JP3037832B2 - Automatic shape control device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a shape control apparatus for flattening a sheet surface in a sheet width direction in rolling of a thin metal sheet.
In particular, the present invention relates to a shape automatic control device provided with a mechanism for storing examples of skilled operator shape pattern determination with variations, so that extremely high-performance shape control can be automatically performed.

[0002]

2. Description of the Related Art Generally, shape control in the rolling of a thin metal sheet is control for flattening the sheet surface in the sheet width direction.

In this type of shape control, there are a plurality of actuators such as an actuator for controlling a symmetrical undulation such as a bending pressure and an actuator for controlling a symmetrical undulation such as a leveling pressure.

Conventionally, a plate-shaped waveform obtained from a signal from a shape meter is decomposed into components that can be controlled by the respective actuators as described above, and the operation amounts of a plurality of actuators are reduced according to the respective components. Determined and controlled.

[0005] However, in this case, (a) a plurality of actuators interfere with each other, and (b) the physical properties of the plate are easily changed by temperature and external pressure. The model differs from the actual operation. Therefore, the control output by the operation of a skilled operator often provides better control results.

In recent years, attempts have been made to store the judgment of a skilled operator in a machine, and to select a control actuator and to determine a control direction and an operation amount based on the judgment result. I have. Usually, the judgment of a skilled operator is based on "experience" and "intuition", and in many cases, the mechanism of the judgment is not clarified. Therefore, the most important point is how to store the judgment of the skilled operator in the machine.

[0007] It is said that the neural network will mimic the determination method by presenting a sufficient number of examples of the determination even if it is difficult to elucidate the mechanism. Therefore, a sufficient number of examples of pairs of input data from the sensor and the operation pattern of the operator are collected,
By teaching this to the neural network,
It becomes possible to store the operation of the operator. By the way, conventionally, the controllable component of the actuator is obtained from the waveform of the plate obtained from the shape sensor, and the control is performed based on the component amount.

On the other hand, the operator looks at the contour (shape) of the surface of the metal plate and classifies it into the most similar form of some typical patterns. Then, the shape is controlled by executing a standard operation procedure for the pattern.

In this case, the operating procedure includes the order in which the actuators to be used are used, the usage of a unique actuator peculiar to each pattern, etc., and the control according to the control target and the state at each time is utilized. Have been. These enable delicate control that is difficult to grasp in dimensions such as decomposition and synthesis of controllable components.

Therefore, the output data from the shape meter and the actual example of the judgment of the operator recognition pattern at that time are taught to the neural network. After teaching a sufficient number of judgment examples, the neural network starts to make a judgment of the operator, and thereafter, the pattern judgment is performed by the neural network, and a standard operation procedure is performed based on the judgment result. By implementing,
Control similar to that of an operator can be realized.

Therefore, according to the present method, it is not possible to cope with the conventional operation only by combining operations on controllable components.
It is possible to perform control that matches the unique control target. However, when trying to store the operation of the operator in the machine, there are still the following problems.

[0012] The judgment of the operator (human) varies, and sometimes the judgment differs from person to person. Therefore, even if an operation example or a judgment example of the operator is input to the machine, a contradiction occurs between the input data. As a result, even if the information is stored in the machine, the accuracy does not increase or the machine cannot store the information.

When the shape of the plate is similar to a typical pattern, the judgment of the operator is always the same, but when the shape is intermediate between a plurality of typical patterns, for example.
The judgment of the operator sometimes varies from person to person.

Therefore, even if an attempt is made to teach to a neural network, the data groups to be taught are inconsistent, causing problems such as (a) learning of the neural network does not converge, and (b) determination of the network is unstable.

[0015]

As described above, in the conventional shape control in the rolling of a thin metal plate, there is a problem that high-performance shape control cannot be performed.

An object of the present invention is to provide a mechanism for storing a skilled operator shape pattern judgment example having variation,
An object of the present invention is to provide a highly reliable automatic shape control device capable of automatically performing extremely high-performance shape control.

[0017]

In order to achieve the above object, in a shape control device for flattening a sheet surface in a sheet width direction in rolling of a thin metal sheet,

First, according to the first aspect of the present invention, as shown in FIG. 1, a sensor 1 for capturing the state of the surface shape of a plate.
Teaching data input means 2 for inputting as teaching data a pair of data input from the sensor 1 and an example of a skilled person's shape pattern determination for the input data, and a plurality of teaching data from the teaching data input means 2 , A teaching data selecting means 3 for obtaining a variation of a judgment value of a skilled person with respect to the input data, and selecting and outputting teaching data so as to reduce the variation, and a teaching output from the teaching data selecting means 3. A circuit construction means 4 for constructing a shape decision circuit for storing an expert's shape recognition decision using data, and a built-in shape decision circuit constructed by the circuit construction means 4, and data inputted from the sensor 1. A shape pattern determining unit 5 for determining which one of the predetermined patterns corresponds to the shape pattern determining unit 5; Based on the pattern determination result by, constituting an output means 7 for outputting seeking control output.

On the other hand, according to the second aspect of the present invention, as shown in FIG. 1, the sensor 1 captures the state of the surface shape of the plate.
Teaching data input means 2 for inputting as teaching data a pair of data input from the sensor 1 and an example of a skilled person's shape pattern determination for the input data, and a plurality of teaching data from the teaching data input means 2 , A teaching data selecting means 3 for obtaining a variation of a judgment value of a skilled person with respect to the input data, and selecting and outputting teaching data so as to reduce the variation, and a teaching output from the teaching data selecting means 3. A circuit construction means 4 for constructing a shape decision circuit for storing an expert's shape recognition decision using data, and a built-in shape decision circuit constructed by the circuit construction means 4, and data inputted from the sensor 1. In response to this, the shape pattern determining means 5 for determining which one of the predetermined patterns is applicable and the teaching data selecting means 3 A pattern determination evaluation for examining data input from the sensor 1 using a selection criterion adopted when selecting teaching data, and evaluating whether or not the shape pattern determination means 5 can determine the data. Means 6 and output means 7 for obtaining and outputting a control output based on the pattern judgment result by the shape pattern judgment means 5 and the pattern judgment evaluation result by the pattern judgment evaluation means 5.

[0020]

In the automatic shape control device of the present invention, first,
Prior to the control, a plurality of sets of data input from the sensor 1 by the teaching data input means 2 and a shape pattern determination example of a skilled person for the data are input as teaching data. These data are stored in the teaching data selection means 3
Sent to Then, in the teaching data selection means 3, a variation between the data from the sensor 1 and the judgment value of the operator is obtained, and the data is selected so as to reduce the variation between the data, and is output to the circuit construction means 4. .

Thus, the circuit construction means 4 constructs a shape determination circuit that derives the same result as the shape pattern determination example of the taught data when referring to a sufficient number of teaching examples having no contradiction. Then, the shape determination circuit constructed by the circuit construction means 4 is output to the shape pattern determination means 5. In addition, the information of the selection criterion employed when the teaching data is selected by the teaching data selection unit 3 is output to the pattern determination evaluation unit 6. Next, after taking the above-described processing procedures, the data taken in from the sensor 1 is input to the shape pattern determination means 5 and the pattern determination evaluation means 6.

Then, the shape pattern judgment means 5 uses the shape judgment circuit constructed by the circuit construction means 4 to perform the same judgment as that of the operator on the input data from the sensor 1 except for a part. Is given. Here, the input data from some of the sensors 1 is a portion corresponding to data that has not been input to the circuit constructing means 4 because of a large variation in the judgment of the operator.

In the pattern judgment evaluation means 6, the same data as the data inputted to the shape pattern judgment means 5 is inputted, and the shape pattern judgment means 5 makes a judgment using the selection criterion used by the teaching data selection means 3. Evaluated for possible data.

Further, the output means 7 obtains and outputs a control output based on the pattern judgment result by the shape pattern judgment means 5 and, if necessary, the pattern judgment evaluation result by the pattern judgment evaluation means 5.

That is, when the input data is “determinable” by the shape pattern determination means 5 in the evaluation by the pattern determination evaluation means 6, the actuator according to the operation pattern of the operator according to the determination pattern of the shape pattern determination means 5. Is driven. Also, if the evaluation by the pattern determination evaluation means 6 is such that the input data is “undeterminable” by the shape pattern determination means 5, the output is not performed or the necessary compensation depends on the degree of inability to determine. Output is performed.

[0026]

DESCRIPTION OF THE PREFERRED EMBODIMENTS In the present invention, when a machine learning operation is performed by an operator, data in which the operator's determination varies from among actual examples of skilled operator's determination is detected and removed, and then stored. In addition, in the actual control, when data belonging to the data in which the judgment of the skilled operator varies appears, the control output is performed in the greatest common divisor. By adding the above two processes to the control device, it is possible to construct an automatic shape control device incorporating the know-how of a skilled operator. Hereinafter, an embodiment of the present invention based on the above concept will be described in detail with reference to the drawings.

FIG. 2 is a block diagram showing an example of the overall configuration of the automatic shape control device according to the present invention. That is, as shown in FIG. 2, the automatic shape control device of the present embodiment includes a shape sensor 11, a data collection device 12, a collected data storage device 13, an operator determination input device 14, a teaching data selection device 15, , Selection criteria storage device 16, selection data storage device 17, neural network learning device 18,
, A network learning command input device 19, a network parameter storage device 20, a pattern determination evaluation device 2
1 and neural network 22 without learning mechanism
And an output device 23.

The shape sensor 11 includes a data collection device 12, a collected data storage device 13,
An operator determination input device 14 is provided in the teaching data input means 2 with the teaching data selection device 15 and the selection reference storage device 1.
6. The selection data storage device 17 includes the teaching data selection device 3, the neural network learning device 18, the network learning command input device 19, and the network parameter storage device 20. The circuit construction device 4 includes the pattern determination evaluation device 21. , The neural network 22 corresponds to the shape pattern determination means 5, the pattern determination evaluation means 6, and the output device 23 correspond to the output means 7.

Here, the shape sensor 11 captures the state of the surface shape of the sheet in the rolling of the sheet metal. The input data from the shape sensor 11 is a set of sensor data of the number (N) of channels arranged in the width direction. The data collection device 12 collects data input from the shape sensor 11 at a constant cycle. Further, the collected data storage device 13 stores data collected by the data collection device 12.

Further, the operator judgment input device 14
Is a display / input device capable of performing an operator determination input, and inputs as data teaching data by pairing data input from the shape sensor 11 and a shape pattern determination example of a skilled person with respect to the input data. That is, the operator determination input device 14 displays data for each set, and the operator assigns a pattern number corresponding to a typical pattern as shown in FIG. 3, for example. In this case, the pattern number is an integer from 1 to K.

On the other hand, the teaching data selection device 15 inputs the above-mentioned plurality of teaching data, finds a variation in the judgment value of the expert with respect to the input data, and selects and outputs the teaching data so as to reduce the variation. Things. Further, the selection reference storage device 16 stores the teaching data selection device 15.
Is used to store the sorting criterion employed when the teaching data is sorted. Further, the selection data storage device 17 stores the teaching data output from the teaching data selection device 15.

On the other hand, the neural network learning device 1
Reference numeral 8 denotes a construction of a shape determination circuit that stores the recognition and determination of the shape of a skilled person using the teaching data stored in the selection data storage device 17. That is, in the neural network learning device 18, when the node of the input layer is N
The output and output layer nodes are composed of K or more networks of three or more layers, and learning is performed by back-promotion. Also, N sensor data is given to the input layer, and the data to be taught therefor is one in which only the node of the pattern number determined by the operator among the K output layers is 1 and the others are 0. . In addition, the network learning command input device 19 is the neural network learning device 1
8 is a display / input device capable of inputting a command for network learning.

Further, the network parameter storage device 20 stores the network parameters of the neural network learning device 18, that is, the shape determination circuit.

On the other hand, the pattern judgment / evaluation device 21 uses the sorting criterion stored in the sorting criterion storage device 16 and adopted at the time of performing the teaching data sorting, to form the shape sensor 11.
This is to examine the data input from and evaluate whether the neural network 22 can determine the data.

The neural network 22 corresponds to any one of predetermined patterns with respect to the data input from the shape sensor 11 based on the network parameters stored in the network parameter storage device 20. Is determined. That is, the neural network 22 is configured as a network having exactly the same number of layers as the neural network learning device 18 and the same number of nodes in each layer. However, these do not have an error feedback mechanism for learning, and only output a determination result using parameters changed by learning in the neural network learning device 18.

Further, the output device 23 obtains and outputs a control output based on the pattern judgment result by the neural network 22 and the pattern judgment evaluation result by the pattern judgment evaluation device 21. Next, the operation of the automatic shape control device according to the present embodiment configured as described above will be described.

In FIG. 2, the data collection device 12
The input from the shape sensor 11 is sampled at regular intervals. Further, by setting, the sampled sensor data is immediately displayed on the operator determination input device 14, and the input of the operator can be simultaneously collected.

Next, the collected data collected by the data collecting device 12 is stored in the collected data storage device 13. The operator determination input device 14 can read data in the collected data storage device 13 and input or correct an operator determination value for sensor data.

Further, the teaching data selection device 15 fetches M pieces of data from the collected data storage device 13 and deletes data for which the operator's judgment varies from almost the same sensor data at a certain threshold value. And
This variation is reduced. Specifically, the processing is performed according to the following procedure. (A) The M pieces of input data are divided into K groups for each pattern number determined by the operator. (B) For each pattern number group, the average of the values of all data for each channel is obtained.

(C) For all data, the distance from each pattern average data is obtained. Here, the distance is 2
It is the sum of the squares of the difference between the values for each of the two data.

That is, if the data is close to the standard form of the operator determined by the operator, the distance from the average data of the pattern determined by the operator is small, and the distance from the average data of other patterns is large.

On the other hand, when the data is an intermediate pattern, the difference between the distance from the average data of the pattern determined by the operator and the distance from the average data of other patterns becomes small.

Therefore, (d) the difference between the distance from the average of the data of the operator judgment pattern number assigned to the data and the average distance of the other data is equal to or more than a fixed value (D) among all the data. Is output to the sorting data storage device 17. in this case,
The constant value (D) is a constant determined for each pattern in consideration of the degree of variation in the judgment of the operator. In addition, the average value of each pattern and the tolerance (D) from the average are stored in the selection criterion storage device 16 as a selection data criterion.
Is output to

On the other hand, the network learning command input device 19
Neural network learning device 1
At 8, data is input from the sorted data storage device 17 and learning is started. That is, in the neural network learning device 18, learning is repeated a predetermined number of times in accordance with an instruction previously input from the network learning command input device 19, or learning is automatically performed when an error in learning becomes a predetermined value or less. finish. In addition,
The network learning command input device 19 can monitor the convergence state during learning in the neural network learning device 18.

In actual control, data input from the shape sensor 11 is input to the neural network 22. In the neural network 22, an answer is obtained using the parameters created by the neural network learning device 18.

Therefore, if there is the same data as the taught data, almost the same answer can be obtained. In addition, the neural network 22 has a property that, even if it is not the same data, if data close to the taught data is shown, an answer close to the taught data is obtained. However, it is effective to some extent.

However, the teaching data selection device 15
Intermediate patterns among the plurality of patterns are deleted from the teaching data, and for such data, the reliability of the determination value of the network is not very high.

Therefore, in the pattern judgment and evaluation device 21,
The distance between the average of the data and the average data for each pattern is calculated in the same manner as in the teaching data selection device 15. And this data is the pattern 1
, The difference in distance from other patterns when belonging to pattern 2, the difference in distance from other patterns when belonging to pattern 2, ...
.., And so on. A candidate for which pattern a stable answer can be obtained by the neural network 22 is obtained and output.

In the output device 23, a control output is determined and output in consideration of the candidate result by the pattern determination evaluation device 21 and the pattern determination result by the neural network 22.

That is, if the pattern number including the candidate for the evaluation evaluation information matches the pattern determination result, it can be said that the pattern belongs to the data learned by the neural network learning device 18 and the network pattern determination result is obtained. Is equal to the judgment value of the skilled operator. In this case, for example, the control output is performed according to the same operation procedure as that of the operator as shown in FIG.

On the other hand, if the pattern number containing the candidate for the evaluation information does not match the pattern determination result, the determination in the neural network 22 is performed for data whose network is unknown. Likely. Therefore, the processing is performed according to the following procedure. (A) The distance from the teaching data average value of the pattern determined by the neural network 22 is obtained. (B) If the calculated distance is less than D1 (<D1), the control output is cut by C1% and output. (C) The obtained distance is D1 or more and less than D2 (D1
If ≦, <D2), the control output is output after being cut by C2%. (D) If the obtained distance is not less than D2 (D2 ≦), the output of the control output is postponed.

As described above, the automatic shape control device according to the present embodiment includes a shape sensor 11 for capturing the state of the surface shape of a sheet in the rolling of a thin metal sheet, data input from the shape sensor 11, and skill in inputting the input data. A data collection device 12, a collected data storage device 13, an operator determination input device 14, which are teaching data input means for inputting as teaching data by pairing with the shape pattern determination example of the user,
Input a plurality of teaching data from teaching data input means,
A teaching data selection device 15, a selection reference storage device 16, and a selection data storage, which are teaching data selection means for obtaining the variation of the expert's determination value with respect to the input data and selecting and outputting the teaching data so as to reduce the variation. A neural network learning device 18 which is a circuit construction means for constructing a shape determination circuit for storing the recognition and determination of a skilled person's shape using the device 17 and the teaching data output from the teaching data selection means; Apparatus 1
9. A built-in network parameter storage device 20 and a shape judging circuit constructed by the circuit constructing means for judging to which of predetermined patterns the data inputted from the shape sensor 11 correspond. The neural network 22 checks the data input from the shape sensor 11 by using a neural network 22 which is a shape pattern determining means to be performed and a selection criterion adopted when the teaching data is selected by the teaching data selecting means. Is based on a pattern determination evaluation device 21 that evaluates whether or not is data that can be determined, a pattern determination result obtained by the neural network 22, and a pattern determination evaluation result obtained by the pattern determination evaluation device 21.
And an output device 23 for obtaining and outputting a control output.

Therefore, since the know-how of the operation of the operator can be utilized without classifying the components into controllable components of the actuator as in the related art, the characteristic peculiar to the equipment, which has been difficult to realize by theory or experiment, And complicated characteristics can be surely dealt with. Further, it is possible to solve a conventional problem that learning is not successfully performed because teaching data varies in a neural network.

This makes it possible to automatically make the same judgment as a skilled operator who has difficulty in mechanization, so that the physical characteristics are complicated and the control which has conventionally been difficult to perform automatically is controlled. A control mechanism for the object can be constructed. In addition, since actual examples of determination may vary, construction is extremely easy. As described above, it is possible to automatically perform high-performance shape control. Note that the present embodiment is not limited to the above embodiment, but can be implemented as follows.

(A) In the above embodiment, the case where the selection of the teaching data is carried out by a statistical method and the circuit construction means is constituted by a neural network has been described. By capturing the variation in some form, the judgment of the expert is
The purpose is to realize the mechanism without having to elucidate it. Therefore, as a method of selecting the teaching data, a technique of presenting the same data to the skilled person a plurality of times and measuring the variation may be employed.

(B) In the above embodiment, the case where the pattern judgment / evaluation device 21 is provided has been described. However, this is not an indispensable element of the present invention, and may be provided if necessary.

[0057]

As described above, according to the present invention, in a shape control apparatus for flattening a sheet surface in a sheet width direction in rolling of a thin metal sheet, a sensor 1 for capturing a state of a sheet surface shape, and a sensor 1 Data input means 2 for inputting as data teaching data by pairing data input from a computer and an example of a skilled person's shape pattern determination with respect to the input data
Teaching data selection means for inputting a plurality of teaching data from the teaching data input means 2, obtaining a variation in the judgment value of the expert with respect to the input data, and selecting and outputting the teaching data so as to reduce the variation. 3, a circuit construction means 4 for constructing a shape determination circuit for storing the recognition and determination of the shape of a skilled person using the teaching data output from the teaching data selection means 3, and a shape constructed by the circuit construction means 4. Shape pattern determining means 5 having a built-in determination circuit and determining which of predetermined patterns corresponds to data input from sensor 1.
The data input from the sensor 1 is examined by using the selection criterion adopted when the teaching data is selected by the teaching data selecting means 3 and the shape pattern determining means 5 is used.
A control output is obtained and output based on a pattern determination evaluation means 6 for evaluating whether or not the data is determinable data, a pattern determination result by the shape pattern determination means 5, and a pattern determination evaluation result by the pattern determination evaluation means 5. And an output means 7 for performing a highly-reliable automatic shape control that can automatically perform extremely high-performance shape control by providing a mechanism for storing examples of skilled operator shape pattern determination with variations. A control device can be provided.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a basic configuration example of a shape automatic control device according to the present invention.

FIG. 2 is a block diagram showing an embodiment of an automatic shape control device according to the present invention.

FIG. 3 is a view for explaining the operation in the embodiment.

FIG. 4 is a view for explaining the operation in the embodiment.

[Explanation of symbols]

REFERENCE SIGNS LIST 1 sensor, 2 teaching data input means, 3 teaching data selection means, 4 circuit construction means, 5 pattern shape determination means, 6 pattern evaluation means, 7 output means, 11
... Shape sensor, 12 ... Data collection device, 13 ... Collection data storage device, 14 ... Operator judgment input device, 15 ...
Teaching data sorting device, 16 ... sorting reference storage device, 17 ...
Sorting data storage device, 18 ... Neural network learning device, 19 ... Network learning command input device, 20 ...
Network parameter storage device, 21: pattern determination / evaluation device, 22: neural network without learning mechanism, 23: output device.

Continued on the front page (72) Inventor Hiroyuki Shinonaga 1-1-1, Shibaura, Minato-ku, Tokyo Inside the Toshiba head office (72) Inventor Takeshi Koyama 1-3-18 Wakihama-cho, Chuo-ku, Kobe-shi, Hyogo Inside Kobe Steel, Ltd. (72) Hiroshi Okashita 1-3-18, Wakihama-cho, Chuo-ku, Kobe City, Hyogo Prefecture Inside Kobe Steel Co., Ltd. (72) 1-3-3 Wakihama-cho, Chuo-ku, Kobe City, Hyogo Prefecture No. 18 Kobe Steel, Ltd. (56) References JP-A-3-20609 (JP, A) JP-A-4-127908 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name ) B21B 37/28

Claims (2)

(57) [Claims] 1. A shape control device for flattening a sheet surface in a sheet width direction in rolling of a thin metal sheet, a sensor for capturing a state of a surface shape of the sheet, data input from the sensor, and skill in input data. Teaching data input means for inputting as teaching data in combination with the shape pattern determination example of the expert, and inputting a plurality of teaching data from the teaching data input means, and determining a variation in the judgment value of the expert with respect to the input data. A teaching data selecting means for selecting and outputting teaching data so as to reduce the variation, and a shape for storing recognition judgment of a skilled person's shape using the teaching data output from the teaching data selecting means. A circuit construction means for constructing a decision circuit; and a shape decision circuit constructed by the circuit construction means. Pattern determining means for determining which of the predetermined patterns corresponds to the data, and output means for obtaining and outputting a control output based on a pattern determination result by the shape pattern determining means And an automatic shape control device, comprising: 2. A shape control device for flattening a plate surface in a plate width direction in rolling of a thin metal plate, a sensor for capturing a state of a surface shape of the plate, data input from the sensor, and skill in the input data. Teaching data input means for inputting as teaching data in combination with the shape pattern determination example of the expert, and inputting a plurality of teaching data from the teaching data input means, and determining a variation in the judgment value of the expert with respect to the input data. A teaching data selecting means for selecting and outputting teaching data so as to reduce the variation, and a shape for storing recognition judgment of a skilled person's shape using the teaching data output from the teaching data selecting means. A circuit construction means for constructing a decision circuit; and a shape decision circuit constructed by the circuit construction means. Shape data determining means for determining which of the predetermined patterns the data corresponds to, and a sorting criterion adopted when the teaching data is sorted by the teaching data sorting means. Pattern determination evaluation means for examining data input from the sensor and evaluating whether the data is data that can be determined by the shape pattern determination means, a pattern determination result by the shape pattern determination means, and the pattern determination Output means for obtaining and outputting a control output based on a pattern determination evaluation result by the evaluation means; and an automatic shape control device, characterized by comprising: