order to control, The regular enforcement division 10 of control is regular with reference to the control stored in controlling rule database DB1.Output judgement database DB3 can be visited The control output determination unit 5 being connected to asking in control executive device 20.

Fig. 2 shows the concrete structure examples of the control rule enforcement division 10 of the embodiment of the present invention.The regular enforcement division of control The input data S1 that 10 inputs are produced by control input data preparing department 2 provides control operating side to control output operational part 3 Operational order S2.Control rule enforcement division 10 has neural network 1 01, is illustrated basically by such as Fig. 6 in neural network 1 01 Patent document 1 method come determine control operating side operational order S2.In the present invention, regular enforcement division 10 is controlled to be also equipped with Neural network selector 102 selects best control by referring to the control rule stored in controlling rule database DB1 Rule starts to execute as the control rule in neural network 1 01.In this way, in the control rule enforcement division 10 of Fig. 2, from by Necessary neural network is selected in multiple neural networks that operating personnel's group, control purpose divide and is used.In the regular number of control Can further include as neural network and good can be selected as the data from control object complete equipment 1 according in the DB1 of library Real data (data etc. of operation group) Si as determinating reference whether good.Further, since in the presence of if neural network is executed Then become the relationship of control rule, so not differentiating between neural network and control rule in the present specification, is come with identical meaning It uses.

Back to Fig. 1, in control method learning device 21, implement the neural network used in controlling executive device 20 101 study.In the case where controlling executive device 20 to the output control operating quantity output SO of control object complete equipment 1, Control effect practically shows as needing the time in the variation of real data Si.Therefore, using making time delay and the time pair Data obtained from the amount answered are implemented to learn.In Fig. 1, Z^-1Indicate the reasonable time delay feature for each data.

Control method learning device 21 be by control result well whether determination unit 6, learning data preparing department 7, control rule Then study portion 11, it is good whether judge that database DB4 is constituted as main element.

Wherein, control result well whether determination unit 6 use real data Si and reality from control object complete equipment 1 Border data previous value Si0 and judging whether well stored in database DB4 it is good whether judge that data S5, judgement are practical Data Si towards good direction change still towards the direction change of deterioration, output control result well whether data S6.

In learning data preparing department 7 in control method learning device 21, using will be made by control executive device 20 The control operating side operational order S2 that goes out, control operating quantity S3, the output of control operating quantity could the input datas difference such as data S4 Postpone same time obtained from data and from control result well whether determination unit 6 control result well whether data S6 is produced on the new teacher data S7a used in the study of neural network, is supplied to control rule learning portion 11.In addition, religion Teacher's data S7a is data corresponding with the regular control operating side operational order S2 of the output of enforcement division 10 of control, learning data system Make portion 7 can will use control result well whether determination unit 6 control result that provides well whether data S6 speculate control Data find out as new teacher's data S7a obtained from the control operating side operational order S2 that regular enforcement division 10 exports.

Fig. 3 shows the concrete structure example in the control rule learning portion 11 of the embodiment of the present invention.Control rule learning portion 11 be by input data preparing department 114, teacher's data creating portion 115, Processing with Neural Network portion 110, neural network selector 113 It is constituted as main structural element.In addition, in control rule learning portion 11, as from external input, obtain Make data S8a made of the input data S1 delay times from input data preparing department 2, obtains coming from learning data preparing department 7 new teacher data S7a, referring additionally to the number put aside in controlling rule database DB1 and learning data database D B2 According to.

In control rule learning portion 11, by input data S1 via input data after compensation reasonable time delay Preparing department 114 is taken into Processing with Neural Network portion 110.

In addition, in control rule learning portion 11, new teacher's data S7a from learning data preparing department 7 is made to teach It is the total religion for further including the past teacher's data S7b for being stored in learning data database D B2 in teacher's data creating portion 115 Teacher's data S7c and be supplied to Processing with Neural Network portion 110.These teacher's data S7a, S7b are suitably stored in learning data Database D B2 and utilize.

Similarly, the input data S8a from control input data preparing department 2 is made to be in input data preparing department 114 Further include the total input data S8c for the past input data S8b for being stored in learning data database D B2 and is supplied to god Through network processes portion 110.These input datas S8a, S8b are suitably stored in learning data database D B2 and are utilized.

Processing with Neural Network portion 110 is made of neural network 1 11 and neural network lea rning control portion 112, neural network 1 11 Be taken into the input data S8c from input data producing device 114, teacher's data S7c from teacher's data creating portion 115, The control that neural network selector 113 is selected is regular (neural network), and the neural network that will eventually determine is stored in control rule Then database D B1.

Neural network lea rning control portion 112 is for input data producing device 114, teacher's data creating portion 115, nerve net Network selector 113, with timing controlled appropriate they and obtain the input of neural network 1 11, in addition control is by handling result It is stored in control rule database DB1.

Here, neural network 1 01 in the control executive device 20 of Fig. 2 and in the control method learning device 21 of Fig. 3 Neural network 1 11 is all the neural network of same concept, when to being illustrated using the difference in upper basic conception, such as with It is lower described.First, the neural network that the neural network 1 01 in executive device 20 is predetermined content is controlled, is being provided The control operating side operational order S2 as corresponding output is found out when input data S1, it may be said that be used in a direction The neural network of processing.In contrast, neural network 1 11 in control method learning device 21 is used for will be about input data Input data S8c, the teacher's data S7c of S1 and control operating side operational order S2 are set as asking by learning when learning data Go out to meet the neural network of its input/output relation.

