JPH07204718A - Rolling device - Google Patents

Abstract

PURPOSE:To reduce a computation time, and to perform rolling with satisfactory accuracy by calculating using a neural network with rolling information before a rolling stock is bitten by a finish rolling mill, and rolling it with the setting value of a rolling load derived. CONSTITUTION:Prescribed pressure information and the actual result data on the outlet side of a roughing mill, are inputted off-line in a neural network 16. Based on this input, the neural network 16 is allowed to learn. Rolling is started, and at the time when the tip part of a rolling stock 1, is rolled out from a final roughing mill 2, actual results (plate speed, plate width and plate temperature) on the outlet side of the roughing mill, the estimated value of plate thickness on the outlet side of the roughing mill and rolling information (the kind of a steel, the content of carbon, a target plate thickness on the outlet side of a finish, etc.), are given as the input values of a rolling load setting device 15. Each normalized input value is inputted in a neural network learned off-line previously. A forward calculation is performed, the output value of the neural network, is determined, and it is restored by a nonormalized arithmetic part. It becomes the output value of a rolling load setting device 15, and rolling is performed by this output value.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a rolling mill for predicting a rolling load and obtaining a predetermined plate thickness with the predicted value.

[0002]

2. Description of the Related Art A block diagram of a rolling mill is shown in FIG. There will be 5 finishing rolling mills. The rolled material 1 is rolled by the final rough rolling mill 2 and then passed through the finish rolling mills 3-1 to 3-5 and wound by the winder 4. On the exit side of the final rough rolling mill 2, a rough rolling mill exit side speed meter 5, a rough rolling mill exit side width meter 6 and a rough rolling mill exit side thermometer 7 are installed. The detection signals of the rough rolling mill exit side speed meter 5, the rough rolling mill exit side width meter 6 and the rough rolling mill exit side thermometer 7 are given to the plate thickness setting device 8. The plate thickness setting device 8 stores various kinds of rolling information such as steel type, carbon content, target finish rolling mill outlet temperature and the like. Final finishing rolling mill 3-
On the output side of 5, a finishing rolling mill output side speed meter 9, a finishing rolling mill delivery side thickness gauge 10 and a finishing rolling mill delivery side thermometer 11 are installed. The detection signals of the finish rolling mill exit side speed meter 9, the finish rolling mill exit side thickness gauge 10 and the finish rolling mill exit side thermometer 11 are given to the plate thickness setting device 8. Further, each of the finish rolling mills 3-1 to 3-5 has a rolling load detector 12-1.
To 12-5 and roll gap setting devices 13-1 to 13-13
-5 is installed.

Next, the operation will be described. The plate thickness setting device 8 sets up the finish rolling machine as follows in order to secure a required final finishing rolling mill delivery side plate thickness for the leading end portion of the rolled material 1 after rough rolling. First, the speed, temperature and strip width of the leading end portion of the rolled material 1 on the exit side of the final rough rolling mill are respectively measured by the rough rolling mill exit side speedometer 5, the rough rolling mill exit side width meter 6 and the rough rolling mill exit side thermometer. 7, and the rolling load necessary to obtain the required final finishing rolling mill outgoing side plate thickness is input as the sheet thickness setting device 8 with the rough rolling mill outgoing side plate thickness obtained by these detected values and schedule calculation. Calculate for each finishing rolling mill. This calculation is performed when the tip of the rolled material 1 leaves the rough rolling mill 2. The roll gap for each finishing rolling mill is calculated from the obtained rolling load setting value, and is input to the roll gap setting devices 13-1 to 13-5. The actual rolling load values detected by the rolling load detectors 12-1 to 12-5 are
It is sent to the plate thickness setting device 8 and used for learning calculation for the next rolled material. For the rolling load prediction calculation in the strip thickness setting device 8, the Smis theoretical formula (Equation 1) is usually used.

[0004]

[Equation 1]

P (rolling load) and R '(flat roll radius)
Solve as a binary simultaneous equation with unknown as. At this time, repeated calculation is required, but various simple formulas for obtaining the rolling load P without repeating calculation have been devised due to constraints such as calculation time.

[0006]

The theoretical formula as described above is based on various approximations or assumptions.
Therefore, in the prediction method using the conventional formula, in order to improve the accuracy and to perform fine control, it is necessary to take measures such as increasing the parameters that complicate the formula, increasing the labor required for tuning and the calculation time. Led to an increase.
Moreover, a constant number depending on the characteristics of the plant requires maintenance every time the state of the plant changes, and a complicated control method imposes a heavy burden on the operator during maintenance. Therefore, by using a neural network that is characterized by non-linear characteristics, learning ability, and parallelism, it is possible to accurately control the non-linearity of the rolling phenomenon and flexibly respond to changes in plant characteristics. Also, by using a neural network learned offline, online, the calculation time can be greatly reduced. It is an object of the present invention to provide a rolling apparatus that can reduce the calculation time and accurately roll the leading end of a rolled material to a required strip thickness without using a complicated mathematical expression by using a neural net. is there.

[0007] As a related art related to this, Japanese Unexamined Patent Publication No. 5-42314 “Rolling material shape control device for rolling mill”
Is to control the shape of the plate thickness and the plate thickness using the learning function. In addition, Japanese Patent Laid-Open No. 5-27806 discloses a "feedforward controller using a neural net".
Feedforward control is performed using a neural net so that it has excellent followability when a target value is changed or a disturbance occurs. Further, Japanese Patent Laid-Open No. 5-27808 discloses a "controller using a neural net model" that uses a neural net to control the non-linear characteristics of a plant and also to control disturbances.

[0008]

According to claim 1 of the present invention,
A rolling device for controlling and rolling a plate thickness of a rolled material to match a target plate thickness is provided with a rolling load setting device having a neural net.

According to a second aspect of the present invention, in the rolling apparatus according to the first aspect, at least a rolling load actual value is input and a neural net learning device for updating the neural net is provided.

According to a third aspect of the present invention, in the rolling apparatus according to the first aspect, at least a rolling load actual value is sequentially input and a history of the rolling load actual value is stored, and a neural net learning data file and the neural net learning data. A neural net learning device for inputting history data of a file and updating the neural net.

According to a fourth aspect of the present invention, in the rolling apparatus of the first aspect, at least a rolling load actual value is sequentially input and a history of the rolling load actual value is stored, and a neural net learning data file and the neural net learning data file. It is provided with a neural net learning device for inputting history data of a file to update the neural net, and a rolling load learning coefficient file for storing a rolling load learning coefficient according to an actual rolling load value.

According to a fifth aspect of the present invention, which comprises a rough rolling mill and a finish rolling mill, before the tip of the rolled material is caught in the finish rolling mill, the final finishing rolling mill delivery side plate thickness matches the target plate thickness. Thus, in the rolling apparatus for setting up the finishing rolling mill and rolling, a rolling load setting apparatus having a neural net is provided.

[0013]

According to claim 1 of the present invention, rolling information such as temperature, speed, strip width and strip thickness at the tip of the rolled material immediately before being bitten into the rolling mill is input, and the rolling load set value is set using the neural net. Is calculated, and the plate thickness is controlled using this rolling load set value.

According to a second aspect of the present invention, the neural net learning device inputs the actual rolling load value for each rolled material,
The neural net is updated, and the rolling load set value is calculated using the updated neural net.

According to a third aspect of the present invention, the history of rolling load actual values for each rolled material is sequentially stored in the neural net learning data file, and the neural net learning device is used by using the stored plural rolling load actual values. The neural net is updated with and the rolling load set value is calculated using the updated neural net.

According to a fourth aspect of the present invention, the history of rolling load actual values for each rolled material is sequentially stored in the neural net learning data file, and the neural net learning device is used by using the stored plural rolling load actual values. The neural net is updated with and the calculation is performed using the updated neural net, and the calculated value is corrected according to the rolling load learning coefficient to derive the rolling load set value.

According to a fifth aspect of the present invention, rolling information such as temperature, speed, strip width and strip thickness at the leading end of the rolled material on the output side of the final rough rolling mill is input, and a rolling load set value is set using a neural network. Is calculated, and the finishing rolling mill is set up so that the target plate thickness can be obtained by using this rolling load setting value, and rolling is performed.

[0018]

【Example】

Example 1. The application range of neural networks is wide-ranging,
There are many practical examples of new methods for recognition, control, prediction, etc. The basic items of the neural network will be briefly explained.

