JP2005152937A - Apparatus and method for generating table value for correcting prediction error in numerical formula model - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a generating apparatus of a table value for correcting a prediction error in a numerical formula model by which the appropriate table value is generated even when actual values are rare as making the reduction of the number of processes in table maintenance possible in the case the correction of the prediction error in a numerical formula model by using a table system and a method therefor. <P>SOLUTION: The generating apparatus 8 of the table value for correction controls the whole of this system 10. Furthermore, an error correcting device of a control model rearranges the actual data stored in a data base 5 so as to monotonically increase according to a stratified index and supplements corrected values ( for example, the corrected value of the coefficient of contraction of an outside diameter ) in a correction table 7 on the basis of the rearranged actual data. A process computer 6 in which the control model for predicting the coefficient of contraction of the outside diameter is incorporated controls a finishing mill 1. A past actual operation data is stored in the data base 5. A steel pipe 2a is rolled with the finishing mill 1 by controlling with the process computer 6. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a technique for correcting an error between a predicted value and an actual value in a mathematical model, and more particularly to a table value generation apparatus and method for a prediction error correction table used in a mathematical model for process control in the manufacturing industry.

In the manufacturing industry, there is a certain limit to the prediction accuracy of the mathematical model used for controlling the manufacturing process. This is because disturbances exceeding the assumptions in the mathematical model occur, such as changes over time in the process to be controlled, sudden changes in operating conditions, and operator intervention. The problem of the prediction error of the mathematical model due to these unexpected disturbances is dealt with by a method of correcting the prediction error outside the mathematical model.
For example, a method of learning a prediction error of a mathematical model using a neural network is disclosed (see Patent Document 1).

  In addition, the prediction error of the mathematical model is corrected by a table that is divided into layers (“layer” means a data group in one division when the entire predetermined data is divided by a factor that affects the prediction value). The method of doing is also often used. However, in such a method, when the layers are subdivided, there is a problem that the actual number corresponding to each layer becomes scarce. In order to solve this problem, a method has been proposed in which learning is performed in consideration of the results of adjacent layers according to the distance between layers, and a correction value is generated (see Patent Document 2).

FIG. 8 is a diagram for explaining a learning control method when rolling a material specified by the conditions of sheet thickness layer 3 and sheet width layer 3 in the prior art using the table of Patent Document 2 described above. is there. FIG. 8B is an example of the older prior art, in which only the learning coefficient of the material specified by the conditions of plate thickness layer 3 and plate width layer 3 is corrected. FIG. 8C shows the learning coefficient of the material specified by the conditions of the plate thickness layer 3 and the plate width layer 3, depending on the distance, the conditions for the other plate thickness layers and the plate width layers. This is an example in which learning is performed in consideration of the learning coefficient of the material specified by
Japanese Patent Laid-Open No. 11-707 JP-A-6-259107

However, in the case of the technique disclosed in Patent Document 1, since the learning content by the neural network is a so-called black box, there is a problem that the person who maintains the mathematical model cannot grasp the progress of learning about the error. In addition, the neural network method has a problem that it is difficult to put it into practical use from the viewpoints of learning stability and extrapolation accuracy.
The prior art disclosed in Patent Document 2 is a technique for solving the problems of the method using the table method. This technology is based on the assumption that there is a correlation between the distance of each layer and the prediction error performance corresponding to each layer, but in reality there is no clear correlation between the two There is a problem that an appropriate correction value cannot be generated.

  Therefore, the present invention has been made in view of the above problems, and when the prediction error in the numerical model is corrected using the table method, the actual value is rare while reducing the man-hours for table maintenance. It is an object of the present invention to provide a prediction error correction table value generation apparatus and method for a mathematical model that can generate an appropriate table value.

The reason why the present invention adopts the table method that has been used well in the past is that this method is still the most familiar technique in the field of manufacturing industry, and the person who maintains the mathematical formula model has its prediction error. This is because it is easy to grasp the transition of learning.
In order to build a table, the type of layer (hereinafter referred to as “index”) is set. Taking the rolling process of steel pipes in the steel industry as an example, the outer diameter, thickness, material, etc. are candidates for the index. A person who maintains the mathematical model sets an index by empirically determining what causes the prediction error of the mathematical model. For each index, the number of layer divisions is determined in consideration of the distribution characteristics of the prediction error results. Hereinafter, an index representing each layer is referred to as a “stratified index”.

For example, if it is determined that the prediction error performance has changed greatly according to the outer diameter, the “outer diameter” index has a layer structure of 100 divisions, and the prediction error of the mathematical model does not change much according to the wall thickness. If it is determined, the “thickness” index has a five-layer structure.
Increasing the number of layer divisions in the table method can cope with individual manufacturing conditions, but on the other hand, depending on the layer, there may be a problem that the number of actual data corresponding to it may decrease, so correction accuracy is generally Cannot be said to improve.

  In view of this, a method has been proposed that utilizes not only the prediction error record corresponding to the corresponding layer but also the prediction error record corresponding to the adjacent layer, as in the prior art of Patent Document 2 above. The method of learning by a neural network is also a kind of method of utilizing the prediction error results corresponding to the adjacent layers.

  However, characteristics (for example, average values) of prediction error results corresponding to adjacent layers are not necessarily similar to each other. For example, the tendency of prediction error results corresponding to the adjacent layers “material 1” and “material 2” in the “material” index may be completely different. That is, the prediction error results corresponding to the adjacent layers cannot always be effectively used. However, it is not practical to further increase the table index by further specifying the influence factors of the prediction error because the table is difficult to maintain.

In the present invention, focusing on the setting method of the index for each index layer, rearranging each index layer as a pre-processing for generating a table value utilizes the prediction error results corresponding to other layers. It was found to be effective.
Here, the steel pipe rolling process in the steel industry will be described again as an example. In this process, in order to increase the accuracy of the product outer diameter, it is important to predict the outer diameter shrinkage ratio after finish rolling with high accuracy. In the following, description will be made on the generation of the table value of the prediction error correction table of the mathematical model for predicting the outer diameter shrinkage rate.
A person who maintains this mathematical model creates a prediction error correction table for the mathematical model that predicts the outer diameter shrinkage rate. Based on previous experience, the accuracy of the mathematical model is judged to be affected by the outer diameter, wall thickness, material, etc., and the correction table for this mathematical model includes the outer diameter, wall thickness, material, etc. Provide multiple indexes. Then, each index is divided into layers, and a stratified index is given. The prediction error results collected in the past operation are the positions corresponding to the combinations of the layers of each index (hereinafter referred to as “in-table positions”). Make it correspond. For the positions in the table where there are a considerable number of corresponding prediction error results, the table value at that position can be generated using the average value thereof. However, there is a problem of how to accurately generate table values at positions in the table where the number of corresponding prediction error results is small or nonexistent (position in the table where the record data is scarce).

  In the present invention, first, prediction error results are made to correspond to positions in the table where the record data is scarce, the layers are rearranged in an order in which the average value of the prediction error results simply increases, and a new stratification index is assigned. I thought. By this processing, each index layer that specifies the position in the table where the record data is scarce is represented by a newly assigned layer index. In order to perform such layer rearrangement processing and newly reassign the stratification index, the position in the table based on the new stratification index display is similar to the mutual distance and the average value of the corresponding prediction error results. There must be a relationship between sex. Therefore, if a position in the table is selected based on the distance between the positions in the table based on the new stratified index display, and a prediction error record corresponding to such a position in the table is used, an accurate table value is generated. be able to.

  As a specific method for selecting a position in the table, a method of evaluating with a norm is adopted. Select a position in the table where the norm satisfies a certain standard, find the average value of the entire prediction error results corresponding to the selected position in the table and the position in the table where the actual data is scarce, and calculate the table value. Generate.

  In order to achieve the above-mentioned object, the prediction error correction table value generation device in the mathematical model according to the present invention includes a layered index assigned to each index layer of the correction table and a prediction error result corresponding to each layer. A new layer is created by rearranging the layers so that the relationship between the average value of the prediction error results corresponding to each layer and the layer is monotonically increased by using the acquired data of the set and the prediction error result acquisition means for acquiring the data of the set. Assign a stratified index, calculate the norm of the position in the table based on the display of the newly assigned stratified index and the position in the table of the table value generation target. A table using table position selection means for selecting a table position to be satisfied, and a prediction error record corresponding to the selected table position and table value generation target table position. And a table value calculating means for calculating.

  Thereby, even if the performance data corresponding to the position in the table of the table value generation target is scarce, the table value can be determined appropriately, and finally the mathematical model can be corrected with high accuracy.

