CN109317522A - Plant control unit, rolling mill control device and its control method, recording medium - Google Patents

Plant control unit, rolling mill control device and its control method, recording medium Download PDF

Info

Publication number
CN109317522A
CN109317522A CN201810804376.6A CN201810804376A CN109317522A CN 109317522 A CN109317522 A CN 109317522A CN 201810804376 A CN201810804376 A CN 201810804376A CN 109317522 A CN109317522 A CN 109317522A
Authority
CN
China
Prior art keywords
control
data
real data
mentioned
object equipment
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201810804376.6A
Other languages
Chinese (zh)
Other versions
CN109317522B (en
Inventor
服部哲
高田敬规
田内佑树
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hitachi Ltd
Original Assignee
Hitachi Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hitachi Ltd filed Critical Hitachi Ltd
Publication of CN109317522A publication Critical patent/CN109317522A/en
Application granted granted Critical
Publication of CN109317522B publication Critical patent/CN109317522B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • GPHYSICS
    • G05CONTROLLING; REGULATING
    • G05BCONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/0265Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion
    • G05B13/027Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion using neural networks only
    • BPERFORMING OPERATIONS; TRANSPORTING
    • B21MECHANICAL METAL-WORKING WITHOUT ESSENTIALLY REMOVING MATERIAL; PUNCHING METAL
    • B21BROLLING OF METAL
    • B21B37/00Control devices or methods specially adapted for metal-rolling mills or the work produced thereby
    • B21B37/16Control of thickness, width, diameter or other transverse dimensions
    • B21B37/18Automatic gauge control
    • BPERFORMING OPERATIONS; TRANSPORTING
    • B21MECHANICAL METAL-WORKING WITHOUT ESSENTIALLY REMOVING MATERIAL; PUNCHING METAL
    • B21BROLLING OF METAL
    • B21B37/00Control devices or methods specially adapted for metal-rolling mills or the work produced thereby
    • BPERFORMING OPERATIONS; TRANSPORTING
    • B21MECHANICAL METAL-WORKING WITHOUT ESSENTIALLY REMOVING MATERIAL; PUNCHING METAL
    • B21BROLLING OF METAL
    • B21B13/00Metal-rolling stands, i.e. an assembly composed of a stand frame, rolls, and accessories
    • B21B13/14Metal-rolling stands, i.e. an assembly composed of a stand frame, rolls, and accessories having counter-pressure devices acting on rolls to inhibit deflection of same under load; Back-up rolls
    • B21B13/147Cluster mills, e.g. Sendzimir mills, Rohn mills, i.e. each work roll being supported by two rolls only arranged symmetrically with respect to the plane passing through the working rolls
    • BPERFORMING OPERATIONS; TRANSPORTING
    • B21MECHANICAL METAL-WORKING WITHOUT ESSENTIALLY REMOVING MATERIAL; PUNCHING METAL
    • B21BROLLING OF METAL
    • B21B37/00Control devices or methods specially adapted for metal-rolling mills or the work produced thereby
    • B21B37/28Control of flatness or profile during rolling of strip, sheets or plates

Landscapes

  • Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Mechanical Engineering (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Automation & Control Theory (AREA)
  • Feedback Control In General (AREA)
  • Control Of Metal Rolling (AREA)

Abstract

The present invention provides plant control unit, rolling mill control device and its control method, recording medium.Plant control unit identifies the combined mode of its real data to control object equipment to implement control, have control method learning device and control executive device, control executive device has the regular enforcement division of control, control output determination unit and control output suppressing portion, control method learning device has the good no determination unit of control result, learning data establishment portion, control regular forgetting portion and control rule learning portion, control method learning device passes through study, the combination of independent real data and control operation is obtained to multiple control targets according to the state of control object equipment, the combination of obtained real data and control operation is used as to the combination of the real data for the control object equipment for controlling regular enforcement division and the decision of control operation, using repairing when the state of control object equipment is similar with the combination of real data and control operation before amendment Learning data after just implements control.