(1) A neural network is a model that literally mathematically models a biological neural network and reproduces its characteristic functions. The neural network is composed of basic elements called neurons and synapses connecting the neurons. (Fig. 1)

(2) The features of the neural network are summarized below. It does not require a program because it has the ability to acquire knowledge by learning examples. It is suitable for interpreting pattern information such as voice and images because it can handle information that is difficult to quantify and incomplete information. Parallel processing is possible. Excellent generalization. It has excellent robustness. Even if there is noise or there is a partial failure or variation in characteristics of the neural network itself, it is unlikely to be affected by them. The present invention focuses particularly on the above.

(3) The "unit model" which is the theoretical model of the neural network will be described with reference to FIG. A neural network can be thought of as a parallel processing system with units as basic processors. The neuron is a multi-input / single-output type arithmetic element, and the synapse is an analog amount memory element. For this "analog amount",
Although there are various names such as synapse load, synapse bond strength, bond strength, and bond coefficient, in the present invention, they are unified as “bond load”. As the output of the unit, a model is often used in which the sum of products of the input signal and the coupling weight is transformed by the response function f and given.

(4) The response function includes a digital type Heaviside function (FIG. 3-a), an analog type sigmoid function (FIG. 3-b), a probabilistic model, a chaotic model, etc., but the sigmoid function should be used. There are many. Because this function well reflects the saturated nature of nerve cells.

(5) Structurally, the neural network model can be roughly classified into a hierarchical type (FIG. 4-a) and an interconnected type (FIG. 4-b). In the present invention, a typical hierarchical type is used. As a model, a backpropagation model having many practical examples was used.

(6) The following elements of the neural network are variable and can be freely adjusted according to the system to be controlled. Number of network layers Number of units Response function Correction coefficient of coupling weight / bias value Number of learning

Learning in the neural network means that a desired output is obtained when a certain input is given.
Strength of connection between neurons (connection weight: synaptic weight)
Is to change. Specifically, the input value is transmitted from the input layer to the output layer, the output error (difference between the output value and the teacher signal) is calculated, and then this output error is propagated from the output layer to the input layer. This means changing the coupling weight and the bias value in the process. In the input layer, the input value is used as the output value as it is, but in the middle layer and thereafter, the output value is obtained by adding the bias value of the own layer to the product sum of the output value of the previous layer and the coupling weight.

The back propagation model mathematically follows an optimization method called "steepest descent method" and modifies the internal coupling state in the direction of minimizing the squared error evaluation function in the output layer. Is. It is desirable to use as much data as possible without bias (a well-represented controlled object) for learning. Using this data, learning is repeated until the output error falls within the desired range (allowable range). At this time, it is necessary to be careful that the neural network learns the learning data too much and loses generalization so that it cannot deal with other data of the same population (same control target). It is to prevent it. This phenomenon is commonly referred to as "overlearning."

In order to avoid "over-learning", the output error (prediction accuracy) obtained when data (test data) that the neural network has not learned is passed through the neural network during learning of the neural network, A method of stopping learning when the prediction accuracy is minimized is generally used.

Learning of a four-layer back propagation model (FIG. 5) will now be described as an example. Although FIG. 5 shows the case where there is one output value, the same applies when learning a plurality of output values. This specific calculation is represented by the formula (Equation 2). This equation is a forward calculation for calculating the output value of the neural net, and the output error (difference between the output value and the teacher signal = target plate thickness) is obtained by the calculation. This output error is calculated in the reverse direction of (Equation 3) and (Equation 4) to update the coupling weight and the bias value. Then, the learning is repeated until the output error falls within the desired range (allowable range).

[0029]

[Equation 2]

[0030]

[Equation 3]

[0031]

[Equation 4]

Next, a data processing method will be described. The data handled in the network must be converted in consideration of the input / output to the response function. Various methods can be considered for the conversion, but here, the output range as the response function is −
Data conversion (normalization) will be described by taking a case of using a sigmoid function of 0.5 to 0.5 as an example.

The input value to the network and the teacher signal are
It is normalized to the range of -0.5 to 0.5 by the following formula. x * = (x-xmin) / (xmax-xmin) x: input value x *: normalized input value xmax: maximum value of x xmin: minimum value of x

The output value is restored by the following equation. y * = y (ymax-ymin) + ymin y: network output value y *: restored output value ymax: maximum y value ymin: minimum y value

The maximum value and the minimum value are set for each stand by searching all the data used for learning the neural network, but predetermined upper and lower limit values may be used.

