Supervision data generation device and supervision data generation method

[0039] First, the basic concept included in the present embodiment will be described.

[0040] In equipment control, actual phenomena that cannot be modeled, operator skills based on past experience, etc. are included in the past operation performance data of equipment. Therefore, it is effective for optimal equipment control to collect past operation performance data of equipment, extract control rules offline, and use the extracted control rules as supervisory data for learning. A control rule is information that associates shape outputs (state quantities) with operations (operation quantities). Here, an implementation procedure for applying the results of machine learning performed by AI offline using supervised data generated based on previously accumulated past facility operation performance data to an actual machine will be described below.

[0041] (1) Collect the actual operation performance data of the equipment.

[0042] (2) Supervision data used in machine learning of AI is extracted from the actual operation performance data of the equipment.

[0043] (3) Implement machine learning of AI using supervised data.

[0044](4) Verification of the results of machine learning of AI using supervised data for experiments.

[0045] (5) Verification of the control performance of the AI ​​completed by learning based on simulation or the like.

[0046] (6) Verify the real-time control performance of the actual machine based on the equipment.

[0047] The present embodiment focuses on the above (2). After that, by implementing (3) to (6), the control performance can be verified in the previous stage of application to an actual device.

[0048] In the present embodiment, control rules are generated based on actual operation performance data of equipment, and the generated control rules are evaluated. Also, new supervision data is generated according to the control rules whose evaluation results are good.

[0049] The operating performance data of equipment includes noise, and there is often an influence of disturbance on the control system. In the present embodiment, the influence of these noises and disturbances can be suppressed, and supervisory data can be efficiently generated from the actual operation performance data.

[0050] In the present embodiment, control rules with good evaluation results are extracted from data having a certain amount of operation among the actual operation performance data of equipment, and monitoring data are generated. In the present embodiment, it is possible to suppress the influence of the variation in the time delay of the control result, appropriately evaluate the change of the equipment state by the control, and extract a control rule with a favorable evaluation.

[0051] In the present embodiment, a supervised data database for learning storing supervised data used in machine learning of AI is constructed using the control rules generated from the operation performance data of the equipment. In the present embodiment, the existing supervised data database for learning can be updated with new supervised data generated based on the control rules.

[0052] Control based on results learned from actual performance data using AI such as a neural network is an inductive method based on learned data. Therefore, control reliability is reduced outside the range that can be handled by learned data. The present invention has the following functions: in order to expand the applicable range of control, to expand the range of states that can be corresponded to the AI ​​learning supervision data database, and to construct a learning supervision data database according to control rules corresponding to more states. , to calculate the value of each control rule.

[0053] When the number of supervisory data stored in the supervisory data database for learning increases, the amount of computation for machine learning using the supervisory data increases, and the computation time also increases. In this embodiment, in order to keep the computation amount and computation time required for the learning constant, the amount of supervised data in the supervised data database for learning is managed. supervised data to update the least valuable supervised data. That is, in the present embodiment, the number of supervision data can be kept constant.

[0054] In the present embodiment, the time for measuring the values ​​of the operation and the state change is extended as a measure to reduce the influence of noise included in the actual operation performance data of the equipment and the influence of the deviation of the time delay between the operation and the state change.

[0055] Considering the influence of noise, the control cycle of shape control is in the range of 0.5 seconds to several seconds, and the operating range of the shape control mechanism that can be controlled per cycle is not so large. The movable range of the shape control mechanism (AS-U, IMR) of the rolling mill is several mm. On the other hand, even in a fixed rolling state, the shape continues to change by a fixed amount. This is due to the transmission of force between the material and the rolling mill, the introduction of oil, the quality deviation of the material, etc., the degree of elongation of the product is constantly fluctuating. This is considered as noise when performing shape evaluation. Therefore, for a small change in shape caused by a small operation of controlling one cycle, the change that always occurs may be larger, and it is difficult to extract only the pure shape change component relative to the operation. As a method for addressing these problems, it is considered that the shape is evaluated as the sum of the operation results for a certain fixed period, and the evaluated result is allocated to the operation performed in the period. At this time, in order to eliminate the influence of noise, a dead zone (a range outside the evaluation target) is set in the shape evaluation, and if it is included in the dead zone, it is considered that the shape does not change.

[0056] Considering the time delay, in the rolling mill, the shape control is periodically performed until a time delay occurs until a shape change actually generated after the shape control is detected. In addition, since new control is performed after a fixed cycle, the shape change that actually occurs is not only the shape change of a certain cycle, but also the shape change based on the overlapping of the effects produced by a series of operations. As the reasons for such a time delay, the delay until the part rolled by the shape control mechanism reaches the measurement position (movement delay), and the delay from the operation input to the shape control mechanism to the completion of the operation of the shape control mechanism (operation delay) are considered. , The delay from the completion of the operation of the shape control mechanism until the rolling state is stabilized (rolling stabilization delay), etc. Figure 12 These delayed images are shown. The above-mentioned movement delay is determined by the movement speed of an object to be rolled, such as a rolled steel sheet. The above-mentioned operation delay needs to take into account the shape change in the operation of the shape control mechanism. The rolling stabilization delay mentioned above is an uncertain element. Therefore, the shape change according to the operation of the shape control mechanism is generated in stages with a certain delay from the input of the operation.

[0057] In the present embodiment, only control rules including operations effective for control are extracted from offline operation performance data, and used to generate supervisory data. At this time, if the operation amount of the operating device (ie, the shape control mechanism) in the control cycle is small, the operation amount itself has noise, and the state change amount also has noise influence, and the evaluation of the operation becomes unreliable. In addition, there is a possibility that the time relationship between the operation and the state change may be incorrect, and the possibility of extracting the wrong relationship is high. On the other hand, the data is not compared in one cycle, but is set as a shape evaluation interval ( Figure 13 ), a large change in state quantity corresponding to a large operation can be evaluated, and the influence of time delay variation is only at the start and end of the extraction time, so it can be suppressed as a ratio.

[0058] In addition, as a method of updating the supervision data, it is considered to generate a control rule using the sum of the operation quantities generated in the operation quantity collection section including a plurality of cycles, and add it to the database as new supervision data. In this case, normalization of the control rule for setting the sum of the operation quantities to one cycle is required. That is, it may be normalized so that the sum of the operation amounts is included in the operation range of the mechanism that moves in one cycle. In addition, in consideration of the time delay of the shape change, it is considered appropriate to lower the weight of the operation immediately before the shape change and distribute it equally to other operations. The evaluation section of the shape and the collection section of the operation amount are shifted by the time delay amount. Figure 13 An image showing the shape evaluation section and the operation collection section.

[0059] In order to construct AI that can correspond to many states with a limited set of actual performance data, the collection of supervisory data is improved. Therefore, supervised data is sorted according to value, and supervised data is removed from data with low order. At this time, the degree of similarity between the supervision data is measured as a ranking criterion, and it is considered that the higher the degree of similarity of the supervision data, the lower the value, so that it is possible to create a supervision data database for learning corresponding to as many states as possible.

[0060] When appending new supervision data, it is sorted according to the value containing the existing supervision data, and the data with lower order is deleted. At this time, the value as the supervision data is set to have no other similar rules (approximate degree) and a large effect (effectiveness) with respect to the operation. Calculate the approximation, that is, the distance between the supervised data as an index for sorting, extract the two closest supervised data according to the calculated distance, and evaluate the operation result of each supervised data as the validity of the two supervised data. Compare and decide which supervised data to delete.

[0061] As for the distance between the supervision data, the distance to all the other supervision data is calculated for one supervision data stored in the learning supervision data database, and the minimum value (minimum distance d) is set. The distance between the supervision data is obtained by taking the squares of the deviations of the input data (shape deviation) and output data (operation amount) of the supervision data, weighting them respectively, and adding them together. The supervised data is sorted according to the minimum distance d calculated above, and the lowest supervised data is deleted from the supervised data database for learning. At this time, since there must be two supervised data to be the minimum distance d, the one with the lower evaluation value of the operation result is deleted. In addition, in the above-described method, the supervisory data for which the operation result evaluation value is erroneously judged to be good may continue to remain without being deleted. To prevent this, the operation result evaluation value may also be reduced by multiplying by a predetermined coefficient K (0

[0062] In the above example, the control rule adopted as new supervision data can be expressed as (state amount)|(operation amount), for example. At this time, the operation amount may be corrected based on the operation result evaluation value and added as learning data. That is, it can also be set to

[0063] When the operation result evaluation value is good: (state amount)|(operation amount)×α(α>1)

[0064] When the evaluation value of the operation result is normal: (state quantity)|(operation quantity)×β(0

[0065] When the operation result evaluation value is bad: (state amount)|(operation amount)×γ(0

[0066] In equipment control, even if the input state is approximately the same, the output sometimes changes, and the same is true when it is close to the optimal state. In order to maintain the optimal state, it is important to deal with fine output changes in a small area near the optimal state. Therefore, in the present embodiment, it is possible to perform management such that the value of the supervision data is weighted by the distance from the optimal state, thereby increasing the density of the supervision data close to the optimal state, and reducing the number of deviations from the optimal state. Supervise the density of the data.