Description

Plant control unit, rolling mill control device and its control method, recording medium
Technical field
It the present invention relates to the use of in the real-time feedback control of the artificial intelligence technology progress of neural network etc., equipment control Device processed and its control method, rolling mill control device and its control method, recording medium.
Background technique
In the past, implemented to obtain desired control result by its control based on various controls in various equipment Theoretical equipment control.
As an example of equipment, such as in rolling mill control, as control theory, using an example as control Son and the shape of the ripple state of control panel control be object, be applicable in fuzzy control, nerve fuzzy control.Fuzzy control is applicable in It is controlled in the shape that coolant is utilized, in addition, the shape that nerve fuzzy control is suitable for Sendzimir rolling mill controls.Its In, as shown in Patent Document 1, find out the true form mode (pattern) and target shape mode detected with shape detector Difference, with the similar ratio of preset reference figure mode, according to this similar to ratio, according to this also by for presetting Reference figure mode control operating side operating quantity come the control rule that shows, find out the control output quantity for operating side come Carry out the shape control for being applicable in nerve fuzzy control.Hereinafter, as conventional art, using having used the gloomy of nerve fuzzy control The shape control of your formula rolling mill of Jimmy.
Fig. 5 indicates the shape control for the Sendzimir rolling mill that Fig. 1 of patent document 1 is recorded.Sendzimir steel rolling In the shape control of machine, nerve fuzzy control is used.In the example, in pattern-recognition mechanism 51, examined using by shape detector 52 The true form measured carries out the pattern-recognition of shape, and true form and preset reference figure mode which is calculated It is a closest.It controls in arithmetical organ 53, is grasped using by the control operating side for preset shape model shown in fig. 6 The control rule that work amount is constituted implements control.If further specifically described to Fig. 6, in pattern-recognition mechanism 51, operation The true form (Xing Zhuan real achievement that is detected out by shape detector 52) and target shape (ε ref) difference (Δ ε) and 1 to 8 Shape model (ε) which is closest, control in arithmetical organ 53, any of 1 to 8 control method selected to execute.
However, in order to control the verifying of rule, having in the method for patent document 1 and requiring operator during the rolling process Member is operated manually, and carries out the situation of verifying of control rule etc., there is the case where showing change in shape contrary to expectations. That is, generating the control rule of above-mentioned decision such as does not meet the situation of reality.The reason is that the research of mechanical property is insufficient, rolls The variation of the mode of operation, mechanical condition of steel machine, however verify whether preset control rule is best rule one by one Condition in need of consideration is more and difficult.Therefore, it is then just tieed up as long as no bad once setting control is regular in most cases It holds.
It due to the variation etc. of operating condition, controls rule and does not meet reality, then because control rule is fixed, so being difficult to Obtain the control precision above to a certain extent.Once then operator is without manual operation in addition, shape control starts to act (becoming noise for control), so being also difficult to find out new control rule by the manual intervention of operator.Also, It is also difficult to set control rule with corresponding to its material in the case where the rolling stock to new specification rolls.
As more than, in previous shape control, controlled using preset control rule, so there is hardly possible With the problem of Correction and Control rule.
In order to solve this problem, as shown in patent document 2, make to control regular random while carrying out shape control Variation, the rule that study shape improves, is achieved in:
1) new control rule is found out while implementing shape and controlling during the rolling process.
2) new control rule has and can not expect in advance, but does not have foreseeable control rule to become best completely The case where, so randomly making to control operating side movement, the control result for this is observed to find out.
Representative shape is set as reference figure mode in advance by above-mentioned conventional art, based on expression and is directed to benchmark The control rule of the relationship of the control operating side operating quantity of waveform pattern is controlled.The study of rule is controlled also and for base The control operating side operating quantity of waveform mode is related, maintains using pre-determined representative reference figure mode.Therefore, There are problems that becoming and only the shape of specific shape model reaction is controlled.
Reference figure mode is that people is in advance based on knowledge related with the rolling mill of object is become, true form and savings hand It moves the experience of interventional procedure to determine, it is difficult to which covering is as the whole generated in the rolling mill and material to be rolled of object Shape.Therefore, also have in the case where producing the shapes different from reference figure mode, cannot implement based on shape control Control, cannot inhibit form variations and remain, or be mistakenly identified as similar reference figure mode, carry out the control behaviour of mistake The case where making, deteriorating shape instead.
Therefore in previous shape control, control using preset reference figure mode and correspondingly is regular To carry out the study of control rule, implementation control, so the problem for having the raising of control precision limited.
Patent document 1: Japan Patent 2804161
Patent document 2: Japan Patent 4003733
Summary of the invention
In order to solve this problem, it such as provides a kind of for control object equipment, the actual number of identification control object equipment According to combined mode, come implement control plant control unit.
The plant control unit has the real data of study control object equipment and the combined controlling party of control operation Calligraphy learning device and the control for implementing the control of control object equipment according to the combination of the real data of study and control operation Executive device.
Control executive device has: mentioning according to the combination of the real data of control object equipment and the decision of control operation For the control rule enforcement division of control output;Determine to control the control output of regular enforcement division output could, and to above-mentioned control Method learning device processed notifies the real data and controls the control output determination unit of operating mistake;Control output determination unit is right In the case where the output control output of control object equipment, in the case where being judged as that the real data of control object equipment deteriorates, resistance Only suppressing portion is exported to the control that control object equipment exports control output.
Control method learning device has: control output being effectively outputed to control object equipment in control executive device In the case of, after the delay until control effect appears to real data, determine that real data is gone back compared with before the control to be good It is the good no determination unit of good no control result of the control result of difference;Use the control result in the good no determination unit of control result Good no and control output obtains the learning data establishment portion of monitoring data;Using above-mentioned real data and above-mentioned monitoring data as The control rule learning portion for practising data to learn, control method learning device is by study, according to the state of control object equipment The combination of independent real data and control operation is obtained to multiple control targets, by obtained real data and control operation The combination as the decision of the real data and control operation that control the control object equipment in regular enforcement division is combined to use.
Here, control rule learning portion is learnt using the learning data being made of multiple real data and monitoring data, However also include past learning data as learning data to learn, so if residual becomes the bad reason of control result Learning data, then generate competition and learn insufficient.
In addition, the study of control rule is implemented afterwards to a certain degree in learning data accumulation, or operation stopped process is medium Implement, so there is a possibility that control output wrong during exporting again so far.Accordingly, there exist Correction and Control rules to obtain The problem of time is needed until optimal.
Therefore, in the present invention, it is therefore intended that, shape control in automatic Correction and Control can be realized with the shortest time by providing The control of shape model and operating method Deng used in is regular and becomes the plant control unit of optimal rule and its control Method, rolling mill control device and its control method and recording medium.
Plant control unit provided by the invention is directed to control object equipment, identifies the real data of control object equipment Combined mode, to implement control.
The plant control unit has the real data of study control object equipment and the combined controlling party of control operation Calligraphy learning device and the control for implementing the control of control object equipment according to the combination of the real data and control operation that are learnt Executive device processed.
Control executive device has: mentioning according to the combination of the real data of control object equipment and the decision of control operation For the control rule enforcement division of control output;Determine to control the control output of regular enforcement division output could, and to controlling party Calligraphy learning device notifies real data and controls the control output determination unit of operating mistake;It is judged as above-mentioned control object equipment In the case that real data deteriorates, the control for exporting control output to above-mentioned control object equipment is prevented to export suppressing portion.
Control method learning device has: controlling the case where executive device exports control output to control object equipment Under, after the delay until control effect appears to real data, determine the control result of the real data compared with before control The good no determination unit of good no control result;Use the good no and control output of the control result in the good no determination unit of control result Obtain the learning data establishment portion of monitoring data;It deletes and the real data and control operation before the amendment of the learning data of creation Composite class as learning data control rule forgetting portion;Real data and monitoring data are learnt as learning data Control rule learning portion.
Control method learning device obtains multiple control targets according to the state of above-mentioned control object equipment by study The combination of independent real data and control operation is held the combination of obtained real data and control operation as control rule The real data of control object equipment in row portion is used with the combination for controlling the decision operated, in the shape of control object equipment In the case of as the composite class that real data and control before state and amendment operate, implement to control using revised learning data System.
In accordance with the invention it is possible to realize shape model used in shape control etc. in automatic Correction and Control with the shortest time Control with operating method is regular and becomes optimal rule.
Detailed description of the invention
Fig. 1 is the figure for indicating the summary of plant control unit of the embodiment of the present invention.
Fig. 2 is the figure for indicating the specific structural cases of the regular enforcement division 10 of the control of the embodiment of the present invention.
Fig. 3 is the figure for indicating the specific structural cases in control rule learning portion 11 of the embodiment of the present invention.
Fig. 4 is neural network structure the case where indicating for the present invention to be suitable for the shape control of Sendzimir rolling mill Figure.
Fig. 5 is the figure indicated in the shape control of Fig. 1 of patent document 1 Sendzimir rolling mill recorded.
Fig. 6 is the control rule indicated in the shape control of Fig. 1 of patent document 1 Sendzimir rolling mill recorded Figure.
Fig. 7 is the figure for indicating the summary of input data establishment portion 2.
Fig. 8 is the figure for indicating the summary of control output operational part 3.
Fig. 9 is the figure for indicating the summary of control output determination unit 5.
Figure 10 is the figure for indicating form variations and control method.
Figure 11 is the figure for indicating to control the summary of good no determination unit 6.
Figure 12 is the figure arranged each portion's data, the relationship of symbol in control output operational part 3 to show.
Figure 13 is the figure for indicating processing stage and process content in learning data establishment portion 7.
Figure 14 is the figure for indicating to be stored in the data example of learning data database D B2.
Figure 15 is the figure for indicating the example of neural network management table TB.
Figure 16 is the figure for indicating to reflect the example of learning data database D B2 for the processing for controlling regular forgetting portion 200.
Figure 17 is the figure for indicating the storage content of forgetting rules database D B5.
Description of symbols
1: control object equipment
2: control input data establishment portion
3: control output operational part
4: control output suppressing portion
5: control output determination unit
6: the good no determination unit of control result
7: learning data establishment portion
10: controlling regular enforcement division
11: control rule learning portion
20: control executive device
21: control method learning device
100: displacement controls regular enforcement division
200: controlling regular forgetting portion
DB1: control rule database
DB2: output determines database
DB3: learning data database
DB4: good no judgement data database
DB5: forgetting rules database
Si: real data
SO: control operating quantity output
S1: input data
S2: control operating side operational order
S3: control operating quantity
S4: control operating quantity output could data
S5: good no judgement data
S6: the good no data of control result
S7a, S7b, S7c: monitoring data
S8a, S8b, S8c: input data (control rule learning portion use).
Specific embodiment
Attached drawing used below, the embodiment that the present invention will be described in detail, before it first by taking the control of the shape of rolling mill as an example Illustrate the opinion in the present invention and obtains process of the invention.
Firstly, being needed to solve the above subject in the present invention:
1) be not be set separately in advance reference figure mode and correspondingly control operation, come learn control operation side Method, but learn the combination of shape model and control operation, and implement control operation using it.
2) new control rule can not expected in advance because having, but inscrutable control rule completely becomes Best situation observes control result correspondingly so randomly making to control operating side movement to find out.
In order to realize above content, it is required that for the shape model of shape control and the combination variation of control operation, Change control operation is so that control result improves.For this reason, it may be necessary to be configured to the combination of study shape model and control operation Neural network, according to control result it is good it is no come change for the neural network of the shape model generated in rolling mill control behaviour The output of work.
If implementing above content while implementing shape control to the rolling mill in operation, also there is output error Control output the case where, have shape deteriorate and generate plate rupture etc. operation exception the case where.If generating plate rupture, rolling mill Used in the exchange of roller need the time, or the material to be rolled in the waste operation of rolling is then damaged very big.Therefore, it is necessary to The control of rolling mill not output error is exported as far as possible.
Therefore, in the present invention, such as the control that neural network exports is verified using simple model of rolling mill etc. and is operated It is good no, whether the output that obvious shape can deteriorate or not to the control operating side of rolling mill, prevents shape from deteriorating.At this point, about Neural network, it is believed that for the shape model control operating mistake and implement to learn.
It is possible that good itself mistake of no verification method of control operation, so by being wrong nerve with certain probabilistic determination Network-based control operation output is also output to the control operating side of rolling mill, so as to learn the shape model and control outside imagining Make the combination of operation.
In order to keep neural network lea rning control regular, many control methods as true form and correspondingly are needed Combined learning data.In the present specification study number will be recorded as with many learning datas of the study of Mr. Yu's neural network According to group.According to which type of learning data group used, the control rule as learning outcome is different.According to above-mentioned method, In the case that change is directed to the control method of shape model, existing learning data is appended to using new learning data is created The method of group.Learning data in learning data group only increases, so the time needed for the study of neural network also increases.Cause This, it is also envisaged that learning data is deleted or random erasure based on period, however also has the control rule of thus learning outcome The case where being taken on an entirely new look.Therefore, it is desirable that retaining the learning data in the learning data group of status and additional new study Data.
In the case where producing new learning data, should have in learning data group as being judged as no control effect The combined learning data of true form and control method correspondingly, it is even if retaining them that new learning data is additional To learning data group, also become that the two competes as a result, the control rule as learning outcome will not become be intended to rule Then (the same control rule with new learning data).
In addition, if learning data increases, correspondingly learning the required time also increases.Plant control unit because To be to implement to learn according to pre-determined timetable, so preferably learning time is almost fixed.
Therefore, in the present invention, while additional new learning data, it will be deemed as the learning data of no control effect from Data group is practised to delete.Thereby, it is possible to limit increasing for learning data, learnt using a certain range of learning data.
In addition, control rule possessed by neural network does not change before being learnt using new learning data group.Cause This, also has using being judged as the control rule of no control effect come the case where executing control again.The study of neural network needs Time, so until obtaining the learning outcome using new learning data group, without using the control for being judged as no control effect Rule processed is controlled, and improves control effect using based on the control method of modified learning data.
[embodiment 1]
Fig. 1 shows the summaries of the plant control unit of the embodiment of the present invention.The plant control unit of Fig. 1 includes: control Object-based device 1;Input the real data Si from control object equipment 1 and by the regular (nerve net of the control illustrated according to Fig. 6 Network) the control operating quantity that determines exports the SO control executive device 20 that is supplied to control object equipment 1 to control;Input carrys out automatic control The real data Si etc. of object-based device 1 processed is learnt, and makes the control rule reflection of study to the control in control executive device 20 Make the control method learning device 21 of rule;Multiple database D B (DB1 to DB3);And the management table TB of database D B.
Control the primary structure element of executive device 20 are as follows: control input data establishment portion 2, control regular enforcement division 10, Control output operational part 3, control output suppressing portion 4, control output determination unit 5, control gimp generating unit 16, displacement control Regular enforcement division 100.
Wherein, in control executive device 20, first according to the real data of the rolling mill as control object equipment 1 Si, using control input data establishment portion 2, creation controls the input data S1 of regular enforcement division 10.Control regular enforcement division 10 Using the neural network (control rule) of the relationship of the real data Si and control operating side operational order S2 of performance control object, Control operating side operational order S2 is created according to the real data Si of control object.In control output operational part 3, based on control behaviour Make end operational order S2, control operating quantity S3 of the operation to control operating side.As a result, according to the actual number of control object equipment 1 According to Si, control operating quantity S3 is created using neural network.
And control in the control output determination unit 5 in executive device 20, use the actual number from control object equipment 1 According to Si and the control operating quantity S3 from control output operational part 3, decision can to the control operating quantity output of control operating side No data S4.In control output suppressing portion 4, according to control operating quantity output could data S4 determine to control operating side Control operating quantity S3 output could, using be determined as can control operating quantity S3 as the control for being supplied to control object equipment 1 Operating quantity exports SO to export.It is judged as that abnormal control operating quantity S3 will not be output to control object equipment 1 as a result,.Control Gimp generating unit 16 generates noise, is supplied to control object equipment 1 to verify plant control unit.
In addition, being controlled in regular enforcement division 100 in displacement, advised by the control in aftermentioned control method learning device 21 Then forgetting portion 200 is judged as the learning data that should be deleted and the revised learning data from forgetting rules database D B5 taking-up Information, be compared with input data S1 and the output data S2 for controlling regular enforcement division 10, thus output control output As a result, become control object equipment 1 state deteriorating control output in the case where, based on the revised learning data Information output control exports the control output to replace controlling regular enforcement division 10.
Control executive device 20 configured as described above executes for its processing, and further as be described hereinafter, reference controls regular number Database DB3 is determined according to library DB1 and output.Control rule database DB1 is held with the control rule controlled in executive device 20 The both sides in the control rule learning portion 11 in row portion 10 and aftermentioned control method learning device 21 are connected in a manner of being able to access that It connects.Control regular (neural network) as the learning outcome in control rule learning portion 11 is stored in control rule database DB1 controls regular enforcement division 10 referring to the control rule for being stored in control rule database DB1.Output determine database DB3 with Control output determination unit 5 in control executive device 20 is connected in a manner of being able to access that.
Fig. 2 indicates the specific structural cases of the regular enforcement division 10 of the control of the embodiment of the present invention.Control rule executes The input data S1 created in 10 input control input data establishment portion 2 of portion provides control operating side to control output operational part 3 Operational order S2.It controls regular enforcement division 10 and has neural network 1 01, substantially utilized as illustrated in Fig. 6 in neural network 1 01 The method of patent document 1 come determine control operating side operational order S2.In the present invention, controls regular enforcement division 10 and be also equipped with nerve Network selector 102, referring to the control rule for being stored in control rule database DB1, thus as the control in neural network 1 01 System rule selects optimal control rule, and executes.In the control rule enforcement division 10 of Fig. 2 in this way, from according to operator Class, control purpose divide multiple neural networks, select need neural network come using.In control rule database DB1, It preferably also include the reality that can select neural network and good no determinating reference as the data from control object equipment 1 Data (data etc. of operation class) Si.In addition, in the relationship for becoming control rule if executing neural network, so this theory Neural network and control rule are not distinguished in bright book, is used with the identical meaning.
Back to Fig. 1, in control method learning device 21, implement neural network 1 01 used in control executive device 20 Study.In the case that control executive device 20 outputs control operating quantity output SO to control object equipment 1, actually control The variation that effect is embodied in real data Si needs the time.Therefore, using the data for postponing the time corresponding with the time To implement to learn.In Fig. 1, Z- 1Indicate the delay function appropriate for being directed to each data.
The primary structure element of control method learning device 21 is the good no determination unit 6 of control result, learning data establishment portion 7, rule learning portion 11, good no judgement database DB4 are controlled, controls regular forgetting portion 200, forgetting rules database D B5.
Wherein, the good no determination unit 6 of control result uses real data Si and real data from control object equipment 1 The previous value Si0 and good no judgement data S5 for being stored in good no judgement database DB4, determines that real data Si is toward good side To variation, or toward bad direction change, export the good no data S6 of control result.
In learning data establishment portion 7 in control method learning device 21, use will be created using control executive device 20 Control operating side operational order S2, control operating quantity S3, the output of control operating quantity could the input datas such as data S4 prolong respectively The data of slow identical time, the good no data S6 of control result from the good no determination unit 6 of control result, creation are used for nerve net The new monitoring data S7a of the study of network is supplied to control rule learning portion 11.In addition, monitoring data S7a is and control rule The corresponding data of control operating side operational order S2 that enforcement division 10 exports, learning data establishment portion 7 can will use control knot The good no data S6 of control result that the good no determination unit 6 of fruit provides controls the control operating side behaviour of the regular output of enforcement division 10 to infer The data for making instruction S2 to obtain are as new monitoring data S7a.
Fig. 3 indicates the specific structural cases in the control rule learning portion 11 of the embodiment of the present invention.Control rule learning The primary structure element in portion 11 are as follows: input data establishment portion 114, monitoring data establishment portion 115, Processing with Neural Network portion 110, mind Through network selector 113.In addition, control rule learning portion 11, makes to create from input data as from external input acquisition The data S8a for the input data S1 delay for building portion 2, the new monitoring data S7a from learning data establishment portion 7, reference are obtained Put aside the data in control rule database DB1 and learning data database D B3.
It controls in rule learning portion 11, input data S1 is after suitably compensation of delay through by input data establishment portion 114 It is input to Processing with Neural Network portion 110.
And in control rule learning portion 11, the new monitoring data S7a from learning data establishment portion 7, which is used as, also to be wrapped Include total supervision that the past monitoring data S7b of learning data database D B3 is stored in monitoring data establishment portion 115 Data S7c is supplied to Processing with Neural Network portion 110.These monitoring datas S7a, S7b is suitably stored in learning data data Library DB3 and be utilized.
In the same manner, the input data S8a from control input data establishment portion 2, which is used as, is also included within input data establishment portion It is stored in total input data S8c of the past input data S8b of learning data database D B3 in 114, is supplied to nerve Network processes portion 110.These input datas S8a, S8b are suitably stored in learning data database D B3 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 Read the input data S8c from input data creating device 114, the monitoring data S7c from monitoring data establishment portion 115, The control that neural network selector 113 selects is regular (neural network), and the neural network finally determined is stored into control rule Database D B1.
Neural network lea rning control portion 112 is directed to input data creating device 114, monitoring data establishment portion 115, nerve net Network selector 113, is in due course and controls them, controls to obtain the input of neural network 1 11, and processing result is stored To control rule database DB1.
Regular forgetting portion 200 is controlled according to input data S8a and new monitoring data S7a and controls regular enforcement division 20 output S2, from learning data database D B2 retrieval for create new monitoring data with original input data S8a with And the learning data that output S2 is similar, implement the replacement processing of the learning data and new monitoring data.