The considerations of basic handling in control method learning device 21 formed as described above, method was as described below.It is first First, control operating quantity output could data S4 content be "available" in the case of, to control object complete equipment 1 export control Operating quantity exports SO, control result well whether data S6 content be that " good " (real data Si becomes towards good direction Change) in the case of, judge that the control operating side operational order S2 that regular enforcement division 10 exports in order to control is correct, so that neural network Be output into control operating side operational order S2 mode make learning data.

On the other hand, control operating quantity output could the content of data S4 be "No" in the case of, or to control The output of object complete equipment 1 control operating quantity output SO and control result well whether data S6 content be "No" (actual number According to Si towards the direction change of deterioration) in the case of, judge the control operating side operational order that regular enforcement division 10 exports in order to control Mistake occurs for S2, and learning data is made in a manner of the not output of output nerve network.At this point, being exported as control, with opposite Identical control operating side output+direction, this 2 kinds of-direction mode exported constitute neural network output, with what is do not exported The mode of the control operating side operational order S2 of side makes learning data.

In addition, in the control rule learning portion 11 that Fig. 3 is illustrated, as the number based on neural network lea rning control portion 112 According to processing as a result, processing as described below.Here, first used as make to control executive device 20 input data S1 prolong The learning data of the combination of S8c obtained from the slow time and the teacher's data S7c produced by teacher's data creating portion 115, it is real It is applied to the study of the neural network 1 01 of control rule enforcement division 10.In fact, having and controlling in control rule learning portion 11 The 01 identical neural network 1 11 of neural network 1 of system rule enforcement division 10 carries out performance test to learn this under various conditions When response, the result as study be confirmed generate more good result control rule.Due to needing using multiple Learning data is learnt, so taken out from the learning data database D B2 for having put aside the learning data produced in the past This learning data is stored in learning data database D B2 by multiple past learning datas to implement study processing. In addition, learn the neural network is stored in control rule database DB1 to be utilized by the regular enforcement division 10 of control.

In the study of neural network, both past learning data can together be made whenever making new learning data For learning, past learning data can also together be made after by learning data savings to a certain degree (such as 100) For learning.

In addition, control result well whether determination unit 6 in, according to from well whether judge that database DB4's is good Whether determinating reference, judge whether implementing good.Control result it is good whether judge, due to being sentenced according to control purpose Disconnected result is different, so multiple neural networks corresponding with multiple control purposes are made, even if input data is identical also according to control Purpose processed makes teacher's data to learn respectively, to make multiple teacher's data for the input data of 1 secondary amounts and be used for The study of the corresponding neural network of each teacher's data, thus, it is possible to study nerve nets corresponding with multiple control purposes simultaneously Network.Here, multiple control purposes refer to, such as want any part is preferentially controlled on plate width direction in the case where shape controls (plate end, central portion, asymmetric portion etc.) thinks preferentially to control multiple control object projects (such as plate thickness and tension, rolling lotus Arbitrary project etc. in again etc.).

In the case where being configured to as described above, once the neural network 1 01 for controlling regular enforcement division 10 is learned When habit, new control operation is not just implemented.Therefore, make new operating method random in due course using control operation interference generating unit 16 Ground occurs, and is applied to control operating quantity S3 and carrys out executive control operation, to learn new control method.

Hereinafter, being controlled as object with the shape in sendzimir mill shown in patent document 1, this whole series is described in detail and sets Standby control method.In addition, being controlled about shape, illustrated using following specifications A, B.

Specification A is the specification about priority, is set as the information with the priority of plate width direction.For example, in shape control In system, it is desired value that the control on entire plate width direction is often difficult in mechanical property.Therefore, under being arranged on plate width direction State 2 specifications A1, A2 about priority.Wherein, it is " keeping plate end preferential " about the specification A1 of priority, about priority Specification A2 be " keeping central portion preferential ", implement according to 2 orders of priority as A1, A2 control.In the feelings for implementing control Under condition, consider about any specification in the specification A1 or A2 of priority.

Specification B is the specification for the condition distinguished in advance about reply.If given an example, shape model and control The relationship of method due to changing under various conditions, so be considered needing for example by by specification B1 be set as plate it is wide, by specification B2 is set as the division of steel grade class to separate.It is changed respectively due to above-mentioned, so the influence journey of the shape to shape manipulation end Degree changes.

In the example, control object complete equipment 1 is sendzimir mill, and real data is shape actual result.This Outside, sendzimir mill is the milling train with multiple roll for carrying out cold rolling to the hard material such as stainless steel.It is rolled in Sendzimir In machine, in order to provide hard material the working roll for rolling and using by force path.Therefore it is difficult to obtain flat steel plate.As Its countermeasure, using the constructing of multiple roll, variously-shaped control unit.In sendzimir mill, it is however generally that, among the upper and lower the 1st Roller has single cone, other than it can shift, also have 6 segmentation rollers and 2 rollers for being referred to as AS-U up and down.In following explanation Example in, as the real data Si of shape, using the detection data of shape detector, and then as input data S1, make It is used as and the form variations of the difference of target shape.In addition, as control operating quantity S3, it is set as the AS-U, upper and lower of #1~#n The roller amount of movement of 1st intermediate calender rolls.