[0067] The present embodiment relates to a method of generating supervisory data used for AI machine learning based on control rules generated from the collected equipment operation performance data having the above-described functions.

[0068] According to the present embodiment, it is possible to construct an artificial intelligence learning supervision data database based on the actual operation performance data of the equipment. Therefore, the AI ​​applied to real-time equipment control can learn control rules offline using past equipment operation performance data, and the performance can also be verified offline. Thereby, the reliability of control is improved, and the application of real-time control using AI can be realized.

[0069] In addition, according to the present embodiment, it is possible to extract new supervisory data having a good control rule that can be expected to be appropriate based on the operation performance data of the equipment including noise and having a deviation in the time delay of the operation and the result of the operation. It evaluates the operation results and implements effective control by letting the AI ​​learn. A monitoring data database for AI learning can be constructed from a large amount of equipment operation performance data. The more equipment operation performance data, the higher the control accuracy using AI. In addition, since the amount of supervision data can be kept constant, it is possible to suppress the expansion of computation time required for AI learning.

[0070] (Example)

[0071] Hereinafter, the configuration of the supervisory data generating apparatus according to the embodiment will be described with reference to the drawings.

[0072] The supervisory data generating apparatus according to the present embodiment analyzes the actual equipment operation performance data, and generates supervisory data used in the learning of artificial intelligence for controlling the above-mentioned equipment, and the above-mentioned equipment operation actual performance data includes the above-mentioned equipment based on the state of the equipment. The operating equipment controls the relevant state quantities and the operation quantities of the control mechanism.

[0073] figure 1 The schematic functional blocks of the supervisory data generating apparatus according to the embodiment are shown. figure 1The supervising data generating means of the 20 have new supervising data extracting means 20 and supervising data database updating means 31 . The supervisory data generating apparatus is configured to include a computer that executes a program stored in the storage apparatus, whereby the computer functions as the new supervisory data extracting apparatus 20 and the supervisory data database updating apparatus 31 . The supervisory data generated by the supervisory data generating means is used for learning of the device control means having a neural network as AI.

[0074] The new supervisory data extracting device 20 refers to the operating speed v(t), the plate shape state quantity S(t), and the operation performance data stored in the facility operating performance data database DB5 (hereinafter simply referred to as "operating performance DB5") as equipment operating performance data. The machine state quantity is O(t), and new supervision data Tnew is created. t represents a moment or a moment-based parameter.

[0075] The supervisory data database updating means 31 acquires new supervisory data Tnew from the new supervisory data extracting means 20 . In addition, the supervisory data database updating means 31 acquires the supervisory data number counter N and supervisory data T(1), ···, T(N stored in the supervisory data database DB2 for neural network learning (hereinafter simply referred to as "supervisory DB2") ). Then, the supervisory data database updating means 31 updates the acquired supervisory data number counter N and supervisory data T(1), . . . , T(N), and overwrites them in the supervisory DB2. Every time the supervision data T is added, the supervision data quantity counter N is sequentially counted from 1 to the maximum number NMAX of supervision data storage.

[0076] figure 2 Schematic functional blocks of the new supervisory data extraction apparatus 20 are shown. The new supervisory data extraction device 20 constitutes, as main elements, an actual performance data reading speed condition setting unit 200, an actual performance data reading counter setting unit 201, a supervisory data input unit state quantity reading unit 202, and an actual performance data noise countermeasure extraction time width Setting unit 203 , supervisory data output unit operation quantity reading unit 204 , post-operation state quantity change delay time setting unit 205 , operation result evaluation unit 206 , new supervisory data generation condition setting unit 207 , new supervisory data generation judgment unit 208 and a new supervision data generation unit 209.

[0077] The actual performance data reading counter setting unit 201 acquires the running speed v(t) from the actual running performance DB 5 , and sets the actual performance data reading counter t. Specifically, the actual performance data reading counter setting unit 201 sequentially reads the running speed v(t) stored in the running actual performance DB 5 while advancing the actual performance data reading counter t from the beginning. That is, the running speed v(t) is sequentially read from the part with the earliest time (earlier in time). Then, when the following reading conditions are satisfied, the actual performance data reading counter t at that time is output.

[0078] Figure 4 The operation speed pattern of the Sendzimir rolling mill targeted in this embodiment is shown. The actual performance data reading counter setting unit 201 is equal to or greater than the actual performance data reading speed condition THv set by the actual performance data reading speed condition setting unit 200 (that is, the condition speed extracted as the supervision data, that is, the supervision data extraction speed THv). In addition, the fact that the running speed is maintained for a fixed period is used as the above-mentioned reading condition, and the actual performance data reading counter t is output.

[0079] The monitoring data input unit state quantity reading unit 202 uses the actual performance data reading counter t output from the actual performance data reading counter setting unit 201 to read the plate shape state quantity S(t) from the actual operation performance DB 5 . In the present embodiment, the plate shape state quantity S(t) includes the plate shape actual performance value spfb and the target plate shape spref detected by the shape detector at time t.

[0080] Figure 5 An example of the plate shape state quantity S(t1) in the actual performance data reading counter t(t=start time t1) is shown. exist Figure 5 Among them, DS represents the drive side, and WS represents the workpiece side. In the upper-level graph, the horizontal axis represents the position i{i=1, . shape spref(i). Here, Ch represents the total number of shape detectors used in the board width direction. In the lower-level graph, the horizontal axis represents the position i of the shape detector in the plate width direction, and the vertical axis represents the shape deviation spdev(i). The shape deviation spdev(i) is obtained by the following formula (1). The following equation (2) represents the supervision data input unit Tin(t1), which is a set of shape deviations spdev(i) in the plate width direction at time t1.

[0081] [Formula 1]

[0082] spdev(i)=spfb(i)-spref(i)...(1)

[0083] Tin(t1)={spdev(1), spdev(2),...,spdev(Ch)}...(2)

[0084] The supervisory data output unit operation amount reading unit 204 uses the actual performance data reading counter t from the actual performance data reading counter setting unit 201 and the extraction time width tband [sec] acquired from the actual performance data noise countermeasure extraction time width setting unit 203 , and read the operating machine state quantities O(t) to O(t+tband) from the actual operation performance DB5. In this embodiment, the operating device state quantity O(t) includes the operating device state quantity Pj(t) of each operating device at time t {j represents the number of the shape control mechanism, j=1, . . . , 10}.

[0085] Image 6 An example of the temporal change of the operating device state quantity Pj(t) is shown. Pj(t) represents the operation amount of the shape control mechanism (number j) at time t. In the present embodiment, each number j corresponds to each shape operating mechanism as shown on the horizontal axis of the lower level graph. That is, the shape manipulation mechanisms indicated by the use numbers 1 to 7 are "AS-U#1" to "AS-U#7". The shape manipulation mechanism indicated with reference number 8 is "Top IMR shift", and the shape manipulation mechanism indicated with reference number 9 is "Bot IMR shift". Use the shape manipulation mechanism "Leveling" denoted by number 10.

[0086] From the extraction time width tband and Pj(t) set by the actual performance data noise countermeasure extraction time width setting unit 203, the shape control mechanism operation amount of the shape control mechanism (number j) at time t is obtained by the following equation (3) Oj(t). Here, tcyc represents the actual performance data sampling period.

[0087] [Formula 2]

[0088] Oj(t1)=(Pj(t2)-Pj(t1))/tband*tcyc...(3)

[0089] Among them, t2=t1+tband

[0090] Here, the extraction time width tband sets the average time required for the shape-improving operation. When it can be determined that the shape is not affected by noise and can be improved by operation, the standard is about ±5 [I-unit], and each shape control required for this can be obtained from equations (4) to (6) using the following parameters The operating time tj of the mechanism.

[0091] CA: AS-U reference operation amount [mm/I-unit]

[0092] VA: AS-U action speed [mm/sec]

[0093] CI: IMR shift reference operation amount [mm/I-unit]

[0094] VI: IMR shift action speed [mm/sec]

[0095] CL: Leveling reference operation amount [mm/I-unit]

[0096] VL: Leveling action speed [mm/sec]

[0097] [Formula 3]

[0098] tj=(5×CA)/VA(j=1,...,/)...(4)

[0099] tj=(5×CI)/VI(j=8, 9) …(5)

[0100] tj=(5×CL)/VL(j=10)…(6)

[0101] Here, the reference operation amount of each shape control mechanism is the operation amount of each shape control mechanism whose influence on the shape (state amount) becomes equal, and can be determined by testing the change amount of the shape with respect to the operation amount of each shape control mechanism . The extraction time width tband sets the maximum time of tj, but in fact, considering the variation in the effect of the operation and the interval of multiple operations, as shown in the following formula (7), set several times the time (time multiplied by the coefficient β) .

[0102] [Formula 4]

[0103]

[0104] β = 2 to 3 times

[0105] Here, the reference that can be judged as being able to be improved by operation is set to 5 [I-unit], but it may be appropriately changed according to the actual rolling conditions. In addition, with regard to tband, the value of tband can be selected by methods other than the present example as necessary, such as the operation duration period that is more than fixed.

[0106] The shape control mechanism operation amount Oj(t) is obtained by using the above-mentioned CA, CI, and CL, and the supervisory data output portion Tout(t1) is obtained by the following equation (8).

[0107] [Formula 5]

[0108] Tout(t1)={O1(t1)/CA,...,O7(t1)/CA,O8(t1)/CI,O9(t1)/CI,O10(t1)/CL}...(8)

[0109] By using CA, CI, and CL, the weight of each operation amount can be normalized (normalized) and used.

[0110] The operation result evaluation unit 206 uses the actual performance data reading counter t from the actual performance data reading counter setting unit 201 , the extraction time width tband from the actual performance data noise countermeasure extraction time width setting unit 203 , and the post-operation state quantity change delay time As for the delay time Δt of the setting unit 205 , the plate shape state quantity S(t) is read from the actual operation performance DB 5 .

[0111] Figure 7 The time delay of the plate shape detection of the rolled steel plate is shown. The plate shape of the material to be rolled 4 rolled using the upper work rolls 1 and the lower work rolls 2 is detected by the shape detector 3 located at the position of the travel distance L in the rolling direction. Therefore, the post-operation state quantity change delay time setting unit 205 uses the running speed v(t), which is the traveling speed of the material to be rolled 4, to obtain the rolling speed in the material to be rolled 4 by the following formula (9). The delay time Δt until the shape of the part is detected by the shape detector.

[0112] [Formula 6]

[0113] Δt=L/v(t)...(9)

[0114] The operation result evaluation is obtained from the plate shape state quantity before and after the operation. The plate shape state quantity is evaluated based on the shape deviation spdev(i). For example, the shape evaluation value V(t) at the time t indicated by the actual performance data reading counter t is calculated by the following equation.

[0115] [Formula 7]

[0116] V(t)=∑|spdev(i)|/Ch...(10)

[0117] Figure 8The graph of , represents the temporal change of the shape evaluation value V(t). From the shape evaluation value V(t1), the extraction time width tband, and the delay time Δt, the operation result evaluation value Tv(t1) in the actual performance data reading counter t(t=t1) is obtained by the following equation (11).

[0118] [Formula 8]

[0119] Tv(t1)=(V(t4)-V(t3))/V(t3)/tband*tcyc...(11)

[0120] Among them, t3=t1+Δt, t4=t1+Δt+tband

[0121] The new supervision data generation determination unit 208 generates the shape evaluation value condition THT and the new supervision based on the operation result evaluation value Tv(t) from the operation result evaluation unit 206 and the new supervision data from the new supervision data generation condition setting unit 207 The data generation operation amount condition THO determines the new supervision data generation flag fT by the following formula (12).

[0122] [Formula 9]

[0123] fT=0(Tv(t1)

[0124] MAX{O1(t1)/CA, ..., O7(t1)/CA, O8(t1)/CI, O9(t1)/CI, O10(t1)/CL}

[0125] fT=1(Tv(t1)≥THT∧

[0126] MAX{O1(t1)/CA, ..., O7(t1)/CA, O8(t1)/CI, O9(t1)/CI, O10(t1)/CL}≥THO)...(12)

[0127] When the operation result evaluation value Tv(t1) is smaller than the new supervision data generation shape evaluation value condition THT or the maximum value of the shape control mechanism operation amount Oj(t1) of each shape control mechanism is smaller than THO, the new supervision data generation flag fT= 0. That is, when the shape change of rolling is small or the operation amount of the shape control mechanism is small, the new supervision data Tnew is not generated as the new supervision data generation flag fT=0.

[0128] When the operation result evaluation value Tv(t1) is equal to or more than the new supervision data generation shape evaluation value condition THT and the maximum value of the shape control mechanism operation amount Oj(t1) of each shape control mechanism is THO or more, the new supervision data generation flag is set. fT=1. That is, when the rolling shape change is large and the operation amount of the shape control mechanism is large, the new supervision data Tnew is generated as the new supervision data generation flag fT=.

[0129] In addition, only the operation result evaluation value Tv(t1) may be determined, and the new supervision data generation flag fT may be set. That is, fT=0 (when Tv(t1)

[0130] Here, the new supervision data generation shape evaluation value condition THT is a condition for extracting the condition after the state quantity of the plate shape has been improved, and it can be determined that the shape is significantly improved by operation rather than a temporary change caused by noise or the like. benchmark. Based on experience, it is considered that a change of ±5 [I-unit] for each shape detector is appropriate as the shape evaluation value V(t) about 10/Ch [I-unit]. Taking these circumstances into consideration, the following formula (13) is used to express the new supervision data generation shape evaluation value condition THT.

[0131] [Formula 10]

[0132] THT=10/Ch...(13)

[0133] In addition, the new supervisory data generation operation amount condition THO is set so as to exclude the case where the shape evaluation value is changed due to influences other than the operation although the operation is hardly actually performed. An operation amount equivalent to a change of ±5 [I-unit] per shape detector was used as a criterion for determination.

[0134] [Formula 11]

[0135] THO=5/tband*tcyc...(14)

[0136] As described above, the new supervision data generation shape evaluation value condition THT and the new supervision data generation operation amount condition THO use values ​​of 10/Ch[I-unit], ±5 [I-unit] for each shape detector, etc., but These numerical values ​​may be appropriately changed according to the rolling state.

[0137] The new supervision data generation unit 209 operates based on the new supervision data generation flag fT from the new supervision data generation determination unit 208, the supervision data input unit Tin(t) from the supervision data input unit state quantity reading unit 202, and the supervision data output unit. The supervisory data output unit Tout(t) of the quantity reading unit 204 and the operation result evaluation value Tv(t) from the operation result evaluation unit 206 generate new supervisory data Tnew.

[0138] When the new monitoring data generation flag fT is 0, it is assumed that the new monitoring data Tnew has not been generated, and the actual performance data reading counter setting unit 201 is requested to read the next actual performance data.

[0139] When the new supervision data generation flag fT is 1, as the generation new supervision data Tnew, the new supervision data Thew is created according to the following formula.

[0140] [Formula 12]

[0141] Tnew={Tin(t1), Tout(t1), Tv(t1)}…(15)

[0142] image 3 Schematic functional blocks of the supervisory data database updating means 31 are shown. The supervisory data database updating device 31 constitutes, as main elements, a supervisory data database data reading unit 311, a supervisory data quantity data storage number setting unit 312, a supervisory data database updating method judgment unit 313, a supervisory data updating unit 314, and a supervisory data value calculation unit 313. Section 315 , update supervision data determination section 316 .

[0143] In the supervision data quantity data storage number setting unit 312, the maximum number NMAX of supervision data storages to be stored in the supervision DB2 is set.

[0144] The supervisory data database data reading unit 311 reads the supervisory data number counter N and supervisory data T(n) stored in the supervisory DB2.

[0145] The supervisory data database update method determination unit 313 sets the update method flag flagud by using the supervisory data storage maximum number NMAX from the supervisory data quantity data storage quantity setting unit 312 and the supervisory data quantity counter N from the supervisory data database data reading unit 311 .

[0146] [Formula 13]

[0147] flgud=1 (N

[0148] flgud=2 (N=NMAX)...(17)

[0149] The supervisory data updating unit 314 extracts the supervising data T(1), . The new supervision data Tnew of the device 20 is used to generate the updated supervision data T'(1), .

[0150] When the update method flag flagud is 1, that is, when the number N of supervisory data T stored in the supervisory DB2 is smaller than the maximum number NMAX of supervisory data storage, the following equations (18) to (20) are used to obtain The updated supervision data T'(1),...,T'(N) and the supervision data quantity counter N'.

[0151] [Formula 14]

[0152] T'(n)=T(n) (n=1,...,N)...(18)

[0153] T'(N+1)=Tnew...(19)

[0154] N'=N+1...(20)

[0155] That is, when the update method flag flagud is 1, the supervision data T(1), . New supervision data Tnew is added to the latter supervision data T'(N+1), and stored in the supervision DB2.

[0156] In the case where the update method flag flagud is 2, that is, when the number N of supervision data T stored in the supervision DB2 is equal to the maximum number of supervision data storage NMAX without increasing the number of more supervision data T, The updated supervision data T'(1), . . . , T'(N) and the supervision data number counter are obtained using the following equations (21) to (24) by the update supervision data counter Nud from the update supervision data determination unit 316 N'.

[0157] [Formula 15]

[0158] T(N+1)=Tnew...(21)

[0159] T’(n)=T(n) (n≠Nud)…(22)

[0160] T'(Nud)=Tnew...(23)

[0161] N'=N...(24)

[0162] In this case, first, the supervisory data value calculation unit 315 adds T(N+1)(=Tnew) to the supervisory data T(1), . 21)), according to these supervision data T, the supervision data value Val(1),...,Val(N+1) will be calculated. The supervisory data value calculation unit 315 expands the area of ​​the corresponding input state based on the limited supervisory data T, so that the state quantity of the supervisory data input unit Tin is different from the others as high value (good evaluation), and In a similar case, the value of the supervised data is determined as a poor value (poor evaluation).

[0163] Figure 9 This is a method of schematically showing the operation of the supervisory data value Val using a graph. As shown in the above formula (2), the supervisory data input unit Tin(n) is the total number of shape detectors (that is, the number of detection areas), that is, the value of the Ch dimension, but here is assumed to be two-dimensional (sp(1) for the sake of explanation. ) and sp(2)). Here, the supervision data value Val(m) of the supervision data T(m) is obtained as the minimum value among the distances with other supervision data T using the following equations (25) to (28).

[0164] [Formula 16]

[0165]

[0166]

[0167]

[0168]

[0169]Here, sp(i)n denotes the shape deviation spdev(i) in the i-th region (detection region corresponding to the i-th shape detector) of the n-th supervisory data. din(n, m) represents the distance between the n-th supervision data and the input state quantity of the m-th supervision data. din0(n) represents the distance between the nth supervised data and the optimal state 0 (that is, the origin of the sp(i) coordinate system, which is the target value where the shape deviation spdev(i) is 0). din0(m) represents the distance between the mth supervised data and the best state 0. By dividing the distance between each supervision data by the distance between each supervision data and the optimal state 0, it can be considered that the direction of the operation amount of the supervision data output unit Tout changes finely in the vicinity of the optimal state 0.

[0170] Then, the update supervision data determination unit 316, based on the supervision data T(1), ···, T(N+1) from the supervision data update unit 314 and the supervision data value Val(1) from the supervision data value calculation unit 315, ..., Val(N+1) to decide to update the supervision data counter Nud. The update supervision data counter Nud is a counter for determining the supervision data T to be updated (supervision data T with the lowest value).

[0171] Figure 10 , Figure 11 It is a figure explaining the determination method of the update supervision data counter Nud. like Figure 10 As shown in the shown table, the supervision data T(n) are rearranged in order from the largest according to the supervision data value Val(n). Here, let n(k) be the k-th supervised data number after reordering. In this rearrangement, the decision is made as Figure 11 The groups of the two supervised data with the least value are shown. By comparing the operation result evaluation values ​​Tv of the Nth and N+1th supervisory data obtained in this way, the supervisory data counter Nud to be updated is determined as follows.

[0172] [Formula 17]

[0173] Nud=n(N) (Tv(n(N))

[0174] Nud=n(N+1) (Tv(n(N))≥Tv(n(N+1))…(30)

[0175] The supervision data T'(Nud) determined by the update supervision data counter Nud determined in this way is updated based on the new supervision data Tnew, and the remaining supervision data T'(n) are used as supervision data read from the supervision data update unit 314. T(n) (wherein, n≠Nud) is covered in supervised DB2. The supervising data T(N+1) added at first is used only for rearranging the supervising data value Val, so it is not overwritten in the supervising DB2 and is discarded.

[0176] As described above, the supervisory data generating apparatus of the present embodiment constructs and updates the supervisory DB2.

[0177] As described above, according to the present embodiment, the operation result evaluation value Tv, which is an evaluation value of the result corresponding to the operation performed during the predetermined extraction time width tband from the predetermined start time t1, is calculated based on the actual equipment operation performance data. The operation result evaluation value Tv is used to judge whether or not the new supervision data Tnew can be generated. When it is determined that the supervisory data T can be generated, a period including the supervisory data input part Tin calculated from the plate shape state quantity S(t) at the start time t1 and a predetermined extraction time width tband from the start time t1 is extracted. The supervisory data T of the supervisory data output unit Tout is calculated by operating the state quantity O(t) of the machine, and the extracted supervisory data T is stored in the supervisory DB2. By setting it as such a structure, the AI ​​supervision DB2 can be constructed based on the actual operation performance data of a device. Therefore, the AI ​​applied to real-time equipment control can learn control rules offline using past equipment operation performance data, and can also verify its performance offline. Thereby, the reliability of control is improved, and the application of real-time control using AI can be realized.

[0178] Since the operation result is evaluated by the predetermined extraction time width tband, the supervisory data T can be generated by reducing the influence of noise included in the actual operation performance data of equipment control and the delay of the state detection with respect to the operation. Therefore, high-precision control can be realized from the early stage of AI-based device control by performing AI machine learning such as a neural network using supervisory data T obtained from the accumulated huge amount of operational performance data.

[0179] In addition, the supervisory data database update device 31 stores the supervisory data T in the supervisory DB2 with the predetermined supervisory data storage maximum number NMAX as the upper limit. By setting it as such a structure, the number of objects of supervision data can be kept constant. Therefore, the time required for AI machine learning can be kept constant, and this time dilation can be suppressed.

[0180] In addition, when the number of the supervision data T in the supervision DB2 reaches the maximum number NMAX of supervision data storage, the supervision data database update means 31 compares the supervision data T in the supervision DB2 with the newly extracted by the new supervision data extraction means 20 from the supervision data T in the supervision DB2. In the set of supervision data obtained by combining a new supervision data Tnew, two supervision data with the smallest distance between the input part and other supervision data are determined. Then, the one with the worse operation result evaluation value Tv is removed from the two supervision data, and the remaining supervision data T is stored in the supervision DB2. By adopting such a configuration, it is possible to preferentially remove supervisory data whose distances of the input units are close, that is, supervisory data that are similar to each other. Therefore, since supervised data that are not similar to each other remains, AI can be generated to perform machine learning based on supervised data corresponding to various states, and to perform ideal control in a wide range.

[0181] In addition, the supervision data database update device 31 weights the distances between the input parts of other supervision data so as to give priority to the remaining supervision data T with the smaller distance from the target value of the supervision data input part Tin, and determines the weighting After the two supervised data T with the smallest distance. With such a configuration, fine control with high accuracy is required in the vicinity of the target value (that is, the optimum operating state) of the supervisory data input unit Tin. As a result, by the distance weighting of the supervisory data input unit Tin, the Surveillance data is preferentially left in the vicinity of the target value, whereby the control accuracy in the vicinity of the target value of the AI ​​can be improved.

[0182] The extraction time width tband is determined based on the operation time tj calculated using the reference operation quantities CA, CI, CL set for each shape control mechanism and the operation speeds VA, VI, and VL of the shape control mechanism. With this configuration, the average time required for the shape-improving operation can be set to the extraction time width tband, so that the influence of noise and time delay on the operation can be effectively reduced.

[0183] The facility is a rolling mill, the state quantity of the facility operation performance data is a plate shape state quantity detected by a shape detector provided in the rolling mill, and the operation amount is an operation amount of a shape control mechanism provided in the rolling mill. By setting it as such a structure, it becomes possible to generate|occur|produce the supervision data T used for AI machine learning which controls a rolling mill which is a facility.

[0184] In addition, the new supervisory data extraction device 20 generates supervisory data T using data obtained when the operation speed v of the rolling mill is equal to or higher than the predetermined actual performance data reading speed condition THv among the plant operation performance data. By setting it as such a structure, it becomes possible to generate|occur|produce appropriate supervision data using the actual operation performance data at the time of operation of a rolling mill.

[0185] In addition, the new supervisory data extracting means 20 generates a new supervisory data generation operation amount condition THT or more in which the operation result evaluation value Tv is a predetermined evaluation threshold value and the shape evaluation value condition THT or more and the shape control mechanism operation amount Oj is a predetermined operation threshold value or more. , it is determined that the supervision data T can be generated. With such a configuration, the monitoring data can be generated using the actual operation data when the shape change due to rolling is large and the operation amount of the shape control mechanism is large, so that the influence of noise can be reduced and the operation of the equipment can be appropriately reflected. Supervision data for state changes caused. In addition, the new supervision data extraction device 20 may determine that the supervision data T can be generated when the operation result evaluation value Tv is equal to or greater than the new supervision data generation shape evaluation value condition THT of a predetermined evaluation threshold value. With such a configuration, it is possible to more easily determine whether or not the supervisory data T can be generated.

[0186] In addition, this invention is not limited to the said embodiment, In the implementation stage, in the range which does not deviate from the summary, a component can be deformed and it can be demonstrated concretely. In addition, various inventions can be formed by appropriate combinations of a plurality of constituent elements disclosed in the above-described embodiments. For example, some components may be deleted from all the components shown in the embodiment. In addition, components related to different embodiments may be appropriately combined.