Here, in the control method learning device 21 of the neural network 1 01 in the control executive device 20 of Fig. 2 and Fig. 3 Neural network 1 11 is all the neural network of identical concept, however is illustrated using the difference in upper basic conception, then as follows. The neural network that the neural network 1 01 in executive device 20 is pre-determined content is controlled first, for providing input data The control operating side operational order S2 as corresponding output is found out when S1, it may be said that be the nerve for the processing in a side direction Network.On the other hand, the neural network 1 11 in control method learning device 21 will be for input data S1 and control operating side When input data S8c, the monitoring data S7c of operational order S2 is set as learning data, the input/output relation will be met Neural network is found out by study.
The idea of the basic processing in control method learning device 21 constituted as described above is as follows.Firstly, controlling Operating quantity processed output could data S4 content be "available" in the case where, control operating quantity output is exported to control object equipment 1 SO is judged as in the case where the content of the good no data S6 of control result is " good " (real data Si toward good direction change) The control operating side operational order S2 for controlling the regular output of enforcement division 10 is correct, is output into control operating side with neural network The mode of operational order S2 creates learning data.
On the other hand, control operating quantity output could data S4 content be "No", alternatively, to control object equipment 1 Output control operating quantity exports SO, and the content of the good no data S6 of control result is "No" (real data Si toward bad direction change) In the case where, it is judged as and controls the control operating side operational order S2 mistake that regular enforcement division 10 exports, with not output nerve net The mode of the output of network creates learning data.At this point, exported as control, with to identical control operating side output+direction ,- The mode of two kinds of outputs in direction constitutes neural network output, not export the control operating side operational order S2 of the side of output Mode create learning data.
In the control rule learning portion 11 that Fig. 3 is illustrated, the knot of the data processing as neural network lea rning control portion 112 Fruit is handled as follows.Here, first used as the S8c for making the input data S1 delay to control executive device 20 With the combined learning data of the monitoring data S7c created in monitoring data establishment portion 115, implement controlling regular enforcement division 10 The study of the neural network 1 01 used.In fact, having the mind with the regular enforcement division 10 of control in control rule learning portion 11 Through the identical neural network 1 11 of network 101, performance test is carried out to learn response at this time with various conditions, is confirmed The control rule for the result that result as study has generated.Study needs to carry out using multiple learning datas, therefore from product The learning data database D B2 of the learning data of creation in the past is stored, multiple past learning datas is taken out, learns and implement to locate Reason, and current learning data is stored into learning data database D B2.In addition, the neural network of study is held in control rule Row portion 10 is utilized, so being stored into control rule database DB1.
At this point, past learning data should be output into the control behaviour on the basis of the learning data specifically updated comprising becoming The learning data for the reason of making end operational order S2, so even if the directly additional learning data specifically updated learns, also at For with opposite learning data study as a result, the control method for interfering neural network learning new.Therefore, to specifically updating, chase after The combination of the input data S1 and control operating side operational order S2 on the basis of the learning data added, execute and delete most similar mistake The processing of the learning data gone is to control regular forgetting portion 200.
The study of neural network can be used together past learning data whenever creating new learning data to learn It practises, past learning data can also be used together to learn after learning data puts aside certain degree (such as 100).The feelings Under condition, it is used in the learning data number that the learning data group of study is included whenever creating new learning data and increases, and Have because opposite learning data there are due to learning effect reduce the case where.
In this patent, it has been observed that increasing for learning data can be prevented by the learning data of deletion error.For example, pre- First it is set as learning using 1000 learning datas, in the case where creating new learning data, deletes and original input The most similar learning data of combination of data S1 and control operating side operational order S2, can not be such that learning data increases, and make The new control method of nervous function e-learning.Thereby, it is possible to the time needed for saving study, as the preservation sky of learning data Between needed for computer resource (memory, hard disk etc.).In addition, being improved learning efficiency by deleting opposite learning data. Therefore, the study of neural network can be carried out with the shortest time.
In addition, in the good no determination unit 6 of control result, based on from the good no good no determinating reference for determining database DB4 To implement good no judgement.Control result it is good it is no determine that judging result is different according to control purpose, so creation it is multiple with it is multiple The corresponding neural network of purpose is controlled, monitoring data is respectively created according to control purpose input data is identical, is learned It practises, so that the input data to 1 secondary amounts creates multiple monitoring datas, for neural network corresponding with each monitoring data It practises, thus, it is possible to study neural networks corresponding with multiple control purposes simultaneously.Here, multiple control purposes are such as shape controls It is which part (plate end, central part, asymmetric portion etc.) in priority acccess control board width direction, or preferential in the case where system Control which of multiple control object projects (for example, plate thickness and tension, rolling load etc.) etc..
In the case where constituting as described above, learned once controlling neural network 1 01 used in regular enforcement division 10 It practises, then cannot implement new control operation.Therefore, by controlling gimp generating unit 16, new operation is randomly generated in due course Method, in addition control operating quantity S3 executive control operation, to learn new control method.
Hereinafter, being controlled with the shape in Sendzimir rolling mill shown in patent document 1 as object, detailed description is originally set Standby control method.It controls about shape, is illustrated using following specification A, B.
Specification A is the specification about priority, the information of the priority with board width direction.Such as it is controlled in shape In, due to being difficult in mechanical property throughout the control of board width direction universe be target value.Therefore, it is arranged in board width direction and closes In specification A1, A2 of following 2 priorities.It is wherein " preferential plate end " about the specification A1 of priority, about priority Specification A2 is " preferential central portion ", implements the control of 2 priorities according to as A1, A2.In the case where implementing control Consider the specification A1 or A2 about priority.
Specification B is the specification about the reply to the condition distinguished in advance.It gives an example, shape model and controlling party The relationship of method is changed with various conditions, so for example, it is believed that needs specification B1 being set as board width, specification B2 is set as steel grade in this way Differentiation separate.Above-mentioned each variation, to change to the influence degree of the shape at shape manipulation end.
Control object equipment 1 is Sendzimir rolling mill in the example, and real data becomes true form.Sendzimir Formula rolling mill is with the rolling mill in order to carry out the multiple roll of cold rolling to the hard material such as stainless steel.Sendzimir rolling mill In, for the purpose of suppressing to hard material application, use the working roll of path.Therefore it is difficult to obtain flat steel plate.As it Countermeasure, using the construction of multiple roll, various shape control units.In general, Sendzimir rolling mill, among the upper and lower the 1st Roller have a taper, can be displaced, in addition to this up and down have 6 segmentation rollers and 2 be referred to as AS-U roller.It says below In bright example, the real data Si as shape is further used as input data using the detection data of shape detector S1, using and target shape difference, that is, form variations.It is the AS-U of #1~#n and as control operating quantity S3, upper and lower the The roller displacement of 1 intermediate calender rolls.
Fig. 4 indicates the neural network structure of the case where shape control for Sendzimir rolling mill.Here so-called Neural network is controlled in regular enforcement division 10 with neural network 1 01, and the mind that neural network 1 11 is used in rule learning portion 11 is controlled Through network, regardless of which construction is all identical.
In the example of the shape control of Sendzimir rolling mill as shown in Figure 4, the reality from control object equipment 1 Data Si is the data (here, the difference that is, form variations to export true form and target shape) comprising shape detector The real data of Sendzimir rolling mill controls in input data establishment portion 2, obtains standardized form as input data S1 Deviation 201, form variations grade 202.Thus the input layer of neural network 1 01,111 is inclined by standardized form deviation 201, shape Poor grade 202 is constituted.In Fig. 4, by form variations grade 202 as the input to neural network input layer, however can also root Switch neural network according to grade.
In addition, output layer controls operating side, AS-U, the 1st intermediate calender rolls pair with the shape as Sendzimir rolling mill Ying Di is made of AS-U operational degree 301 and the 1st intermediary operation degree 302.About AS-U, each operational degree is each AS-U has AS-U evolution to (direction that roller gap (interval between the roller of operation up and down of rolling mill) is opened), the side of closing AS-U To (direction of roller gap-closing).In addition, about the 1st intermediate calender rolls, have the 1st intermediate calender rolls evolution to (the in upper and lower 1st intermediate calender rolls The direction that 1 intermediate calender rolls are acted towards the outside at rolling mill center), the 1st intermediate calender rolls close direction, and (the 1st intermediate calender rolls are towards rolling mill The direction of central side movement).For example, in the region of shape detector 20, by form variations grade 202 be 3 grades (it is big, in, It is small) in the case where, input layer is 23 inputs.In addition, it is assumed that the saddle of AS-U is 7, upper and lower 1st intermediate calender rolls can be wide in plate Direction displacement is spent, then the AS-U operational degree 301 of output layer is 14, and 1 intermediary operation degree is four, amounts to 18.About The neuron number of the number of plies of middle layer and each layer is set in due course.In addition illustrate below referring to Fig. 8, about as output layer The shape of Sendzimir rolling mill controls operating side, to each control operating side output+direction, two kinds of outputs in-direction Mode constitute neural network output.
Figure 10 indicates form variations and control method.Here the top Figure 10 indicates the controlling party of the big situation of form variations Method, the lower part of Figure 10 indicate the control method of the small situation of form variations.Short transverse is the size of form variations, X direction It is board width direction, the two sides display plate end of board width, central display plate central portion.As shown in the top of the Figure 10, in shape In the case that shape deviation is big, all shapes of preferential amendment compared with the form variations of the part in board width direction.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 change control method according to the size of form variations, so setting form variations etc. as shown in Figure 4 Grade 202 is supplied to neural network 1 01,111, the size of predicting shape deviation.It can be regardless of form variations about form variations Size, and use the deviation for being for example standardized as 0~1.This is an example, it is also considered that form variations standardization is not direct It is input to the input layer of neural network, according to the size of form variations, changes neural network itself (for example, preparing 2 nerve nets Network, be divided in form variations it is big in the case where the neural network that uses and it is small in the case where the neural network that uses).
To the neural network 1 01,111 of the structure of Fig. 4 as described above, study is made to be directed to the operation side of shape model Method implements shape control using the neural network of study.Even the neural network of identical structure, according to the condition of study Also become different characteristics, different control output can be exported to identical shape model.
Therefore, multiple neural networks can be used separately according to other conditions of true form, thus to a variety of conditions Constitute optimal control.This is the reply to specification B.The representation of the Fig. 2 illustrated before carries out the case where specification Concrete example.In the structural cases of Fig. 2, about neural network 1 01 used in regular enforcement division 10 is being controlled, in fact according to rolling Border achievement, rolling mill operator name, the steel grade of material to be rolled, board width etc. prepare independent neural network, log on to control Rule database DB1 processed.In neural network selector 102, selection meets the neural network of the condition at the moment, is set as controlling Make the neural network 1 01 of regular enforcement division 10.It, can be from control as the condition at the moment in neural network selector 102 The data that board width is obtained in real data Si in object-based device 1, correspondingly select neural network.In addition, making here As long as multiple neural networks have input layer shown in Fig. 4, output layer, the number of plies of middle layer, the unit number of each layer can be with It is different.
Fig. 7 indicates to create data S1 (the standardized form deviation for being input to the input layer of neural network 1 01,111 201, form variations grade 202) control input data establishment portion 2 summary.Here as real data Si, will test conduct The shape detector data of the shape detector of plate shape when rolling in the Sendzimir rolling mill of control object equipment 1 As input, firstly, finding out the maximum value of the testing result in each shape detector region with form variations PP value arithmetic unit 210 With difference that is, form variations PP value (Peak To Peak value) SPP of minimum value.Form variations grade operational device 211 is according to shape Form variations, are divided into 3 large, medium and small grades by shape deviation PP value SPP.Shape is the board width of the elongation of material to be rolled Directional spreding uses elongation using the I-UNIT that 10-5 unit indicates as unit.For example, as following formula is classified.