Fig. 4 shows the neural network structure in the case of use in the control of the shape of sendzimir mill.Here, conduct Neural network is neural network 1 01 when for controlling regular enforcement division 10, and god is indicated when for controlling rule learning portion 11 Through neural network shown in network 111, but construct all identical.

In the example of the shape control of sendzimir mill shown in Fig. 4, the reality from control object complete equipment 1 Data Si be include that (form variations for being set as the difference as true form and target shape herein are defeated for the data of shape detector The data gone out) the real data of sendzimir mill, as input data S1, obtained in control input data preparing department 2 Standardized form deviation 201, form variations grade 202.The input layer of neural network 1 01,111 is inclined by standardized form as a result, Poor 201, form variations grade 202 is constituted.In addition, in Fig. 4, form variations grade 202 is used as to neural network input layer Input, but neural network can also be switched according to grade.

In addition, output layer and the shape as sendzimir mill control the AS-U of operating side, the 1st intermediate calender rolls accordingly by AS-U operational degrees 301 and the 1st intermediary operation degree 302 are constituted.In respective operational degree, about AS-U, for each AS- There is U AS-U evolutions to close direction (roll gap closing to (roll gap (interval between the roller of operation up and down of milling train) open direction), AS-U Direction).In addition, about the 1st intermediate calender rolls, have the 1st intermediate calender rolls evolution to (the 1st intermediate calender rolls are to roll upper and lower 1st intermediate calender rolls The direction that machine center is acted towards outside), the 1st intermediate calender rolls close direction (direction that the 1st intermediate calender rolls are acted towards milling train central side). For example, in the case where shape detector is 20 subregions and form variations grade 202 is set as 3 grade (large, medium and small), Input layer is 23 inputs.In addition, AS-U seat board be 7 pieces and upper and lower 1st intermediate calender rolls can when being shifted on plate width direction, Output layer is the 18 total of 14 AS-U operational degrees, 301,41 intermediary operation degree.In time set the number of plies of middle layer And the neuron number of each layer.Although in addition, carrying out aftermentioned, the shape about the sendzimir mill as output layer with reference to Fig. 8 Operating side is controlled, it is defeated to be constituted neural network in a manner of relatively each control operating side output+direction, the output of this 2 kinds of-direction Go out.

Figure 10 shows form variations and control method.Here, on the top of Figure 10, show in the case where form variations are big Control method the control method in the case where form variations are small is shown in the lower part of Figure 10.In addition, short transverse is shape The size of shape deviation, X direction are plate width direction, the wide both sides display plate end of plate, central display plate central portion.Such as figure It is preferential to correct entirety compared to the form variations of the part of plate width direction in the case where form variations are big shown in 10 top Shape.On the other hand, as shown in the lower part of Figure 10, in the case where form variations are small, the preferential form variations for reducing part.

As such, it is desirable to according to the size modification control method of form variations, so setting form variations grade as shown in Figure 4 202 and it is supplied to neural network 1 01,111, the size of predicting shape deviation.About form variations, no matter the size of form variations How, such as all using the result for being standardized as 0~1.This is an example, both it is contemplated that not carried out to form variations The input layer of neural network is standardized and is directly inputted to, it is also contemplated that certainly according to the big minor change neural network of form variations Body (such as prepares 2 neural networks, distinguishes the neural network that is used in the case where form variations are big and small in form variations In the case of the neural network that uses).

So that neural network 1 01,111 study described above such as the structure of Fig. 4 is directed to the operating method of shape model, makes Implement shape control with the neural network learnt.It, also can be according to learning even mutually isostructural neural network Condition and different control output is exported with different characteristic and for same shape pattern.

Therefore, multiple neural networks are used separately by the other conditions according to shape real data, can be directed to each Kind condition constitutes best control.This is the correspondence for specification B.The representation of previously described Fig. 2 carries out above-mentioned specification In the case of concrete example.In the structural cases of Fig. 2, for the neural network 1 01 for controlling regular enforcement division 10, according to The individual neural networks of preparations such as actual conditions, mill personnel's name, the steel grade class of material to be rolled, plate be wide are rolled, are registered To control rule database DB1.In neural network selector 102, the neural network of selection and the condition coupling at the time point And set the neural network 1 01 of regular enforcement division 10 in order to control.In addition, as the time point in neural network selector 102 Condition, the wide data of plate can be taken into from the real data Si in control object complete equipment 1, correspondingly selection god Through network.In addition, if multiple neural networks as used herein have input layer as shown in Figure 4, output layer, middle layer The number of plies, the unit number of each layer can also be different.

Fig. 7 shows to make data S1 (the standardized form deviations of the input layer for being input to neural network 1 01,111 201, form variations grade 202) control input data preparing department 2 summary.Here, as real data Si, detection is controlled Plate shape when object complete equipment 1 processed is the rolling in sendzimir mill, the shape detector data of shape detector As input, the testing result as each shape detector subregion is found out first with form variations PP values arithmetic unit 210 Form variations PP values (Peak To Peak values, peak-to-peak value) S of the difference of maximum value and minimum value_PP.It is transported in form variations grade It calculates in device 211, according to form variations PP values S_PPForm variations are classified as this large, medium and small 3 grades.Shape is rolled The plate width direction of the elongation of material is distributed, and the I-UNIT office of elongation will be indicated with 10-5 units.For example, such as following formula Classify like that.

Here, classifying as follows:By the establishment of (1) formula, form variations grade become (it is big=1, in=0, small= 0), by the establishment of (2) formula, form variations grade become (it is big=0, in=1, small=0), by the establishment of (3) formula, shape is inclined Poor grade become (it is big=0, in=0, small=1).In addition, form variations about each subregion herein, using being set as S_PM=S_PP's S_PMCarry out execution standardization.

[Formula 1]

S_PP≥50I-UNIT…(1)

[Formula 2]

50I-UNIT > S_PP≥10I-UNIT…(2)

[Formula 3]

10I-UNIT > S_PP…(3)

As previously discussed, the standardized form deviation 201 and shape as the input data to neural network 101 are made Deviation levels 202.Standardized form deviation 201 and form variations grade 202 are the input datas of control rule enforcement division 10 S1。

Fig. 8 shows the summary of control output operational part 3.Control output operational part 3 is according to from control rule enforcement division 10 The output of interior neural network 1 01 is to control operating side operational order S2 (in the example of the shape control of sendzimir mill AS-U operational degrees 301, the 1st intermediary operation degree 302 are suitable with its), it makes as the operation for controlling operating side to each shape The control operating quantity S3 of instruction.In addition, herein about there are multiple AS-U operational degrees 301, the 1st intermediary operation degree 302, Show that each 1 data example, each data are constituted from evolution to degree with a pair of of the data for closing direction degree.

In control output operational part 3, the AS-U operational degrees 301 of input have each AS-U evolutions to, close the defeated of direction Go out, so being multiplied by conversion gain G by the difference to them_ASUTo export the operational order to each AS-U.Due to the control to each AS-U System output is the positions AS-U amount of change (unit is length), so conversion gain G_ASUIt is the transformation from degree to position amount of change Gain.

In addition, the 1st intermediary operation degree 302 equally inputted has the output of outside, inside among the 1st, so passing through Conversion gain G is multiplied by their difference_1STTo export the operational order shifted to each 1st intermediate calender rolls.Due to each 1st intermediate calender rolls Control output be the 1st intermediate calender rolls shift position amount of change (unit is length), so conversion gain G_1STIt is from degree to position The conversion gain of amount of change.

By the above, being capable of operation control operating quantity S3.Operating quantity S3 is controlled by the positions #1~#nAS-U amount of change (n bases In the seat board number of AS-U rollers) and upper 1st intermediate bit position amount of change, lower 1st intermediate bit position amount of change composition.In addition, In fig. 8, it is illustrated that the interference data from control operation interference generating unit 16 are applied to control operating side operational order S2 System.

Fig. 9 shows the summary of control output determination unit 5.Control output determination unit 5 is repaiied by rolling phenomenon model 501 and shape Determination unit 502 is constituted whether just good, is obtained the real data Si from control object complete equipment 1, is exported fortune from control The information of the control operating quantity S3 and output judgement database DB3 in calculation portion 3, the control operating quantity provided to control operating side are defeated Going out could data S4.By above structure, in control exports determination unit 5, by the way that the control for exporting operational part 3 and calculating will be controlled The variation of shape in the case that operating quantity S3 processed is output to the milling train as control object complete equipment 1 is input to known control The model (in the case of the embodiment of fig. 9 be rolling phenomenon model 501) of object complete equipment 1 processed is predicted, pre- Want in the case of deteriorating for shape, inhibits control operating quantity to export SO, prevent shape from substantially deteriorating.

If described in more detail, control operating quantity S3 is input to rolling phenomenon model 501 to predict based on control The change in shape of operating quantity S3, operation form variations correction amount prediction data 503.On the other hand, by coming from control object The shape detector data Si (the form variations real data 504 of current point in time) of complete equipment 1 is added form variations amendment It measures prediction data 503 and obtains form variations prediction data 505, by evaluating form variations prediction data 505, can predict How shape changes when control operating quantity S3 is output to control object complete equipment 1.According to the form variations actual number of present situation According to 504 and form variations prediction data 505, whether shape corrects good in determination unit 502 predicting shape towards good direction Towards the direction change of deterioration, obtaining control operating quantity output could data S4 for variation or shape.

Whether shape corrects good in determination unit 502, specifically carry out as described below shape it is modified well with No judgement.First, as shown in specification A1, A2 of the priority controlled about shape, consider the control priority on plate width direction, So the weight coefficient w (i) of plate width direction is set to each specification of specification A1, specification A2 in output judgement database DB3.Make With its, such as using following (4) formulas evaluation function J come predicting shape variation it is good whether.In addition, in (4) formula, w (i) it is weight coefficient, ε fb (i) are that practical 504, the ε est (i) of form variations are that form variations predict that 505, i is shape detector Subregion, rand are random several.

[Formula 4]

In the case of the evaluation function J of use (4) formula, when evaluation function J deteriorates into just, shape becomes when shape becomes good Evaluation function J is negative.In addition, rand is random several, the evaluation result of evaluation function J is made randomly to change.Even if existing as a result, In the case that shape deteriorates, still there is the situation for becoming positive as evaluation function J, so not about rolling phenomenon model 501 Correct situation can also learn the relationship of shape model and control method.Here, in the control object as test running originally Increase maximum value in the case of the model of complete equipment 1 is insecure, want to implement stable control with certain degree learning control method The mode that 0 is set as in the case of system in time changes rand.

Whether shape corrects good in determination unit 502, operation evaluation function J, to be set as control operating quantity in J >=0 Output could data S4=1 (can), in J<Be set as when 0 control operating quantity output could data S4=0 (no) mode, output control Operating quantity output processed could data S4.

In control exports suppressing portion 4, exported according to the control operating quantity of the judgement result as control output determination unit 5 Could data S4, determining has undirected control object complete equipment 1 output control operating quantity output SO.Controlling operating quantity output could Data S4 is the amount of change output of the positions #1~#nAS-U, the output of upper 1st intermediate bit position amount of change, lower 1st intermediate bit position Amount of change output is set, determination of such as getting off is passed through.

IF (control operating quantity output could data S4=0) THEN

The positions #1~#nAS-U amount of change output=0

Upper 1st intermediate bit position amount of change output=0

Lower 1st intermediate bit position amount of change output=0

ELSE

The positions the #1~#nAS-U amount of change output=positions #1~#nAS-U amount of change

Upper 1st intermediate bit position amount of change exports=upper 1st intermediate bit position amount of change

Lower 1st intermediate bit position amount of change exports=descends the 1st intermediate bit position amount of change

ENDIF

In controlling executive device 20, according in the real data Si execution of control object complete equipment 1 (milling train) Operation is stated, control operating quantity output SO is output to control object complete equipment 1 (milling train), to implement shape control.

Next, illustrating the action summary of control method learning device 21.In control method learning device 21, using The time delay data of the data used in control executive device 20.Time delay Z^-1Mean e^-TS, indicate that delay is preset Time T.Control object complete equipment 1 has time response, so being sent out in real data according to control operating quantity output SO Changing is previously present time delay.Therefore, using the reality at the time point for passing through delay time T after executive control operation Data are implemented to learn.In shape control, detected to shapometer after being exported for the operational order of AS-U, the 1st intermediate calender rolls Until change in shape, several seconds are needed, it is possible to which being set as T=2 to 3 seconds or so, (delay time is also according to shape detector Type, mill speed and change, as long as so will control operating side change become change in shape until the best time It is set as T).

Figure 11 shows the action summary of determination unit 6 whether control is good.Change in shape well whether determination unit 602 in make With following formula it is good whether judge evaluation function J_C。

[Formula 5]

In addition, in (5) formula, ε fb (i) are the form variations real data for including, ε last (i) in real data Si It is form variations real data previous value, the plate width direction weight coefficient of judgement whether wC (i) is good.Here, based on good Database DB4 is judged whether good, according to the weight system of specification A1, A2 of the priority about control setting judgement whether good Number wC (i).Using well whether judge whether evaluation function Jc judgement control result good.In addition, being exported about as control The control operating quantity output of the judgement result of determination unit 5 could data S4 the case where being 0 (control output can not), although it is practical to The control operating quantity output=0 of control object complete equipment 1, but still it is judged as that shape deteriorates.

Here, control operating quantity output could data S4=0 in the case of, data S6 whether being set as control result well =-1.In addition, according to threshold condition (LCU >=0 >=LCL), preset threshold value upper limit LCU and threshold value add and subtract LCL.At this point, with Judge the result of the comparison of evaluation function Jc if it is Jc&gt whether good;LCU, then data S6 whether being set as control result well =-1 (shape deterioration), if it is LCU >=Jc >=0, then (shape is towards deterioration by data S6=0 whether being set as control result well Direction change), if it is 0>Jc >=LCL, then (shape becomes data S6=1 towards good direction whether being set as control result well Change), if it is Jc<LCL, then data S6=0 whether being set as control result well (shape is good).

Here, control result well whether data S6=-1 indicate to inhibit exported control output since shape deteriorates The case where, control result well whether data S6=0 indicate since amorphism variation or shape are good and keep being exported The case where control output, although control result well whether data S6=1 indicate that shape has towards good direction change into one The case where step becomes good possibility and exported controlled quentity controlled variable is made to increase.

In this way, according to specification A1, A2 of the priority about control, the weight coefficient wC (i) of plate width direction changes, Judge that evaluation function Jc is different whether so good.Therefore, the judgement result of data S6 whether being considered control result well It is different.Therefore, in control method learning device 21, control knot is implemented to this 2 kinds of specification A1, A2 of the priority about control Fruit well whether data S6 judgement.

Next, illustrating the summary of learning data preparing department 7.As shown in Figure 1, in learning data preparing department 7, according to next The judgement result (control result well whether data S6) of determination unit 6 whether from control result well utilizes control operating side behaviour The judgement result (output of control operating quantity could data S4) for making instruction S2, control operating quantity S3, control output suppressing portion, makes For teacher's data S7a of the neural network 1 11 used in control rule learning portion 11.

Teacher's data S7a of the situation is output, that is, AS-U operations of the output layer shown in Fig. 4 from neural network 1 11 Degree 301,1 intermediary operation degree 302.Control operating side of the learning data preparing department 7 used as the output of neural network 1 01 Operational order S2 (AS-U operational degrees 301,1 intermediary operation degree 301) and #1~#nAS- that SO is exported as control operating quantity The amount of change output of the positions U, the output of upper 1st intermediate bit position amount of change, the amount of change output of lower 1st intermediate bit position, make For teacher's data S7a of the neural network 1 11 used in control rule learning portion 11.

In the action summary for illustrating learning data preparing department 7, by each section number in the control output operational part 3 of Fig. 8 It is Figure 12 to be arranged according to the relationship of, mark.Here, the control operating side operational order S2 about the output as neural network 1 01, AS-U operational degrees 301 are representatively shown, the data of operational degree positive side are set as OPref, by the number of operational degree negative side According to OMref is set as, the operational degree occurred at random from control operation interference generating unit 16 is set as operational degree random number Conversion gain is set as G by Oref, and control operating quantity output SO is set as Cref and is illustrated.In this way, herein to put it more simply, The output of the output layer of neural network 1 01 from control rule enforcement division 10 is set as operational degree positive side and operational degree The operational degree occurred at random from control operation interference generating unit 16 is set as operational degree random number by negative side.In addition, will It is set as operational order value for the control operating quantity output SO of control operating side.

Figure 13 shows the processing step in learning data preparing department 7 and process content.Here, if according to Figure 12 mark Agreement illustrate, then utilize (6) formula to find out operational order value Cref in initial processing step 71.

[Formula 6]

Cref=G (OPref-OMref+ORref) ... (6)

In next processing step 72, whether according to control result well data S6 correct operational order value Cref and It is set as C ' ref.Specifically, control result well whether data S6=-1 when utilize (7) formula, control result well whether When data S6=0 utilize (8) formula, control result well whether data S6=1 when utilize (9) formula, be set as operational order value Correction value C ' the ref of Cref.

[Formula 7]

[Formula 8]

C ' ref=Cref ... (8)

[Formula 9]

In processing step 73, according to revised operational order value C ' ref, (10), (11) formula is utilized to find out operation journey Spend correction amount Oref.

[Formula 10]

C ' ref=G ((OPref+ Δ Oref)-(OMref- Δ Oref)) ... (10)

[Formula 11]

In processing step 74, (12) formula of utilization finds out teacher's data OP ' ref, OM ' ref for neural network 1 11.

[Formula 12]

In this way in learning data preparing department 7, as shown in figure 12, control object complete equipment 1 is exported for reality Operational order value Cref, the control result of judgement result whether according to as control result well in determination unit 6 well whether Data S6, arithmetic operation command value correction value C ' ref.Specifically, control result well whether data S6=1 the case where Under, control direction is OK, but insufficient in the case that judging to export in order to control, and operational order value is made to increase Δ to the same direction Cref.Conversely, control result well whether data S6=-1 in the case of, be judged as control direction there is a situation where mistake Under, so that operational order value is reduced Δ Cref round about.Conversion gain G is known, institute due to being preset With if it is known that operational degree positive side and the value of operational degree negative side, then can find out correction amount Oref.Here, sharp in advance Value appropriate is found out with emulation etc. to set Δ Cref.By above step, it can be found out and controlled using above-mentioned (12) formula Teacher's data OP ' ref, the OM ' ref used in rule learning portion 11.

In addition, being illustrated in fig. 13 with simple example, but actually implement the AS-U for #1~#nAS-U The whole of operational degree 301 and the 1st intermediary operation degree 302 shifted for the displacement of upper 1st intermediate calender rolls, lower 1st intermediate calender rolls, Teacher's data of the neural network 1 11 used in control rule learning portion 11 are set as (in AS-U operational degree teachers data, 1 Between operational degree teacher data).

Figure 14 shows the data example preserved in learning data database D B2.For learning neural network 111, need to be permitted The combination of multi input data S8a and teacher's data S7a.Therefore, the teacher's data S7a that will be produced by learning data preparing department 7 (AS-U operational degree teachers data, the 1st intermediary operation degree) and control rule execution is input in controlling executive device 20 The time delay data S8a of the input data S1 (standardized form deviation 201 and form variations grade) in portion 10 is combined and is made For one group of learning data, it is saved in learning data database D B2.

In addition, various database D B1, DB2, DB3, DB4, Figure 15 is used to show in the complete equipment control device of Fig. 1 Structure for the neural network management table TB for being associated ground management application to each database D B1, DB2, DB3, DB4.Pipe Reason table TB has the management table of specification.Specifically, in management table TB, according to (B1) plate wide, (B2) steel grade class with And specification A1, A2 of the priority about control divides specification.It is wide as (B1) plate such as wide using 3 feet, meter is wide, 4 Foot is wide, 5 feet wide this 4 divisions use steel grade class (1)~steel grade class (10) this 10 to divide degree as steel grade class.Separately Outside, for the specification A of the priority about control, it is set as this 2 kinds of A1 and A2.In this case, 80 divisions are set as, according to Rolling condition is used separately 80 neural networks.

Neural network lea rning control portion 112 is defeated by conduct as shown in figure 14 according to the neural network management table TB of Figure 15 The learning data for entering the combination of data and teacher's data is associated with corresponding neural network No. (number), be stored in as Learning data database D B2 shown in Figure 16.

When controlling executive device 20 to the execution shape control of control object complete equipment 1,2 groups of learning datas are made.It is former Because being exported for identical input data, control, the specification A1 and specification A2 this 2 of the priority about control is used Evaluation criteria judges whether carrying out control result well, so making 2 kinds of teacher's data.If teacher's data are put aside certain Degree (such as 200 groups) or new savings arrive learning data database D B2, then neural network lea rning control portion 112 indicates nerve The study of network 111.