FIG. 6 shows an example of the structure of the rolling force prediction neural net. Regarding the input to the input layer, FIG.
This will be described below.

FIG. 7 is a block diagram of a rolling apparatus showing Embodiment 1 of the present invention. The number of finish rolling mills is 5, and the rolled material 1 is rolled by the final rough rolling mill 2 and then finished rolling mills 3-1 to 3-3.
It is taken up by the winder 4 through -5. On the exit side of the final rough rolling mill 2, a rough rolling mill exit side speed meter 5, a rough rolling mill exit side width meter 6 and a rough rolling mill exit side thermometer 7 are installed. The detection signals of the rough rolling mill exit side speed meter 5, the rough rolling mill exit side width gauge 6, and the rough rolling mill exit side thermometer 7 are given to the plate thickness setting device 14.

The strip thickness setting device 14 includes a rolling load setting device 1
5 is built in and stores steel type, carbon content, target finish rolling mill outlet temperature, and other various rolling information. The rolling load setting device 15 has a neural net 16 built therein. A finish rolling mill exit side speed meter 9, a finish rolling mill exit side thickness gauge 10 and a finish rolling mill exit side thermometer 11 are installed on the exit side of the final finishing rolling mill 3-5. The detection signals of the finish rolling mill exit side speed meter 9, the finish rolling mill exit side thickness gauge 10 and the finish rolling mill exit side thermometer 11 are given to the plate thickness setting device 14. Further, the finish rolling mills 3-1 to 3-5 have rolling load detectors 12-1 to 12-1 respectively.
2-5 and roll gap set values 13-1 to 13-5 are set.

FIG. 8 shows a detailed block diagram of the control state of the first embodiment shown in FIG. The rolling load setting device 15 transmits plant constants, steel types, carbon contents, target finish rolling mill outlet temperatures, and other various rolling information necessary for setting rolling loads. The rolling load setting device 15 is input with actual product data (each detection signal such as strip speed, strip width, strip temperature) obtained on the exit side of the final rough rolling mill, and various rolling information such as steel type and carbon content rate. In addition, the neural network 16 in which the actual rolling load value is learned as a teacher signal is incorporated. Further, a normalization calculation unit 17 and a non-normalization calculation unit 18 are incorporated.

Next, the operation will be described with reference to the flowchart of FIG. (1) Predetermined pressure information and performance data on the rough rolling mill output side are input to the neural net 16 offline (step S1). (2) The neural net 16 is learned based on this input (step S2). (3) Rolling is started (step S3). (4) It is determined whether or not the tip of the rolled material 1 has passed through the final rough rolling mill 2, and if the tip is pulled out (step S4), (5) the tip of the rolled material 1 moves through the final rough rolling mill 2. At the time of exiting, the rough rolling mill output side results (plate speed, plate width, plate temperature), rough output side plate thickness estimated value and rolling information (steel type, carbon content rate, finishing output side target plate thickness etc) are set as rolling load. It is given as an input value of the device 15 (step S5).

(6) Each input value is normalized with a normalized calculation value by using preset upper and lower limit values. The upper and lower limit values at this time are the same as the values used for learning the neural network (step S6). (7) Each normalized input value is input to a neural network that has already been learned offline (step S
7). (8) The forward calculation according to (Equation 2) is performed to determine the output value of the neural network (step S
8).

(9) The output value of this neural network is restored by the denormalization calculation unit and becomes the output value of the rolling load setting device 15 (step S9). (10) Using the output value of the rolling load setting device 15 as the rolling load setting value, the rolling load of the rolling mill is set and rolling is performed (step S10).

The effect of the first embodiment is very simple as compared with the conventional mathematical model by performing the calculations (6) to (8) using the neural net.
There is an advantage that the calculation time is significantly reduced. Furthermore,
Conventional flat roll radius (R '), average deformation resistance (k
Since an intermediate variable such as pm) is not used, highly accurate plate thickness control can be realized without being affected by calculation accuracy.

Example 2. FIG. 9 shows a block diagram of the control state of the second embodiment of the present invention. This block diagram is obtained by providing the neural net learning device 19 in FIG. 8 of the first embodiment.

Next, the operation will be described with reference to the flow charts of FIGS. Since the operation from step S1 to step S10 is the same as that of the first embodiment, the description thereof will be omitted, and description will be given from step S21 of FIG. (1) It is determined whether the tip of the rolled material 1 has passed through the final finishing rolling mill 3-5 (step S21).

(2) When the tip of the rolled material 1 leaves the final finishing rolling mill 3-5, the rolling load detecting devices 12-1 to 12-1
Each detected value 12-5 is input to the rolling load setting device 15 (step S22). (3) This rolling load actual value is normalized by the normalization calculation unit to be a teacher signal of the neural network 16 (step S2).
3). (4) The teacher signal of step S23 and step S7
The input value of is used as the input value of the neural network learning device 19, the neural network 16 is relearned, and the neural net 16 is updated. As the learning method, as described above, the forward calculation by the above (Equation 2) is performed, and if it is not within the set prediction error, the coupling weight and the bias value are updated by the backward calculation by the above (Equation 3). , Repeat the calculation so that it is within the prediction error. At this time, even if the error does not fall within the prediction error, the online calculation time is limited, so the maximum number of learning times is set in advance in consideration of the capacity of the computer (step S24). (5) When calculating the rolling load setting for the tip of the next material, the rolling load setting value is determined using the updated neural net 6 (step S25).

The effect of the second embodiment is that, in addition to the advantage of the first embodiment that the calculation time is greatly shortened and the calculation accuracy is not affected, the neural net is updated for each rolled material. As a result, it is possible to flexibly control the plate thickness even if there is a temporal change in the plant including thermal expansion and wear of the roll.

Example 3. FIG. 10 shows a block diagram of the control state of the third embodiment of the present invention. This block diagram is obtained by providing the neural net learning device 19 and the neural net learning data file 20 in FIG. 8 of the first embodiment.

Next, the operation will be described with reference to the flowcharts of FIGS. Since the operation from step S1 to step S10 is the same as that of the first embodiment, it is omitted, and description will be given from step S31 of FIG. (1) It is determined whether the tip of the rolled material 1 has passed through the final finishing rolling mill 3-5 (step S31), and (2) the tip of the rolled material 1 has passed through the final finishing rolling mill 3-5. At the time point, each detected value of the rolling load detecting devices 12-1 to 12-5 is input to the rolling load setting device 15 (step S32).

(3) This rolling load actual value is normalized by the normalization calculation unit to be a teacher signal of the neural network 16 (step S33). (4) The teacher signal of step S33 and the input value of step S7 are added to the end of the neural network learning data file, and the head data, that is, the oldest data is deleted from the neural network learning file. Therefore, in this neural network learning data file, the data (a plurality of pairs of teacher signals and input values) used for off-line learning of the neural network are first put, and then,
The data used for learning the neural net online is sequentially stored. The number of data stored at this time is set in advance in consideration of the capacity of the computer (step S
34).

(5) In step S34, the neural net 16 is re-learned by using the neural net learning data file 20 in which the data is newly organized, and the neural net 16 is updated. As the learning method, as described above, the forward calculation by the above (Equation 2) is performed, and if it is not within the set prediction error, the coupling weight and the bias value are updated by the backward calculation by the above (Equation 3). , Repeat the calculation so that it is within the prediction error. At this time, even if the error does not fall within the prediction error, the online calculation time is limited. Therefore, the maximum number of learning times is set in advance in consideration of the ability of the computer (step S35). (6) When calculating the rolling load setting for the tip of the next material, the rolling load setting value is determined using the updated neural net 16 (step S36).

The effect of the third embodiment is that, in addition to the advantage of the first embodiment that the calculation time is greatly shortened and the calculation accuracy is not affected, the existing neural net learning data is added to each rolled material. Data is newly added to the neural network and learning of the neural network is performed using these data, so that the neural network learning data is newly organized every time rolling of one rolled material is completed, and the neural network is updated. By doing so, it is possible to realize plate thickness control that flexibly responds to changes over time in plant characteristics based on the plant history. Compared to Example 2, the results are not so different from those of Example 2 when there is a certain change with time, but when there are many changes with time, past data is also used, and therefore Example 3 is more preferable. Stable results are obtained. Further, in the second embodiment, when an abnormality detection value is generated due to a sensor failure, noise, or the like, the abnormal value is learned, but in the third embodiment, since the history up to this point is taken into consideration, the influence of the abnormality. Can be very small.

Example 4. FIG. 11 shows a block diagram of the control state of the fourth embodiment of the present invention. This block diagram is obtained by providing the neural net learning device 19, the neural net learning data file 20, and the rolling load learning coefficient file 21 in FIG. 8 of the first embodiment.

Next, the operation will be described with reference to the flow charts of FIGS. 12, 14 and 15. Since steps S1 to S10 are the same operations as those of the first embodiment and steps S31 to S35 are the same operations as those of the third embodiment, they will be omitted and description will be made from step S41 of FIG.

(1) The ratio between the rolling load set value in step S10 and the actual rolling load value in step S33 is calculated, and this value is used as a rolling load learning coefficient.
1 (step S41). (2) At the time of calculating the rolling load setting of the tip of the next material, the updated output value of the neural net 16 is added to the step S4.
The rolling load learning coefficient of 1 is multiplied to determine the rolling load setting value (step S42).

The effect of the fourth embodiment is that the operation time, which is an advantage of the first embodiment, is significantly shortened and is not affected by the operation accuracy, and that the advantage of the third embodiment is that the existing neural network has the advantage. Data is newly added to the net learning data for each rolled material, and the neural net learning is performed by using these data, so that the neural net learning data is obtained every time rolling of one rolled material is completed. By newly forming and updating the neural net, in addition to the advantage that it is possible to realize plate thickness control flexibly corresponding to changes over time in plant characteristics based on plant history, store rolling load learning coefficient, Since it is used when the rolling load is set for the next rolled material tip, there is an effect that the plate thickness accuracy can be further improved.

The rolling load learning coefficient is the ratio of the rolling load actual value and the rolling load set value, but there is a case where the difference is taken. In short, the rolling load learning coefficient depends on the actual rolling load value. It should take a value.

Example 5. The calculation of the rolling load set value using the neural net of the above embodiment is not limited to the case where the rolling load set values of all the rolling mills of the finishing rolling mill are calculated by the neural net, but the rolling load setting value of the final finishing rolling mill. If the value is calculated by neural net and the schedule of other rolling mills is calculated, the rolling load setting value of any other desired rolling mill (one or more rolling mills) can be calculated by neural net and other schedules can be scheduled. When it is calculated, it can be applied in any case.

When a plurality of rolling mills or all rolling mills are calculated by the neural net, it is better to install a thickness gauge on the output side of the target rolling mill and input the measured strip thickness to the neural net. Rolling accuracy can be improved rather than entering the calculated plate thickness.

Example 6. In the above examples, the continuous rolling mill having the rough rolling mill and the finish rolling mill was described, but the present invention can be applied to other rolling mills, and is also applied to reverse rolling or the like in which rolling material is reciprocally rolled. it can.

[0061]

As described above, according to claim 1 of the present invention, the rolling load is calculated by using the neural net without using a complicated mathematical expression, so that the calculation time can be greatly shortened. Also, the rolling mill can be set with high accuracy. Therefore, there is an effect that it is possible to perform plate thickness control that is responsive and accurate.

According to the second aspect of the present invention, since the neural net is updated by using the neural net learning device every time one rolled material is rolled, it is possible to flexibly cope with the change in plant characteristics. It is possible to realize the controlled plate thickness.

According to claim 3 of the present invention, the neural net is updated every time one rolling material is rolled in consideration of past rolling history by using the neural net data file and the neural net learning device. Since this is done, it is possible to realize plate thickness control that flexibly responds to changes in plant characteristics, including changes over time in the plant.

According to claim 4 of the present invention, each time one rolling material is rolled in consideration of the past rolling history by using the neural net data file, the neural net learning device, and the rolling load learning coefficient file. Since the neural net was updated and the set value was corrected using the rolling load learning coefficient, more accurate plate thickness control that flexibly responds to changes in plant characteristics including changes over time in the plant can be performed. realizable.

According to the fifth aspect of the present invention, the rolling load is calculated by using the neural net without using a complicated mathematical expression, so that the calculation time can be greatly shortened.
Further, the rolling mill can be set up with high accuracy. Therefore, there is an effect that it is possible to perform plate thickness control that is responsive and accurate.

[Brief description of drawings]

FIG. 1 is a diagram showing a basic configuration of a neural network.

FIG. 2 is a diagram showing a unit model.