Here, be classified as according to the establishment of (1) formula and form variations grade be (it is big=1, in=0, it is small=0), according to (2) The establishment of formula and form variations grade be (it is big=0, in=1, it is small=0), according to the establishment of (3) formula, form variations grade is (big=0, in=0, small=1).Here, to the form variations in each region, S is usedPM=SPPSuch SPMCarry out execution standardization.
[formula 1]
SPP≥50I-UNIT…(1)
[formula 2]
50I-UNIT > SPP≥10I-UNIT…(2)
[formula 3]
10I-UNIT > SPP…(3)
As described above, standardized form deviation 201 and shape of the creation as the input data to neural network 1 01 Deviation levels 202.Standardized form deviation 201 and form variations grade 202 are to control the input data of regular enforcement division 10 S1。
Fig. 8 shows the summaries of control output operational part 3.Control output operational part 3 according to control it is in regular enforcement division 10, Output from neural network 1 01 that is, the control operating side operational order S2 (thing of the shape control of Sendzimir rolling mill In example, AS-U operational degree 301, the 1st intermediary operation degree 302 are equivalent to this), creation controls operating side as to each shape Operational order control operating quantity S3.Here, for there are multiple AS-U operational degrees 301, the 1st intermediary operation degree 302, each data example is shown, 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 degree 301 of input have each AS-U evolution to, close direction Output, so in the poor multiplied by conversion gain G of themASU, thus export the operational order to each AS-U.Conversion gain GASUIt is The position AS-U amount of change (unit is length) is output into the control of each AS-U, so becoming from degree to position amount of change Conversion gain.
In addition, the 1st intermediary operation degree 302 inputted in the same manner has the output of outside, inside among the 1st, so right Their difference is multiplied by conversion gain G1ST, thus export the operational order to the displacement of each 1st intermediate calender rolls.Conversion gain G1STIt is to each The control of 1st intermediate calender rolls is output into the 1st intermediate calender rolls displaced position amount of change (unit is length), so becoming from degree in place Set the conversion gain of amount of change.
It as described above, being capable of operation control operating quantity S3.Operating quantity S3 is controlled by the position #1~#n AS-U amount of change (the saddle number that n is AS-U roller), upper 1st intermediate displaced position amount of change, lower 1st intermediate displaced position amount of change are constituted.In addition, The system that noise data from control gimp generating unit 16 is added to control operating side operational order S2 is shown in Fig. 8.
Fig. 9 indicates the summary of control output determination unit 5.Control output determination unit 5 is repaired by rolling phenomenon model 501 and shape Just good no determination unit 502 is constituted, and obtains the real data Si from control object equipment 1, the control from control output operational part 3 Operating quantity S3 processed and output determine the information of database DB3, and the control operating quantity output of control operating side could be counted by providing According to S4.It, will be to control the control operating quantity that output operational part 3 calculates through the structure in control output determination unit 5 The variation of shape when S3 is output to the rolling mill as control object equipment 1 is input to the mould of known control object equipment 1 Type (in the case where the embodiment of Fig. 9, for roll phenomenon model 501) is predicted, is predicted as inhibiting control in the case that shape deteriorates Operating quantity processed exports SO, prevents shape severe exacerbation.
More specifically, control operating quantity S3 rolling phenomenon model 501 is input to carry out PREDICTIVE CONTROL operating quantity S3 and cause Change in shape, operation form variations correction amount prediction data 503.On the other hand, it is examined in the shape from control object equipment 1 Device data Si (the form variations real data 504 at current time) is surveyed plus form variations correction amount prediction data 503 to obtain Form variations prediction data 505 evaluates form variations prediction data 505, to be output to control pair that will control operating quantity S3 When as equipment 1, can predicting shape how to change.Number is predicted according to the form variations real data 504 and form variations of status According to 505, corrected in good no determination unit 502 in shape, predicting shape is obtained toward good direction change, or toward bad direction change It could data S4 to control operating quantity output.
It is corrected in good no determination unit 502 in shape, specifically carries out the modified good no judgement of shape as described below.It is first First as shown in specification A1, A2 of the priority controlled about shape, the control priority on board width direction is considered, so defeated Determine in database DB3 out, to the weight coefficient w (i) in each specification setting board width direction of specification A1, specification A2.Use it , changed using the evaluation function J of example (4) formula described as follows come predicting shape good no.(4) in formula, w (i) is weight coefficient, ε fb (i) is form variations real data 504, and ε eSt (i) is form variations prediction data 505, and i is shape detector region, Rand is random entry.
[formula 4]
In the case where the evaluation function J of (4) formula of use, evaluation function J is positive when shape improves, evaluation function J when degenerating It is negative.In addition, rand is random entry, change the evaluation result of evaluation function J at random.As a result, in the case where shape deteriorates, Also generating as evaluation function J becomes positive situation, so can also learn in the case where rolling the incorrect situation of phenomenon model 501 The relationship of shape model and control method.Here rand as test running initially, it is unreliable in the model of control object equipment 1 In the case where increase maximum value, think learning control method in a way come in the case where implementing stable control as 0 Mode changes in due course.
It is corrected in good no determination unit 502 in shape, operation evaluation function J, it could to control operating quantity output in J >=0 Data S4=1 (can), when J < 0 output of control operating quantity could data S4=0 (no) mode export control operating quantity output can No data S4.
In control output suppressing portion 4, exported according to the control operating quantity of the judgement result as control output determination unit 5 Could data S4, determine to control object equipment 1 control operating quantity output SO output whether there is or not.Controlling operating quantity output could Data S4 is the output of #1~#n AS-U position amount of change, upper 1st intermediate displaced position amount of change output, the displacement of lower 1st centre The output of position amount of change, is determined by following,
IF (control operating quantity output could data S4=0) THEN
#1~#n AS-U position amount of change output=0
Upper 1st intermediate displaced position amount of change output=0
Lower 1st intermediate displaced position amount of change output=0
ELSE
The position the #1~#n AS-U amount of change output=position #1~#n AS-U amount of change
Upper 1st intermediate displaced position amount of change exports=upper 1st intermediate displaced position amount of change
Lower 1st intermediate displaced position amount of change exports=descends the 1st intermediate displaced position amount of change
ENDIF。
In control executive device 20, according to the real data Si for coming from control object equipment 1 (rolling mill), execute above-mentioned Operation, control operating quantity output SO is output to control object equipment 1 (rolling mill) and implements shape control.
Next, the movement summary to control method learning device 21 is illustrated.In control method learning device 21, Use the delay data of the data utilized in executive device 20.Be delayed Z- 1Mean e- TS, when indicating that delay is preset Between T.Control object equipment 1 has time response, so exporting SO by control operating quantity, exists until actual data change Delay.Therefore, study use be after control operation execution, and real data at the time of by delay time T is implemented.In shape In control, after the operational order output of AS-U, the 1st intermediate calender rolls, the several seconds is needed until shapometer detects change in shape, It is advantageous to be set as T=2 to 3 seconds degree (delay time also changes according to the type of shape detector, mill speed, so The optimal time until the change for controlling operating side is become change in shape is set as T.).
Figure 11 indicates to control the movement summary of good no determination unit 6.In the good no determination unit 602 of change in shape, such as following formula is used Good no judgement evaluation function JC
[formula 5]
(5) in formula, ε fb (i) is contained within the form variations real data of real data Si, and ε last (i) is form variations Real data previous value, wC (i) are the board width direction weight coefficients of good no judgement.Here, the weight system of good no judgement Number wC (i) is based on good no judgement database DB4, sets according to specification A1, A2 of the priority to control.Pass through good no judgement Evaluation function Jc determines the good no of control result.In addition, in the control operation of the judgement result as control output determination unit 5 Amount output could be actually defeated to the control operating quantity of control object equipment 1 in the case that data S4 is 0 (control output can not) Out=0, however it is judged as that shape deteriorates.
Here, control operating quantity output could data S4=0 in the case where, the good no data S6=-1 of control result.Separately Outside, LCL is added and subtracted come preset threshold value upper limit LCU and threshold value based on threshold condition (LCU >=0 >=LCL).At this point, if with good no The comparison result for determining evaluation function Jc is Jc > LCU, then is set as the good no data S6=-1 (shape degenerates) of control result, If LCU >=Jc >=0, it is set as the good no data S6=0 of control result (shape toward bad direction change), if 0 > Jc >=LCL, It is set as the good no data S6=1 of control result (shape toward good direction change), if Jc < LCL, is set as the good no number of control result According to S6=0 (shape improves).
Here, the good no data S6=-1 of control result is that shape degenerates, so in the control output for inhibiting the phenomenon that output Under, the good no data S6=0 of control result is that shape is unchanged or shape improves, so keeping the feelings of the control output of output Under condition, the good no data S6=1 of control result is shape toward good direction change, however has a possibility that further improving, so It is the case where increasing the control amount of output.
In this way, the weight coefficient wC (i) in board width direction changes, institute according to specification A1, A2 of the priority about control It is different with good no judgement evaluation function Jc.It is therefore contemplated that the judgement result of the good no data S6 of control result is also different.Therefore, exist In control method learning device 21, for the priority about control specification A1, A2 both, implement the good no number of control result According to the judgement of S6.
Next, being illustrated to the summary of learning data establishment portion 7.As shown in Figure 1, in learning data establishment portion 7, Based on the judgement result (the good no data S6 of control result) from the good no determination unit 6 of control result, according to the operation of control operating side Instruct S2, control operating quantity S3, control output suppressing portion judgement result (control operating quantity output could data S4), create needle To the monitoring data S7a of the neural network 1 11 used in control rule learning portion 11.
The monitoring data S7a of the situation become the output shown in Fig. 4 as the output layer from neural network 1 11, AS-U operational degree 301,1 intermediary operation degree 302.Output of the learning data establishment portion 7 used as neural network 1 01 It controls operating side operational order S2 (AS-U operational degree 301,1 intermediary operation degree 301), export SO as control operating quantity The output of #1~#n AS-U position amount of change, upper 1st intermediate displaced position amount of change output, lower 1st intermediate displaced position change More amount output, monitoring data S7a of the creation for the neural network 1 11 used in control rule learning portion 11.
In the movement summary for illustrating learning data establishment portion 7, by Fig. 8 control output operational part 3 in each portion's data, It is Figure 12 that the relationship of symbol, which arranges,.Here, the representative control operating side operation shown for the output as neural network 1 01 Instruct S2 AS-U operational degree 301, using the data of operational degree positive side as OPref, operational degree negative side data as OMref, the operational degree being randomly generated from control gimp generating unit 16 are as operational degree random number Oref, transformation Gain controls operating quantity output SO and is illustrated as Cref as G.In this way, here, to put it more simply, will be from control rule Then the output of the output layer of the neural network 1 01 of enforcement division 10 will come from as operational degree positive side and operational degree negative side The operational degree of gimp generating unit 16 being randomly generated is controlled as operational degree random number.In addition, will be for control behaviour Make the control operating quantity output SO at end as operational order value.
Figure 13 indicates processing stage and process content in learning data establishment portion 7.Here, if according to Figure 12 symbol Agreement is illustrated, then in initial processing stage 71, operational order value Cref is found out according to (6) formula.
[formula 6]
Cref=G (OPref-OMref+ORref) ... (6)
In processing stage 72 below, operational order value Cref is modified and is made according to control result good no data S6 For C ' ref.Specifically, according to (7) formula when the good no data S6=-1 of control result, root when control result good no data S6=0 According to (8) formula, according to (9) formula when the good no data S6=1 of control result, it is set as the correction value C ' ref of operational order value Cref.
[formula 7]
[formula 8]
C ' ref=Cref ... (8)
[formula 9]
In processing stage 73, using modified operational order value C ' ref, according to (10), (11) formula finds out operational degree and repairs Positive quantity Δ Oref.
[formula 10]
C′ref=G ((OPref+ Δ Oref)-(OMref- Δ Oref)) ... (10)
[formula 11]
In processing stage 74, the monitoring data OP ' ref to neural network 1 11, OM ' ref are found out according to (12) formula.
[formula 12]