In controlling rule database DB1, according to management table TB as shown in figure 15, multiple neural networks are stored, The neural network No. for needing to learn, neural network selector 113 is specified to be advised from control in neural network lea rning control portion 112 The neural network then is taken out in database D B1, is set as neural network 1 11.Neural network lea rning control portion 112 is to input data Preparing department 114 and the instruction of teacher's data creating portion 115 are taken out corresponding with the neural network from learning data database D B2 Input data and teacher's data make the study for being used to implement neural network 1 11.In addition, the study about neural network Method, it is proposed that various methods can also use arbitrary method.

When the study of neural network 1 11 is completed, neural network lea rning control portion 112 is by the nerve net as learning outcome Network 111 writes back to the position of the neural network No. of control rule database DB1, to learn to complete.

Learn both be directed to all neural networks defined in fig.15 with Fixed Time Interval (such as daily) together Implement, new learning data can also only be made to be existed by the neural network for putting aside the neural network No. of (such as 100 groups) to a certain degree The time point learns.

More than, it can be realized without substantially upsetting as the shape of the milling train of control object complete equipment 1:

1) and it is non-predetermined be set separately reference figure pattern and needle control of which operation come learn control operating method, and It is to learn the group merging that shape model is operated with control to operate to implement control using it.

2) because new control rule and it is non-predetermined can envision, perfect forecast sometimes less than control rule be best , so randomly acting control operating side and being found while observing the control result to it.

In addition, the neural network used in being stored in control executive device 20 in controlling rule database DB1, but if Stored neural network is to carry out then the study of neural network obtained from only randomly implementing initialization process, until The time is spent until capable of being controlled accordingly.Therefore, when having constructed control unit to control object complete equipment 1, according to The Controlling model for the control object complete equipment 1 distinguished at the time point advances with the study of simulation implementation control rule, will The neural network that study in emulator is completed is stored in database, can be from the startup of control object complete equipment when from the beginning of reality Apply the control of performance to a certain degree.

The complete equipment control device of the present invention is practical to be implemented as computer system, but in this case in department of computer science Multiple program groups are formed in system, wherein these program groups can also be stored in storage medium.

These program groups be, for example, for realizing control executive device processing, control rule execute program, control it is defeated Go out decision procedure and control output inhibits program, wherein control rule executes reality of the program according to control object complete equipment The determined combination of data and control operation exports to provide control;Control output decision procedure judgement control rule executes journey Can sequence export control output, and notify that the real data and control operation are wrong to control method learning device;Control System output inhibits program in control output decision procedure to be judged as that the feelings for being output to control object complete equipment will be being controlled In the case that the real data of control object complete equipment deteriorates under condition, prevent control being output to the control Object complete equipment processed, these program groups be, for example, for realizing control method learning device processing, control result it is good Whether decision procedure, learning data production process and control rule learning program, wherein control result well whether judge journey Sequence for realizing control result well whether the processing that judges, wherein control output is effectively outputed to control in control executive device Judge about reality after the time delay for showing real data up to control effect in the case of object complete equipment processed Data become compared with before the control well still become deteriorate control result it is good whether;Learning data production process makes Control result whether with the control result well in decision procedure it is good whether with control output obtain teacher's data;Control Rule learning program processed learns real data and teacher's data as learning data.

In addition, being learnt by control method learning device, according to the state of control object complete equipment for multiple Control targe and the combination for obtaining real data alone and control operation, by the combination of obtained real data and control operation The determined combination of the real data and control operation of the control object complete equipment in program is executed as control rule.

In addition, by apparatus of the present invention be applied to practical complete equipment when it needs to be determined that neural network initial value, about This point can be passed through before the control in implementing control object complete equipment using the Controlling model of control object complete equipment The real data in control object complete equipment and control operation are shortened in the combination of simulation making real data and control operation During the study of combination.

Industrial availability

The present invention is for example related to control method and the part of the milling train as one of rolling device, simultaneously in practical application Without special problem.

Claims (12)

1. a kind of complete equipment control device identifies the real data of control object complete equipment for control object complete equipment The pattern of combination implement control, which is characterized in that have:

Control method learning device learns the combination of the real data and control operation of control object complete equipment;And

Executive device is controlled, control object complete equipment is implemented according to the combination of the real data and control operation that learn Control,

The control executive device has:

The regular enforcement division of control is carried according to the determined combination of the real data and control operation of control object complete equipment It is exported for control;

Control output determination unit judges that can the control rule enforcement division export control output, and to the controlling party science of law It practises device and notifies that the real data and control operation are wrong;And

Control output suppressing portion is judged as being output to control object the whole series will control in control output determination unit In the case that the real data of the control object complete equipment deteriorates in the case of equipment, prevent to export control It is output to the control object complete equipment,

The control method learning device has:

Control result well whether determination unit, the control executive device will control output be effectively outputed to control object the whole series In the case of equipment, after the time delay until control effect shows real data, judge about real data With control before compared to become well still become deterioration control result it is good whether;

Learning data preparing department, control result whether using the control result well in determination unit it is good whether and the control System output obtains teacher's data;And

Rule learning portion is controlled, the real data and teacher's data are learnt as learning data,

Learnt by the control method learning device, multiple controls are directed to according to the state of the control object complete equipment Target processed and the combination for obtaining real data alone and control operation, the group of obtained real data and control operation is shared Make the determined combination of the real data and control operation of the control object complete equipment in the regular enforcement division of the control.

2. complete equipment control device according to claim 1, which is characterized in that

In order to replace the combination of real data and control operation according to the size of the real data of control object complete equipment, make With information related with the size of real data and it is easy to be standardized real data to implement the information of pattern-recognition, learns The combination of real data and control operation is practised to be controlled.

3. complete equipment control device according to claim 1 or 2, which is characterized in that

The regular enforcement division of control protects the determined combination of the real data of control object complete equipment and control operation It holds as the 1st neural network, the combination of real data and control operation is remained the 2nd nerve net by the control rule learning portion Network holds obtained 2nd neural network of the result of the study in the control method learning device as the control rule The 1st neural network in row portion.

4. the complete equipment control device according to any one in claims 1 to 3, which is characterized in that

The control executive device has the control operation interference generating unit for providing the control output interference, the controlling party Calligraphy learning device is also learnt when including being applied in interference.

5. the complete equipment control device according to any one in Claims 1-4, which is characterized in that

The control method learning device obtains real data and control by the study based on predetermined multiple specifications Multiple combinations of operation, the control executive device are whole according to control object from multiple combinations of real data and control operation The operating condition of complete equipment selects multiple combinations of 1 real data and control operation to provide the control output.

6. complete equipment control device according to claim 3, which is characterized in that

According to the size of real data, the nerve net that the combination to be used to real data and operating method is learnt is changed Network.

7. the complete equipment control device according to any one in claim 1 to 6, which is characterized in that

According to the changes control such as the state of the control object complete equipment or the experience of operator of control object complete equipment Result processed well whether determinating reference, find out the relationship of the real data and operating method for control object complete equipment respectively And it is separately stored in database, to according to the state of the control object complete equipment or the behaviour of control object complete equipment The experience etc. of work person, is controlled using different control methods.

8. the complete equipment control device according to any one in claim 1 to 7, which is characterized in that

Before the control in implementing control object complete equipment, using the Controlling model of control object complete equipment, by imitative Very come the combination for making the real data with controlling operation, shorten the real data in control object complete equipment and control During the study for making the combination of operation.

9. a kind of rolling mill control apparatus, using the complete equipment control device described in any one in requirement 1 to 8 of having the right, The rolling mill control apparatus is characterized in that,

The control object complete equipment is milling train, and the real data is the outlet side shape of the milling train.

10. a kind of complete equipment control method, which is applied to complete equipment control device, for control Object complete equipment identifies the pattern of the combination of the real data of control object complete equipment to implement control, which is characterized in that The complete equipment control device has:

Control method learning device learns the combination of the real data and control operation of control object complete equipment;And

Executive device is controlled, control object complete equipment is implemented according to the combination of the real data and control operation that learn Control,

The control executive device is carried according to the determined combination of the real data and control operation of control object complete equipment It is exported for control, can judgement export control output, and the real data and control are notified to the control method learning device Operation be it is wrong, be judged as will control be output to the control object complete equipment in the case of it is described control pair In the case of as the real data of complete equipment deteriorating, prevent control being output to described control object the whole series Equipment,

Control output is effectively outputed to control object the whole series in the control executive device and set by the control method learning device In the case of standby, after the time delay until control effect shows real data until, judgement about real data and Control before compared to become well still become deteriorate control result it is good whether, using the control result it is good whether with The control output obtains teacher's data, and the real data and teacher's data are learnt as learning data, By study, obtained for multiple control targes according to the state of the control object complete equipment real data alone and The combination of obtained real data and control operation is used as the control pair in the control executive device by the combination for controlling operation As the determined combination of the real data and control operation of complete equipment.

11. a kind of milling train control method, using the complete equipment control method having the right described in requirement 10, which is characterized in that

The control object complete equipment is milling train, and the real data is that the milling train goes out side shape.

12. a kind of storage medium storing program, which is to realize to set for control object the whole series using computer system When complete equipment control device of the pattern of the combination of the real data of standby identification control object complete equipment to implement control Program, the storage medium be characterized in that,

Computer system has:

Control method learning device learns the combination of the real data and control operation of control object complete equipment;And

Executive device is controlled, control object complete equipment is implemented according to the combination of the real data and control operation that learn Control,

Described program is to execute program, control output judgement journey for realizing the control rule of the processing of the control executive device Sequence, control output inhibit program, wherein it is described control rule execute program according to control object complete equipment real data and The determined combination of control operation provides control output;The control output decision procedure judges that the control rule executes program Control output can be exported, and notifies that the real data and control operation are wrong to the control method learning device; Control output inhibits program to be judged as that be output to control object a whole set of will control in control output decision procedure In the case that the real data of control object complete equipment deteriorates in the case of equipment, prevent control output output To the control object complete equipment,

Described program be the processing for realizing the control method learning device, control result well whether decision procedure, Learning data production process, control rule learning program, wherein the control result well whether decision procedure for making control As a result the processing judged whether good is realized, wherein control output is effectively outputed to control object the whole series in control executive device In the case of equipment after the time delay until control effect shows real data judgement about real data with Control before compared to become well still become deteriorate the control result it is good whether;The learning data production process makes Control result whether with the control result well in decision procedure it is good whether with control output obtain teacher's data;Institute Control rule learning program is stated to learn the real data and teacher's data as learning data,

Learnt by control method learning device, multiple control mesh are directed to according to the state of the control object complete equipment It marks and obtains the combination of real data alone and control operation, the combination of obtained real data and control operation is used as institute State the determined combination of the real data and control operation of the control object complete equipment in control rule execution program.