FIG. 3 is a diagram showing a response function.

FIG. 4 is a diagram showing a neural net model.

FIG. 5 is a diagram showing a structure of a pack propagation model.

FIG. 6 is a diagram showing a structure of a rolling force prediction neural network.

FIG. 7 is a configuration diagram of a rolling apparatus according to the first embodiment of the present invention.

FIG. 8 is a block diagram showing a control state of the plate thickness setting device according to the first embodiment of the present invention.

FIG. 9 is a block diagram showing a control state of the plate thickness setting device according to the second embodiment of the present invention.

FIG. 10 is a block diagram showing a control state of the plate thickness setting device according to the third embodiment of the present invention.

FIG. 11 is a block diagram showing a control state of the plate thickness setting device according to the fourth embodiment of the present invention.

FIG. 12 is a flowchart of the operation in the first embodiment of the present invention.

FIG. 13 is a flowchart of the operation in the second embodiment of the present invention.

FIG. 14 is a flowchart of the operation in the third embodiment of the present invention.

FIG. 15 is a flowchart of the operation in the fourth embodiment of the present invention.

FIG. 16 is a configuration diagram of a conventional rolling apparatus.

[Explanation of symbols]

 1 Rolled material 2 Final rough rolling mill 3-1 to 3-5 Finish rolling mill 4 Winding machine 5 Rough rolling mill exit side speed meter 6 Rough rolling mill exit side width gauge 7 Rough rolling mill exit side thermometer 8 Plate thickness setting Device 9 Finishing rolling mill exit side speed meter 10 Finishing rolling mill exit side thickness gauge 11 Finishing rolling mill exit side thermometer 12-1 to 12-5 Rolling load detector 13-1 to 13-5 Roll gap setting device 14 Sheet thickness Setting device 15 Rolling load setting device 16 Neural network 17 Normalization calculation unit 18 Denormalization calculation unit 19 Neural net learning device 20 Neural net learning data file 21 Rolling load learning unit

Claims (5)

[Claims] 1. A rolling apparatus for controlling and rolling a plate thickness of a rolled material so as to match a target sheet thickness, comprising a rolling load setting device having a neural net, and a rolling material tip portion immediately before being bitten into a rolling mill. The rolling information such as temperature, speed, strip width and strip thickness is input, the rolling load set value is calculated using the neural net, and the strip thickness is controlled using this rolling load set value. And rolling equipment. 2. The rolling apparatus according to claim 1, further comprising a neural net learning device for inputting at least the rolling load actual value and updating the neural net, and the neural net learning device provides the rolling load actual value for each rolled material. A rolling mill characterized by updating a neural net as an input and using the updated neural net to calculate a rolling load set value. 3. The rolling mill according to claim 1, wherein at least a rolling load actual value is sequentially input and a history of rolling load actual values is stored, and a history data of the neural net learning data file is input. And a neural net learning device for updating the neural net, and sequentially stores the history of rolling load actual values for each rolled material in the neural net learning data file, and stores the stored plural rolling load actual values. A rolling device, wherein the neural net learning device is used to update the neural net, and the rolling load set value is calculated using the updated neural net. 4. The rolling mill according to claim 1, wherein at least a rolling load actual value is sequentially input and a history of the rolling load actual value is stored, and a history data of the neural net learning data file is input. And a rolling load learning coefficient file that stores a rolling load learning coefficient according to the actual rolling load value, and the neural net learning data file includes rolling load for each rolled material. The history of actual values is sequentially stored, the neural net is updated by the above-mentioned neural net learning apparatus using the stored plural rolling load actual values, and the updated neural net is used for calculation and this calculation is also performed. The value should be corrected according to the rolling load learning coefficient to derive the rolling load setting value. And a rolling device. 5. A finish rolling mill comprising a rough rolling mill and a finish rolling mill, so that the final strip thickness of the final rolling mill matches the target strip thickness before the tip of the rolled material is caught in the finish rolling mill. In the rolling mill that sets up and rolls, a rolling load setting device with a neural net is provided, and rolling information such as the temperature, speed, strip width and strip thickness of the above-mentioned rolled material tip on the exit side of the final rough rolling mill is displayed. As an input, the rolling load set value is calculated using the neural network, and the finish rolling machine is set up so that the target plate thickness is obtained by using the rolling load set value and rolling is performed. And rolling equipment.