In this way, in learning data establishment portion 7, as shown in figure 12, for the behaviour actually exported to control object equipment 1 Make instruction value Cref, according to the good no data S6 of control result as the judgement result in the good no determination unit 6 of control result, operation Instruction value correction value C ' ref is operated out.Specifically, being to be judged as controlling party in the case where control result good no data S6=1 To OK, but the insufficient situation of output is controlled, operational order value is increased into Δ Cref to identical direction.On the contrary, control result is good The case where no data S6=-1 is that operational order value is reduced Δ round about the case where being judged as steering vector error Cref.Conversion gain G is preset, if so being it is known that therefore known operation degree positive side and operational degree negative side Value, it will be able to find out correction amount Oref.Here simulation etc. is first passed through in advance finds out value appropriate to set Δ Cref.By with On sequence, can will control rule learning portion 11 used in monitoring data OP ' ref, OM ' ref according to above-mentioned (12) formula To find out.
It is illustrated, however in fact, operated for the AS-U to #1~#n AS-U with simple example in Figure 13 Degree 301 and the 1st intermediary operation degree 302 being displaced to the displacement of upper 1st intermediate calender rolls, lower 1st intermediate calender rolls implement its whole, make For monitoring data (the AS-U operational degree monitoring data, 1 centre behaviour of neural network 1 11 used in control rule learning portion 11 Make degree monitoring data).
Figure 14 indicates the data example for being stored in learning data database D B2.For learning neural network 111, many is needed The combination of input data S8a and monitoring data S7a.Therefore, the monitoring data S7a (AS-U created in learning data establishment portion 7 Operational degree monitoring data, the 1st intermediary operation degree) control regular enforcement division 10 with being input to by controlling executive device 20 The delay data S8a of input data S1 (standardized form deviation 201 and form variations grade) combines the study as one group Data S11 is saved in learning data database D B2.
At this point, if such as above-mentioned newly created one group of learning data is only appended to learning data database D B2, The group number of learning data is caused to increase.If learning data increases, not only the study of neural network 1 11 needs the time, also to exist Residual, which becomes, in learning data database D B2 grasps the study of neural network 1 11 for the control of above-mentioned modified input data S1 Make the group of the learning data of the mistake on the basis of end operational order S2, a possibility that study cannot be carried out efficiently is big.
Therefore, in controlling regular forgetting portion 200, infer the learning data of mistake, and deleted out of learning database DB2 It removes, thus prevents learning data from increasing, and improve learning efficiency.
Figure 16 indicates to reflect the structural example of the learning data database D B2 for the processing for controlling regular forgetting portion 200.Study It, will also be to nearest study in addition to the combined learning data as input data S1 and monitoring data T1 in data database DB2 Neural network 1 11 have input input data in the case where neural network output that is, control the S2 conduct of operating side operational order To storing.
Other than the good no data S6=0 of control result, the study number for correcting neural network 1 11 is created According to taking savings in study number so obtaining input data S8a at this time and controlling the delay S2a of operating side operational order S2 According to the correlation of the input data S1 and control operating side operational order S2 of each learning data of database D B2, the degree of correlation is selected Highest learning data S10.The relevant method for taking 2 data series is usually related coefficient, however without being limited thereto, can also be with Using the method for using square, appropriate evaluation function.
The storage content of Figure 17 expression forgetting rules database D B5.It is stored in forgetting rules database D B5 and wants to forget Data S10 and the knowledge newly obtained data S11 become a pair.Want the data S10 forgotten storage as input data S1 and The data S11 storage of the combined learning data of monitoring data T1, the knowledge newly obtained is used as input data S1 and modified number According to the combined learning data of T1.Also the neural network 1 11 learnt recently is inputted with storage in these data S10, S11 Neural network output that is, control operating side operational order S2 in the case where input data.
It controls regular forgetting portion 200 and deletes the learning data S10 of selection from learning data database D B2, and to something lost Forget rule database DB5 write-in learning data S10 and modified learning data S11.It is combined in forgetting rules database D B5 Preserve the good no determination unit 6 of control result be judged to needing modified neural network 1 11 learning data S10 and modified Practise data S11.In addition, the data in forgetting rules database D B5 in the study for executing neural network 1 11, are completed to utilize amendment Learning data study after all delete.
Even if being judged to needing to correct in the good no determination unit 6 of control result, modified learning data S11 is created to add Into learning data database D B2, before the study for executing neural network 1 11, also do not reflect the regular enforcement division 10 of control, So actual control is executed according to original control rule.Displacement controls regular enforcement division 100 and passes through in neural network 1 11 The modified learning data S11 of reflection carrys out the raising of early excise control precision before study.
It is controlled in regular enforcement division 100 in displacement, takes to the input data S1 for controlling regular enforcement division 10 and carry out automatic control The control operating side operational order S2 and savings for making regular enforcement division 10 answer modified in forgetting rules database D B5 The correlation for practising data S10 operates the control operating side of modified learning data S11 in the case where being judged as that the degree of correlation is big Instruction S2 is output to control output operational part 3.
In the plant control unit of Fig. 1, various database D B1, DB2, DB3, DB4, DB5, Figure 14 expression have been used The structure of neural network management table TB for mutual management application each database D B1, DB2, DB3, DB4.Management table TB has The management table of specification.Specifically, management table TB is about specification according to (B1 board width, (B2) steel grade and about the excellent of control Specification A1, A2 first spent divides.As (B1) board width, for example, using 3 feet wide, meter Kuan, be 4 inches wide, 5 feet wide 4 Kind, as steel grade, use 10 kinds or so of steel grade (1)~steel grade (10).In addition, the specification A of the priority about control is A1 And two kinds of A2.In this case, becoming 80 kinds, 80 neural networks are distinguished according to rolling condition and are used.
Neural network lea rning control portion 112 by shown in Figure 14, as combined of input data and monitoring data It practises data and table TB is managed according to the neural network of Figure 15, establish and be associated with corresponding neural network No., store to shown in Figure 16 Learning data database D B2.
When controlling executive device 20 to the execution shape control of control object equipment 1,2 groups of learning datas are created.This is because For identical input data, control output, control result it is good it is no determine using about control priority specification A1 and 2 evaluation criteria of specification A2 carry out, so 2 kinds of monitoring datas of creation.If monitoring data puts aside to a certain degree (such as 200 Group), or new savings to learning data database D B2, then neural network lea rning control portion 112 indicates neural network 1 11 It practises.
In control rule database DB1, management table TB according to Figure 15 stores multiple neural networks, in nerve Specified to need the neural network No. that learns in e-learning control unit 112, neural network selector 113 is from control regular data Library DB1 takes out the neural network, is set in neural network 1 11.Neural network lea rning control portion 112 is to input data establishment portion 114 and the instruction of monitoring data establishment portion 115 from learning data database D B2 take out corresponding with neural network input data And monitoring data, implement the study of neural network 1 11 using them.Further it is proposed that the learning method of various neural networks, Arbitrary method can be used.
If completing the study of neural network 1 11, neural network lea rning control portion 112 is by the nerve as learning outcome Network 111 writes back to the position of the neural network No. of control rule database DB1, to complete to learn.
The whole neural networks for learning to define Figure 15 are implemented simultaneously in regular intervals (such as every 1 day), It can also be at the moment only to the new learning data savings neural network of the neural network No. of (such as 100 groups) to a certain degree Study.
As above, do not make the shape large change ground as the rolling mill of control object equipment 1, be accomplished by
1) in advance be set separately reference figure mode and in view of this control operation, be not study control operating method, But learn the combination of shape model and control operation, implement control operation using them.
2) also there is new control rule that can not expect in advance, but completely inscrutable control rule becomes most Good situation finds out the control result of observation in view of this so randomly making to control operating side movement to find.
3) make the study shortest time for the neural network for needing the time by inhibiting the increase of learning data, and can It is controlled before completing study using new control rule.
Neural network used in executive device 20 is controlled in addition, being stored in control rule database DB1, however if The neural network of storage is that the network of preliminary treatment is implemented with random number, then carries out the study of neural network, until can It carries out needing the time until controlling accordingly.Therefore, to control object equipment 1, when constructing control unit, based on sentencing at the moment The Controlling model of bright control object equipment 1 first passes through simulation in advance, implements the study of control rule, will complete in simulator The neural network of study is stored into database, so as to implement a degree of property from the initial start stage of control object equipment The control of energy.
In addition, by the way that learning data number is limited in certain value, the time required to study can be shortened.
In addition, reducing the class in learning data database D B2 using method identical with regular forgetting portion 200 is controlled As learning data, so as to reduce learning data quantity.
Plant control unit of the invention can actually be used as computer system to realize, however at this time in computer system It is interior to form multiple program groups.
These program groups are, for example: for realizing control executive device processing, reality according to control object equipment The control rule that the combination of the decision of data and control operation provides control output executes program;Determine that control rule executes program Output control output could, and by the real data and control operation for mistake situation notify to the above-mentioned controlling party science of law The control for practising device exports decision procedure;The case where control output decision procedure exports control output to control object equipment Under, in the case where being judged as that the above-mentioned real data of control object equipment deteriorates, prevents to export to above-mentioned control object equipment and control The control output of system output inhibits program.
These program groups are: for realizing control method learning device processing, control executive device it is practical to control In the case where object-based device output control output, after the delay until control effect is revealed in real data, sentence Determine real data to improve compared with before the control, the processing of the good no judgement of good no control result of the control result still to degenerate The good no decision procedure of control result;Use the good no and control output of the control result in the good no decision procedure of the control result Obtain the learning data creation program of monitoring data;Real data and control before the amendment for the learning data deleted and created are grasped The control rule forgetting portion program of learning data as the composite class of work;Using above-mentioned real data and above-mentioned monitoring data as The control rule learning program for practising data to learn.
Moreover, control method learning device passes through study, according to the state of above-mentioned control object equipment to multiple control mesh Mark obtains the combination of independent real data and control operation, using the combination of obtained real data and control operation as above-mentioned Control rule executes the combination of the real data of the control object equipment in program and the decision of control operation to use, and is controlling In the case of as the state of object-based device and the real data before amendment and the composite class of control operation, revised study is used Data implement control.
When apparatus of the present invention are suitable for physical device, need to preset the initial value of neural network, however about this Point can create reality by simulating using the Controlling model of control object equipment before implementing the control in control object equipment The real data in control object equipment and the combined learning period of control operation are shortened in the combination of border data and control operation Between.
In the present invention as described above, in the apparatus control system by study Correction and Control rule, for height Control is implemented on effect ground, and the data of creation for deleting the study for meeting modified control rule (forget unnecessary actual number According to) method and learning again for control rule will not be implemented immediately, so temporarily using modified before proposing implementation study The method of control method.In accordance with the invention it is possible to efficiently carry out the study of control rule, and can prevent from implementing control rule The control method for learning to use pervious control effect few before again then.Therefore, control precision improves, the starting of control unit The effect that period shortens, copes with aging etc..
As the background constituted as in the present invention, in the case where having used Deep Learning in the study of control rule Time (several hours ranks) are needed, thus it is not all right in the method for additional learning data, it needs to delete using either method Except learning data (forgetting), and need with from during division come method that throw away the method for the learning data gradually increased different To manage.
Industrial utilizability
Control method and control unit of the present invention for example about one rolling mill as rolling equipment, without special Practical application when the problem of.