Furthermore, in order to achieve the above object, a process control system according to the present invention provides:
A prediction error correction table value generation device for a predetermined mathematical model, a database device and a process computer connected via a network, wherein the table value generation device is a layer assigned to each index layer of the correction table The relationship between the prediction error record acquisition means for acquiring data of a set of another index and the prediction error record corresponding to each layer, and the average value of the prediction error record corresponding to each layer using the acquired set of data is monotonous. Rearrange the layers to increase and assign a new stratification index, and calculate the norm of the table position to be generated for the table position based on the display of the newly assigned stratification index In-table position selection means for selecting a position in the table in which the calculated norm satisfies a certain criterion, the selected position in the table, and Table value calculating means for calculating a table value using a prediction error record corresponding to a table value generation target table value, and the database device further comprises storage means for storing the prediction error record, The process computer includes process control means for controlling the process to be controlled using the table value received from the table value generation device.

  As a result, the correction table value generation apparatus according to the present invention is connected to the process computer and the database via the network, so that it is not necessary to own past performance data required by the table value generation apparatus. In addition, remote control from the process computer becomes possible.

  In order to achieve the above object, the present invention can be realized as a table value generation method having the characteristic constituent means of the above apparatus as a step, or as a program including all steps of those methods. it can. The program is not only stored in a ROM or the like provided in an apparatus capable of realizing the above method, but can also be distributed via a recording medium such as a CD-ROM or a transmission medium such as a communication network.

  According to the present invention, with respect to a table for correcting a prediction error of a mathematical model used for control of a manufacturing process, a table value can be determined with high accuracy even at a position in the table where there is little or no corresponding actual data. . By using this table, the prediction error of the mathematical model can be corrected appropriately, and an optimum process control system can be realized.

Hereinafter, embodiments according to the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram showing a functional configuration of a process control system 10 according to the present invention. As shown in FIG. 1, the process control system 10 includes a correction table value generation device 8, a database 5, a process computer 6, a finish rolling mill 1, an outer diameter meter 3, and an ultrasonic flaw detector 4.

  The process computer 6 is a device that controls the entire system 10. The correction table value generation device 8 generates a table value (for example, an outer diameter shrinkage rate correction table value) in the correction table 7 based on the result data stored in the database 5 according to an instruction from the process computer 6. . The process computer 6 incorporates a mathematical model for predicting the outer diameter shrinkage rate, receives the table value from the correction table value generation device 8, corrects the mathematical model, and controls the finishing mill 1. The database 5 stores past operation result data. The finish rolling mill 1 rolls the steel pipe 2 a based on the control of the process computer 6.

The outer diameter meter 3 is arranged on the output side of the finish rolling mill 1 and measures the outer diameter of the rolled steel pipe 2a. The ultrasonic flaw detector 4 measures the outer diameter of the steel pipe 2b after cooling.
The process computer 6 and the database 5 in FIG. 1 may be connected to the correction table value generation device 8 via a network or the like.
In the present embodiment, an example in which a mathematical model for predicting the outer diameter shrinkage rate is applied to pipe rolling will be described. An outer diameter meter 3 capable of measuring the outer diameter of the steel pipe 2 after hot rolling is disposed on the output side of the finish rolling mill 1 in the pipe rolling process. The finish rolling mill 1 is one that can manufacture a pipe having an arbitrary outer diameter by changing a roll, changing a gap, or the like.

From the completion of rolling in the finish rolling mill 1 to the completion of measurement in the ultrasonic flaw detector 4, it takes time for material conveyance, heat treatment, air cooling, etc., and there is a time lag of several hours. This is not reflected in the correction of the outer diameter shrinkage rate prediction in the same operation, and is used after the next operation.
Here, the process control system 10 shall roll about 1000 steel pipes having a finished outer diameter Φ200 to 300 mm × wall thickness t20 to 40 mm. As for the pipe production result data in the same production condition range, data on about 5000 pieces are accumulated in the database 5 in the past half year. The correction table value based on this invention is produced | generated using the pipe production performance data, and shall be used for correction | amendment of the prediction error of a numerical formula model at the time of this pipe production. That is, a product obtained by adding the prediction error correction value ΔαT generated by the method of the present invention to the outer diameter shrinkage rate α1 calculated by the theoretical model is the outer diameter shrinkage rate predicted value α2, and the product target outer diameter Dct and its outer diameter are obtained. A target outer diameter Dht at the time of hot rolling is obtained from the shrinkage rate prediction value α2.

α2 = α1 + ΔαT (1)
α1: Theoretical model predicted outer diameter shrinkage [%]
α2: Corrected model predicted outer diameter shrinkage [%]
ΔαT: Outer diameter shrinkage correction value [%]
Dht = (1 + 0.01 × α2) × Dct (2)
Dht: Target outer diameter during hot rolling [mm]
Dct: Product target outer diameter [mm]
The correction table 7 is provided with “outer diameter” and “material”, which are considered to have a strong correlation with the outer diameter shrinkage rate, based on experience so far, as indexes.

  It is assumed that the correction table value generation device 8 starts generating the table value of the correction table 7 in accordance with an instruction from the process computer 6 and obtains the table value of the position in the table under the condition of outer diameter Φ250 mm and high carbon steel. The position in the table of this condition is the index “17” for the “outer diameter” index, and the index “9” for the “material” index (hereinafter, the position in the table is displayed in a matrix (the “outer diameter” index layer, (Layer of “material” index), and (corresponding to (17, 9)). There are nine outer diameter shrinkage rate prediction error results corresponding to the positions in the table. As a precondition, it is assumed that there is a rule in the correction table value generation device 8 that “300 or more performance data is required to calculate a table value”. Then, for the in-table positions (17, 9), the table value cannot be generated only by the corresponding outer diameter shrinkage rate prediction error results, so the outer diameter shrinkage rate prediction error results at other table positions are utilized. I will do it.

First, the correction table value generation device 8 investigates the characteristics of each index layer with respect to the data ΔαE of the outer diameter shrinkage rate prediction error results accumulated in the database 5 (see FIGS. 2A and 2B). ).
ΔαE = αJ−α1 (3)
αJ = (Dhj−Dcj) / Dcj × 100 (4)
ΔαE: Outer diameter shrinkage prediction error [%]
αJ: Outer diameter shrinkage rate actual [%]
Dhj: Hot outer diameter record [mm]
Dcj: Actual cold outer diameter [mm]
The neighboring layers of the “outer diameter” index stratification index “17” are stratification indices “15” and “18” (“16” has no past performance data). The neighboring layers of the layer index “9” of the “material” index are the layer indexes “7” and “8”. It can be seen that there is no clear correlation between the characteristics of the outer diameter shrinkage rate prediction error performance for the layer of the “outer diameter” or “material” index. Therefore, it is difficult to generate a highly accurate table value even if actual data corresponding to adjacent layers is utilized as it is. Therefore, the layers are rearranged so that the average value of the outer diameter shrinkage rate prediction error for each layer is monotonously increased, and a layer-specific index is newly added (see FIGS. 3A and 3B).

Then, each index layer at the position (17, 9) in the table is replaced with a new layer index as follows.
"Outer diameter" index index "17" → "6"
“9” → “6” by layer index of “Material” index
The neighboring layers after being sorted are as follows.

Layers adjacent to the “Outside Diameter” index stratification index “6”
Tier index “5” (original “outer diameter” index is “12”)
Tier index “7” (original “outer diameter” index is “18”)
Layers in the vicinity of the “Material” index stratification index “6”
Stratification index “5” (original “materials” index is “3”)
Stratification index “7” (original “materials” index is “4”)
Based on the “new stratification index” changed in this way, a close table position is selected, and a table value is calculated using actual data corresponding to the position.

  As a method for selecting adjacent table positions, a method (6, 6) in which the table positions (17, 9) are expressed using a new stratified index and a norm of other table positions is obtained and determined. adopt.

ｄ() ：テーブル内位置の近接度を表す距離関数
Ｘ ：近接度判定対象のテーブル内位置
ｘ ：テーブル値生成対象のテーブル内位置
Ｍ ：重み
Ｘ、ｘともに、行列形式で表わされている。重みＭについては、たとえば、対角行列として設定することで、各インデックスの層に任意に重み付けが可能となる。本実施例では、重みＭを単位行列とし、各層に均等に重み付けしながらテーブル内位置間、近接度を判定することとなる。

d (): Distance function representing the proximity of the position in the table
X: Table position for proximity determination
x: Position in the table for generating table values
M: Weights X and x are both expressed in a matrix format. For example, by setting the weight M as a diagonal matrix, it is possible to arbitrarily weight each index layer. In this embodiment, the weight M is a unit matrix, and the proximity between the positions in the table is determined while equally weighting each layer.

As a precondition, it is determined that the in-table position X where the norm with x (6, 6) is 1.0 or less has high proximity to x (6, 6). According to this criterion, four positions in the table of X (7, 6), X (5, 6), X (6, 5), X (6, 7) are selected.
The distance function between x (6,6) and X (7,6) is

ｘ（６、６）とＸ（５，６）との距離関数は、

The distance function between x (6,6) and X (5,6) is

ｘ（６、６）とＸ（６，５）との距離関数は、

The distance function between x (6,6) and X (6,5) is

ｘ（６、６）とＸ（６，７）との距離関数は、

The distance function between x (6,6) and X (6,7) is

この４ヶ所のテーブル内位置に対応する予測誤差実績数は３８２点となり、必須データ数３００点という前提条件も満たしている（図９参照）。

The number of prediction error results corresponding to the four positions in the table is 382 points, and the precondition that the number of essential data is 300 points is also satisfied (see FIG. 9).

  Finally, the correction table value generation device 8 obtains an average value of all the actual data specified by the positions in the four tables and the actual table data where the actual data is scarce, and expresses it in the original stratified index display. Is recorded as a table value at the table position (17, 9). The average value of nine performance data corresponding to the in-table positions (17, 9) was -0.04, but the prediction error results specified by the above four in-table positions were used. The average value was found to be -0.03.

  By the way, after the rearrangement of the layers of each index described above, it is one of the more effective methods to normalize and display the layer index notation by “0 to 1”. Further, it is a more effective method to change the stratified index so that the average value of the outer diameter shrinkage rate prediction error results of each layer is not simply increased due to the relationship with the layer (linear processing). FIG. 4 shows an example of a result obtained by normalizing and displaying the stratified index of each index, and further changing the stratified index so that the average value of the prediction error performance of each tier is linear in relation to the tier.

Hereinafter, normalization and linear processing will be specifically described.
FIG. 4 shows an outer diameter shrinkage ratio prediction error using a layered index βj (j = 1 to k, k: number of layers) normalized as a new layered index in the range of (0 to 1). FIG.
First,
Layer with the smallest average prediction error → Stratified index β1: 0.0
Layer with the largest average prediction error → Stratified index βk: 1.0
And set. For the remaining layers, a layer-by-layer index βj with an error average satisfying the following formula in a range of 0 to 1 is given.

Average prediction error = βj × (Δαmax−Δαmin) + Δαmin (8)
βj: Index for each layer (0.0 <βj <1.0)
Δαmax: layer prediction index = average prediction error performance average of 1.0 layer Δαmin: layer prediction index = 0.0 layer prediction error performance average βj After calculating βj, the above layer rearrangement index after layer rearrangement By replacing with another index βj, a similar table value calculation process can be performed. More specifically, in the above-described case, the outer-diameter index βj having an outer diameter of Φ250 mm is “0.32” as shown in FIG. The stratified index of the neighboring layer is as follows.

The stratification index β of the neighboring layer of the stratification index “0.32” normalized by the “outer diameter” index is
“0.30” (Original outer layer index is “12”)
“0.33” (Original outer diameter index is “18”)
Since this new stratification index that has been normalized and linearized has a linear relationship with the average value of the corresponding actual data, in the above example, the difference in stratification index between neighboring layers (“0.02”) And “0.01”) and the corresponding difference ratio of the actual data average value correspond to each other in the other optimal table in the proximity determination by the above-described norm equation (ie, equation (5)). The position can be selected.

The flow of processing until the table value in the correction table is generated is shown in FIGS.
The correction table value generation device 8 refers to the database 5 and reads past result data (S501), and extracts a prediction error result for each position in the table (S502). Then, the actual number of prediction errors corresponding to each in-table position is determined, and the in-table position that satisfies the prescribed N number is separated from the in-table position that does not satisfy it (S503). Subsequently, the table value generation operation is started, and for the in-table position that satisfies the reference N number, the average value of the corresponding prediction error results is calculated as a table value (S504), and the in-table position that does not satisfy the reference N number For, a table value is generated utilizing prediction error results corresponding to other positions in the table (S505). The table value generation processing for the in-table position that satisfies the specified N number and the in-table position that does not satisfy the specified N number may be performed sequentially or in parallel.

  FIG. 6 is a flowchart showing the flow of the table value generation process (S505) for the position in the table where the actual data that does not satisfy the reference N number is scarce. The correction table value generation device 8 performs layer rearrangement processing, layer-by-layer index replacement processing, layer-by-row index normalization display processing, or linearization according to preset conditions (or condition settings from an operator). Processing is performed (S603 to S605).

  Furthermore, the correction table value generation device 8 selects a table position with a high degree of proximity from the table position based on the newly assigned layer indicator display (S606, S607), and the selected table position And a table value is calculated based on the prediction error track record corresponding to the position in a table with scarce track record data (S608).

  As described above, the outer diameter shrinkage rate prediction accuracy when the pipe is manufactured by applying the outer diameter shrinkage rate correction value calculated based on the method according to the present invention, and the outer diameter shrinkage rate prediction accuracy by other methods. Comparison was made using the histogram of FIG.

  FIG. 7A shows the accuracy of the mathematical expression prediction model itself. FIG. 7B is a histogram showing the results when the layer rearrangement process is not performed. FIG. 7C is a histogram showing the results when the layers are rearranged so that the prediction error actual average value for each layer increases monotonously. FIG. 7D is a histogram showing the results when the normalized display processing and linearization processing of the stratified index are performed. As is apparent from these histograms, a significant accuracy improvement effect was obtained by applying the table value generated based on the method according to the present invention as compared with the conventional technique (see Table 1).

  Moreover, since the correction value is automatically generated in the correction table value generation device 8 connected via a network to a process computer, a long-term storage database, etc., it is possible to reduce maintenance man-hours, which is a conventional problem. It becomes. In addition, by executing the rearrangement of layers, it is possible to appropriately interpolate correction values for layers in which past performance data is scarce, so that the final result can be obtained regardless of the number of past performance data or similar case data. The accuracy of the product can be improved.

  As described above, the table value generation apparatus and method of the prediction error correction table of the mathematical model according to the present invention can be applied to a stand-alone type or network type control system. Suitable for controlling error values and the like.

It is a block diagram which shows the function structure of the process control system containing the prediction error correction table which concerns on this invention. (A) is a figure which shows the relationship between the layer of an "outer diameter" index, and an outer diameter shrinkage | contraction rate prediction error. (B) is a figure which shows the relationship between the layer of a "material" index, and an outer diameter shrinkage | contraction rate prediction error. (A) is the figure which rearranged the layer so that the average value of the outer diameter shrinkage | contraction rate prediction error which concerns on each layer of an "outer diameter" index may increase monotonously. (B) is the figure which rearranged the layer so that the outer diameter shrinkage | contraction rate prediction error average value regarding each layer of a "material" index may increase monotonously. (A) is a figure at the time of performing normalization and a linearization process with respect to the outer diameter shrinkage | contraction rate prediction error average value which concerns on each layer of an "outer diameter" index. (B) is a figure at the time of performing normalization and a linearization process with respect to the outer diameter shrinkage | contraction rate prediction error average value which concerns on each layer of a "material" index. It is a flowchart which shows the flow of a process until it calculates a correction table value. It is the flowchart which showed the flow of the process of table value generation | occurrence | production of the position in a table with few prediction error results which concern on this Embodiment. (A) is the histogram which showed the outer diameter shrinkage | contraction rate prediction error in the conventional theoretical model. (B) is a histogram of the outer diameter shrinkage rate prediction error when the rearrangement process is not performed according to the stratified index. (C) is a histogram of the outer diameter shrinkage rate prediction error when rearrangement is performed according to the stratified index so that the average value of the outer diameter shrinkage rate prediction error, which is a model error characteristic, monotonically increases. (D) is a histogram of the outer diameter shrinkage rate prediction error when expressed according to the normalized layered index so that the relationship between the layer index and the average value of the outer diameter shrinkage rate prediction error is linear. (A) is a figure for demonstrating the learning control method at the time of rolling the material which belongs to the group of the thickness layer classification 3 and the sheet width classification 3 in a prior art with five roll stands. (B) is an example in which only the learning coefficients of the plate thickness layer 3 and the plate width layer 3 are corrected. (C) is an example in which the value is changed in accordance with the proximity of the distance between the sheet thickness layers 3 and the sheet width layers 3.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Finish rolling mill 2a, 2b Steel pipe 3 Outer diameter meter 4 Ultrasonic flaw detector 5 Database 6 Process computer 7 Correction table 8 Correction table value generator 10 Process control system

Claims (10)

A prediction error record acquisition means for acquiring data of a set of a stratified index assigned to each index layer of the correction table and a prediction error record corresponding to each layer;
In-table position selection means for selecting a position in the table with high similarity of the corresponding prediction error actual average value based on the acquired data;
Table value calculating means for calculating the table value of the table position of the table value generation target using the prediction error record corresponding to the selected position in the table and the table position of the table value generation target. Table value generation device for correcting prediction error in model. The prediction error record acquisition means includes
For the acquired set of data, a layer rearrangement processing unit that rearranges the layers so that the relationship between the layer and the prediction error actual value corresponding to each layer increases monotonously,
The table value generation apparatus for prediction error correction in the mathematical model according to claim 1, further comprising: a layer-by-layer index reassigning processing unit that gives a new layer-by-layer index to the rearranged layers. The layer-by-layer index reassignment processing unit further has a range of 0 to 1 for the newly assigned layer-by-layer index, and the relationship between the layer and the average value of the prediction error performance corresponding to each layer is linear. A new stratification index is assigned again to the prediction error correction table value generation device in the mathematical model according to claim 2. The in-table position selecting means includes
For a position in the table based on the display of the newly assigned stratified index, a norm calculation unit that calculates a norm with the position in the table of the table value generation target, and the table value generation target based on the calculated norm The table value generation apparatus for correcting a prediction error in a mathematical model according to claim 2, further comprising: an in-table position determining unit that determines another in-table position close to the in-table position. A process control system comprising a prediction error correction table value generation device in a predetermined mathematical model, a database device and a process computer connected via a network,
The table value generation device includes:
A prediction error record acquisition means for acquiring data of a set of a stratified index assigned to each index layer of the correction table and a prediction error record corresponding to each layer;
In-table position selecting means for selecting another in-table position having a high degree of proximity to the in-table position of the table value generation target;
Table value calculation means for calculating a table value of the table value generation target table position using the prediction error results corresponding to the other selected table position and table value generation target table position;
The database device includes:
Storage means for storing the prediction error results,
The process computer is
A process control system comprising process control means for controlling the process to be controlled using a table value received from the table value generation device. The prediction error result acquisition means of the table value generation device is:
For the acquired set of data, a layer rearrangement processing unit that rearranges the layers so that the relationship between the average value of the prediction error performance corresponding to each layer and the average increases monotonically,
The process control system according to claim 5, further comprising: a stratified index transfer processing unit that assigns a new stratified index to the rearranged layers. The layer-by-layer index reassignment processing unit further has a range of 0 to 1 for the newly assigned layer-by-layer index, and the relationship between the layer and the average value of the prediction error performance corresponding to each layer is linear. The process control system according to claim 6, wherein a new stratification index is assigned again. The in-table position selecting means includes
For a position in the table based on the display of the newly assigned stratified index, a norm calculation unit that calculates a norm with the position in the table of the table value generation target, and the table value generation target based on the calculated norm The process control system according to claim 6, further comprising: an in-table position determining unit that determines another in-table position close to the in-table position. A step of acquiring a prediction error result to acquire data of a set of a stratified index assigned to each index layer of the correction table and a prediction error result corresponding to each layer;
Selecting a position in the table for selecting another position in the table having a high degree of proximity to the position in the table of the table value generation target;
A table value calculating step of calculating a table value of the table position of the table value generation target using the prediction error results corresponding to the other selected table position and the table position of the table value generation target. A table value generation method for correcting a prediction error in a mathematical model. A step of acquiring a prediction error result to acquire data of a set of a stratified index assigned to each index layer of the correction table and a prediction error result corresponding to each layer;
Selecting a position in the table for selecting another position in the table having a high degree of proximity to the position in the table of the table value generation target;
A table value calculating step of calculating a table value of the table position of the table value generation target using the prediction error results corresponding to the other selected table position and the table position of the table value generation target. A program for a table value generation device for correcting a prediction error in a mathematical model.

Images