Claims (12)

1. a kind of plant control unit identifies the combined mould of the real data of control object equipment for control object equipment Formula, to implement control, which is characterized in that
The plant control unit has the real data of study control object equipment and the combined controlling party science of law of control operation It practises device and implements the control of the control of control object equipment according to the combination of the above-mentioned real data and control operation that are learnt Executive device processed,
Above-mentioned control executive device has: mentioning according to the combination of the real data of control object equipment and the decision of control operation For the control rule enforcement division of control output;Determine the control output of control rule enforcement division output could, and to above-mentioned Control method learning device notifies the real data and controls the control output determination unit of operating mistake;It is judged as above-mentioned control pair In the case where deteriorating as the above-mentioned real data of equipment, the control that above-mentioned control output is exported to above-mentioned control object equipment is prevented Suppressing portion is exported,
Above-mentioned control method learning device has: exporting the feelings of control output to control object equipment in above-mentioned control executive device Under condition, after the delay until control effect appears to real data, the control knot of the real data compared with before control is determined The good no determination unit of good no control result of fruit;Use the good no and above-mentioned of the control result in the good no determination unit of the control result Control output obtains the learning data establishment portion of monitoring data;Delete with creation learning data amendment before real data and The control rule forgetting portion of learning data as controlling the composite class of operation;Using above-mentioned real data and above-mentioned monitoring data as The control rule learning portion of learning data study,
Above-mentioned control method learning device obtains multiple control targets according to the state of above-mentioned control object equipment by study The combination of independent real data and control operation advises the combination of obtained real data and control operation as above-mentioned control Then the real data of the control object equipment in enforcement division is used with the combination for controlling the decision operated, in control object equipment State and amendment before real data and control operation composite class as in the case of, use revised learning data real Apply control.
2. plant control unit according to claim 1, which is characterized in that
In order to according to the size of the real data of control object equipment, replace the combination of real data and control operation, using with The relevant information of the size of real data and the information that real data standardization is made it easy to implement pattern-recognition, study are practical The combination of data and control operation, is controlled.
3. plant control unit according to claim 1 or 2, which is characterized in that
Above-mentioned control rule enforcement division is using the combination of the real data of control object equipment and the decision of control operation as the 1st mind It is kept through network, above-mentioned control rule learning portion keeps the combination of real data and control operation as the 2nd neural network, will 2nd neural network obtained from the result of study in above-mentioned control method learning device is as in above-mentioned control rule enforcement division Above-mentioned 1st neural network come using.
4. plant control unit according to claim 1 or 2, which is characterized in that
Above-mentioned control executive device has the control gimp generating unit for exporting to above-mentioned control and applying noise, the imperial side of above-mentioned system The case where when calligraphy learning device will be applied with noise, learns with being also included.
5. plant control unit according to claim 1 or 2, which is characterized in that
Above-mentioned control method learning device obtains real data and control behaviour by the study based on pre-determined multiple specifications The multiple combinations made, above-mentioned control executive device is from multiple combinations of real data and control operation according to control object equipment Operating condition select multiple groups of merging of real data and control operation to provide above-mentioned control output.
6. plant control unit according to claim 3, which is characterized in that
According to the size of real data, the combined neural network of real data and operating method that change study uses.
7. plant control unit according to claim 1 or 2, which is characterized in that
The experience of the operator of state or control object equipment based on above-mentioned control object equipment, changes control result Good no determinating reference finds out the relationship of the real data and maneuver for control object equipment respectively, is stored into data respectively Library, thus according to the experience of the operator of the state of above-mentioned control object equipment or control object equipment, with different controlling parties Method controls.
8. plant control unit according to claim 1 or 2, which is characterized in that
Before implementing control in control object equipment, above-mentioned reality is created by simulation using the Controlling model for making imperial object-based device The above-mentioned real data in control object equipment and the combined learning period of control operation are shortened in the combination of data and control operation Between.
9. a kind of rolling mill control device using plant control unit of any of claims 1 or 2, which is characterized in that
Above-mentioned control object equipment is rolling mill, and above-mentioned real data is the outlet side shape of above-mentioned rolling mill.
10. a kind of apparatus control method is to identify control object equipment for control object equipment in plant control unit The combined mode of real data, to implement the apparatus control method of control, which is characterized in that
Above equipment control device has the real data of study control object equipment and the combined control method of control operation Learning device and the control for implementing according to the combination of the above-mentioned real data that is learnt and control operation control object equipment Executive device is controlled,
Above-mentioned control executive device: control is provided according to the combination of the real data of control object equipment and the decision of control operation System output;Determine the control output could, and to above-mentioned control method learning device notify the real data and control operate Mistake;In the case where being judged as that the above-mentioned real data of above-mentioned control object equipment deteriorates, prevent to above-mentioned control object equipment Above-mentioned control output is exported,
Above-mentioned control method learning device: the case where above-mentioned control executive device exports control output to control object equipment Under, after the delay until control effect appears to real data, determine the control result of the real data compared with before control It is good no;It exports to obtain monitoring data using the good no and above-mentioned control of the control result;It deletes and the learning data of creation Real data before amendment and learning data as the composite class of control operation;Above-mentioned real data and above-mentioned monitoring data are made Learn for learning data, by study, multiple control targets is obtained according to the state of above-mentioned control object equipment independent The combination of real data and control operation, using the combination of obtained real data and control operation as above-mentioned control executive device In the real data of control object equipment and the combination of decision of control operation use, control object equipment state with In the case of as the composite class that real data and control before amendment operate, implement control using revised learning data.
11. a kind of rolling mill control method using apparatus control method described in any one of claim 10, which is characterized in that
Above-mentioned control object equipment is rolling mill, and above-mentioned real data is the outlet side shape of above-mentioned rolling mill.
12. a kind of recording medium, be stored with it is that computer system-readable takes, for being realized by the computer system for controlling Object-based device processed identifies the combined mode of the real data of control object equipment, come when implementing the plant control unit controlled Program, which is characterized in that,
Computer system has the real data of study control object equipment and the combined control method study dress of control operation It sets and is held according to the combination of the above-mentioned real data and control operation that are learnt to implement the control of the control of control object equipment Luggage is set,
Above procedure is the processing for realizing above-mentioned control executive device, real data and control according to control object equipment The control rule for making the combination offer control output of the decision of operation executes program;Determine that the control rule executes program output Control output could, and by the real data and control operate for error notification to above-mentioned control method learning device control Export decision procedure;In the case where the control exports decision procedure and exports control output to control object equipment, it is judged as control In the case that the above-mentioned real data of object-based device processed deteriorates, the control that control output is exported to above-mentioned control object equipment is prevented Output inhibits program,
Above procedure is the processing for realizing above-mentioned control method learning device, control executive device actually to control object In the case where equipment output control output, after the delay until control effect is revealed in real data, determine real Border data improve compared with before control, the control of the processing of the good no judgement of good no control result of the control result still to degenerate As a result good no decision procedure;It is supervised using the good no and control output of the control result in the good no decision procedure of the control result Superintend and direct the learning data creation program of data;Delete the group with real data and control operation before the amendment of the learning data of creation Close the control rule forgetting portion program of similar learning data;Using above-mentioned real data and above-mentioned monitoring data as learning data Come the control rule learning program learnt,
Control method learning device obtains independence to multiple control targets according to the state of above-mentioned control object equipment by study Real data and control operation combination, using obtained real data and control operation combination held as above-mentioned control rule The real data of control object equipment in line program is used with the combination for controlling the decision operated, in control object equipment In the case of as the composite class that real data and control before state and amendment operate, implemented using revised learning data Control.
CN201810804376.6A 2017-08-01 2018-07-20 Equipment control device, rolling mill control device, control method thereof, and recording medium Active CN109317522B (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2017-149011 2017-08-01
JP2017149011A JP6899273B2 (en) 2017-08-01 2017-08-01 Plant control device and its control method, rolling mill control device and its control method and program

Publications (2)

Publication Number Publication Date
CN109317522A true CN109317522A (en) 2019-02-12
CN109317522B CN109317522B (en) 2020-02-14

Family

ID=65020188

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810804376.6A Active CN109317522B (en) 2017-08-01 2018-07-20 Equipment control device, rolling mill control device, control method thereof, and recording medium

Country Status (3)

Country Link
JP (1) JP6899273B2 (en)
CN (1) CN109317522B (en)
DE (1) DE102018212140A1 (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112767380A (en) * 2021-01-28 2021-05-07 东北大学 Wide and thick plate end shape prediction method
CN113050550A (en) * 2019-12-27 2021-06-29 株式会社日立制作所 Plant control device and control method thereof, rolling mill control device and control method thereof
CN115516472A (en) * 2020-05-20 2022-12-23 三菱电机株式会社 Data creation device, machine learning system, and machining state estimation device

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP7156982B2 (en) * 2019-03-11 2022-10-19 株式会社日立製作所 Real-time controllers and distributed control systems using them, industrial machines
JP7179700B2 (en) * 2019-08-30 2022-11-29 株式会社日立製作所 plant controller
JP7396117B2 (en) * 2020-02-27 2023-12-12 オムロン株式会社 Model update device, method, and program
US20220333857A1 (en) * 2020-05-27 2022-10-20 Chiyoda Corporation Method and system for determining operating conditions of liquefied natural gas plant
CN113569343A (en) * 2021-05-19 2021-10-29 苏州大学 Method, device and system for controlling rolling thickness of plate and storage medium

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1607045A (en) * 2003-10-15 2005-04-20 株式会社日立制作所 Rolling control device and rolling control method
CN1609736A (en) * 2003-10-22 2005-04-27 株式会社日立制作所 Controlling device, system utilizing the same controller and controlling method
CN101885003A (en) * 2009-05-14 2010-11-17 株式会社日立制作所 The control device of rolling equipment and control method
JP4878340B2 (en) * 2007-08-20 2012-02-15 株式会社神戸製鋼所 Method for preventing meandering of metal sheet
CN103376085A (en) * 2013-07-05 2013-10-30 燕山大学 Intelligent and comprehensive evaluation method for cold-rolled strip steel plate shapes
CN103394521A (en) * 2013-08-02 2013-11-20 中冶南方工程技术有限公司 Method for controlling strip shape of cold-rolled strip steel

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPS633733A (en) 1986-06-23 1988-01-08 株式会社クボタ Water feed control unit for crops
JP2804161B2 (en) 1990-06-04 1998-09-24 株式会社日立製作所 Method and apparatus for controlling shape of Sendzimir mill
JP2677964B2 (en) * 1994-03-11 1997-11-17 川崎製鉄株式会社 Rolling mill shape control initial setting method
JP2763495B2 (en) * 1994-10-24 1998-06-11 三菱電機株式会社 Rolling equipment
JP3244982B2 (en) * 1995-01-30 2002-01-07 株式会社日立製作所 Distributed control system
JP3223856B2 (en) * 1997-04-17 2001-10-29 日本鋼管株式会社 Rolling mill control method and rolling mill control device
JP2000056807A (en) * 1998-08-11 2000-02-25 Hitachi Ltd Model prediction controller
JP3646726B1 (en) * 2004-04-14 2005-05-11 有限会社山口ティー・エル・オー Disaster prevention business support system
JP6109035B2 (en) * 2013-10-10 2017-04-05 株式会社神戸製鋼所 Rolling process support system

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1607045A (en) * 2003-10-15 2005-04-20 株式会社日立制作所 Rolling control device and rolling control method
CN1609736A (en) * 2003-10-22 2005-04-27 株式会社日立制作所 Controlling device, system utilizing the same controller and controlling method
JP4878340B2 (en) * 2007-08-20 2012-02-15 株式会社神戸製鋼所 Method for preventing meandering of metal sheet
CN101885003A (en) * 2009-05-14 2010-11-17 株式会社日立制作所 The control device of rolling equipment and control method
CN103376085A (en) * 2013-07-05 2013-10-30 燕山大学 Intelligent and comprehensive evaluation method for cold-rolled strip steel plate shapes
CN103394521A (en) * 2013-08-02 2013-11-20 中冶南方工程技术有限公司 Method for controlling strip shape of cold-rolled strip steel

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113050550A (en) * 2019-12-27 2021-06-29 株式会社日立制作所 Plant control device and control method thereof, rolling mill control device and control method thereof
CN113050550B (en) * 2019-12-27 2024-04-16 株式会社日立制作所 Plant equipment control device and method thereof, rolling mill control device and method thereof
CN115516472A (en) * 2020-05-20 2022-12-23 三菱电机株式会社 Data creation device, machine learning system, and machining state estimation device
CN112767380A (en) * 2021-01-28 2021-05-07 东北大学 Wide and thick plate end shape prediction method
CN112767380B (en) * 2021-01-28 2024-02-02 东北大学 Prediction method for end shape of wide and thick plate

Also Published As

Publication number Publication date
JP2019028824A (en) 2019-02-21
CN109317522B (en) 2020-02-14
DE102018212140A1 (en) 2019-02-07
JP6899273B2 (en) 2021-07-07

Similar Documents

Publication Publication Date Title
CN109317522A (en) Plant control unit, rolling mill control device and its control method, recording medium
JP6813416B2 (en) Plant control device and its control method, rolling mill control device and its control method and program
US20200327411A1 (en) Systems and Method on Deriving Real-time Coordinated Voltage Control Strategies Using Deep Reinforcement Learning
DE102016011532A1 (en) Machine learning device and machine learning method for optimizing the frequency of tool correction of a machine tool and machine tool with the machine learning device
JP7033639B2 (en) Plant control device and its control method, rolling mill control device and its control method and program
JP7034046B2 (en) Control device and control method
JP4047875B2 (en) Method for determining optimum values of design parameters and method for determining rolling pass schedule using this method
Sanz-García et al. Parsimonious support vector machines modelling for set points in industrial processes based on genetic algorithm optimization
CN109409780A (en) Changing process method, device, computer equipment and storage medium
CN106067075B (en) Building energy load prediction model building and load prediction method and device
JP6952018B2 (en) Control device and control method
KR102440698B1 (en) Plant control device and controlling method thereof, rolling mill control device and controlling method and program thereof
CN105956680A (en) Frame for generating and managing adaptive rule based on reinforcement learning
Ding et al. Multi-policy deep reinforcement learning for multi-objective multiplicity flexible job shop scheduling
Fan et al. Survey of the selection and evaluation for dispatching rules in dynamic job shop scheduling problem
Gornitzka et al. Making sense of change in university governance
JP2022183827A (en) Plant control system, rolling machine control device, plant control method, and plant control program
Gwiazda et al. Parallel processes planning using computer supporting system
JP7495874B2 (en) PLANT CONTROL SYSTEM, PLANT CONTROL METHOD, AND PROGRAM
Suleimenov et al. Concept of developing an intelligent system for control and operational diagnostics of technological equipment condition
JP2022182538A (en) Plant control system, plant control method and program
Garg Predicting Uncertain Behavior in Critical Engineering Systems Under Vague Environment.
Zhang et al. Optimization Algorithm for Construction Procedure Scheduling in Prefabricated Construction Based on NSGA-III
Komal et al. RAM analysis of the press unit in a paper plant using genetic algorithm and Lambda-Tau methodology
CN117013547A (en) Multi-layer scheduling method based on power system active correction intelligent model

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant