JP2804161B2 - Method and apparatus for controlling shape of Sendzimir mill - Google Patents

Description

Description: BACKGROUND OF THE INVENTION The present invention relates to a shape control method and apparatus for controlling the shape of a material to be rolled by a Sendzimer mill.

[Conventional technology]

In rolling of a sheet material, controlling the shape is an important matter in obtaining a uniform sheet material without distortion, and various shape control methods have been proposed for various types of rolling mills so far. .

For example, a Sendzimer mill is suitable for rolling a hard thin plate, and is used in many practical applications. Such a sendimer mill is described in, for example, JP-A-59-232608. Conventionally, shape control in a rolling mill is performed by operating an actuator corresponding to the output of a shape detector. However, the shape control performance was limited because interference between actuators could not be clarified.
In particular, in the Sendzimer mill, many rolls are complicatedly entangled, and the relationship between the operation of the actuator and the shape of the material to be rolled was qualitatively understood, but it was not quantitatively clear and a control model could not be constructed It is a fact.

For this reason, a control theory that enables high-precision control if an accurate control model exists cannot be put to practical use.

[Problems to be solved by the invention]

The prior art does not consider the overall shape and pattern of the material to be rolled and the effective use of the vague but high-precision control possessed by the operator, resulting in local control. However, there is a problem that high-precision shape control cannot be performed.

SUMMARY OF THE INVENTION An object of the present invention is to provide a shape control method and apparatus capable of controlling a shape of a sendzimir mill with high accuracy.

[Means for solving the problem]

The present invention provides a neural network that obtains an operation amount of an actuator by a difference between a detected shape pattern and a target shape pattern detected by a shape detector arranged on the output side of a sendzimir mill, and a similarity ratio of a late reference shape pattern, At the time of learning the neural network, the actuator is actuated from the difference between the detected shape pattern and the target pattern to obtain a reference shape pattern capable of turning the material to be rolled into the target shape pattern and input to the neural network.
Learning is performed on the relationship between the amount of operation of the actuator and the amount of operation of the actuator to form the reference shape pattern.

[Action]

The reference shape pattern used by the neural network for learning is a shape pattern that can correct the shape of the material to be rolled by operating an actuator, and the difference between the detected shape pattern and the target shape pattern during actual operation and the reference shape pattern The similarity ratio indicates the component amount with respect to the target shape pattern, and since the operation amount of each actuator that performs the correction operation is determined based on the component amount, the shape control accuracy can be improved.

〔Example〕

 Hereinafter, an embodiment of the present invention will be described with reference to FIG.

A sendzimer mill (hereinafter abbreviated as ZR mill) 1 composed of a plurality of rolls is driven by an electric motor 3 through a gear 2 to roll a material 4 to be rolled. The material 4 to be rolled is supplied from one of the left tension reel 5 and the right tension reel 6 and wound on a local reel. These left and right tension reels 5 and 6 are connected to gears 7 and 8 and electric motors 9 and 10, respectively, and by controlling the speeds of the electric motors 3, 9 and 10, respectively, the entrance and exit sides of the rolling mill 1 are controlled. Is controlled on the material 4 to be rolled. The thickness and shape of the material 4 to be rolled are measured by a thickness gauge 11 and a shape detector 12, and the measured values are input to a control mechanism 14. Control mechanism 14
Is a shape detection mechanism 15 which processes the output of the shape detector 12 and outputs it as shape data, a pattern recognition which processes the shape data and outputs the degree of coincidence (similarity ratio) with a pre-stored pattern (reference shape pattern) Control operation mechanism for reflecting ambiguous operation of mechanism 16 and operator in control
Consists of 17 The control mechanism 14 generates an operation command for the actuator 13 of the ZR mill 1.

 FIG. 2 shows a configuration diagram of the ZR mill 1.

The ZR mill 1 directly contacts the material 4 to be rolled, and a pair of upper and lower work rolls 18 that thinly extend the material 4 to be rolled, a first intermediate roll 19 that sandwiches the work roll 18 from above and below, and applies a load to the work roll 18. The second intermediate roll 20 that applies a load across the roll 19 and the second intermediate roll 20 that sandwiches the second intermediate roll 20 from above and below
It is composed of an AS-U roll 21 which applies a load to the intermediate roll 20.

AS-U roll 21 is divided into split roll 22, shaft 23 and saddle (S
addle) 24. The saddle 24 is moved up and down by the actuator 13-1 according to a command of the control mechanism 14,
The shaft 23 is bent according to the amount of movement. By operating the plurality of saddles 24, the shaft 23 is deformed into an arbitrary shape. Axis 23
Of the second intermediate roll via a plurality of split rolls 22
20, the first intermediate roll 19, the work roll 18, and the material to be rolled 4 are sequentially transferred, whereby the shape of the material to be rolled 4 is controlled. One end of the first intermediate roll 19 is tapered so that the load applied to the material to be rolled 4 immediately below the taper is reduced, and the end of the material to be rolled 4 is operated to be thick. First
The intermediate roll 19 is moved in the axial direction by the actuator 13-2 according to the operation command of the control mechanism 14. 1st intermediate roll
The shape of the end of the material 4 to be rolled is controlled by the movement of the member 19 in the axial direction.

 Next, the operation of the ZR mill 1 will be described with reference to FIG.

FIG. 3 (a) shows the upper right area of the roll shown in FIG. 2, and the direction of movement is indicated by adding an arrow to the part moved by the actuator 13. FIG.

FIG. 3 (b) shows a cross section of the sheet 4 to be rolled when the saddle 24 is operated to be pushed down.

FIG. 3 (c) shows the shape of the material 4 to be rolled when the first intermediate roll 19 having a tapered end is moved and the position of the taper is applied to the end.

 FIG. 4 shows an example configuration of the control mechanism 14.

In FIG. 4, the output of the shape detector 12 is input to a waveform calculation mechanism 26 via a shape detection pre-processing mechanism 15 for removing noise and the like. On the other hand, the output of the target shape generation mechanism 25 for generating the target shape of the material to be rolled 4 is also input to the waveform calculation mechanism 26. The waveform calculating mechanism 26 subtracts the target shape (target shape pattern) from the shape detection result (detected shape pattern), obtains a shape deviation Δε to be corrected by control, and inputs the shape deviation Δε to the pattern recognition mechanism 16. The pattern recognition mechanism 16 calculates the shape deviation Δε
The similarity ratio of each waveform component (a reference shape pattern stored in the pattern recognition mechanism 16 in advance) included in the data is output to the control operation mechanism 17. The control operation mechanism 17 makes the operator's ambiguous operation knowledgeable and uses fuzzy control for control,
The control model is output to the actuator 13 using a control algorithm such as multivariable control when the control model is clear.

FIG. 5 (a) shows an example of a neuro-computer 27 constituting the pattern recognition mechanism 16.

NeuroComputer 27 is a Neural Network 2
8, an input switching switch 29, an output switching switch 30, and a learning mechanism 31. The learning mechanism 31 includes a learning control mechanism 32 for controlling each element at the time of learning, an input signal generation mechanism 3
4, a teacher signal generating mechanism 35 and a parameter changing mechanism 33 for changing the load of the neural net 28.

FIG. 5 (b) shows the configuration of the Neuron 36 which constitutes the neural network 28. Niuron 36 input signal x _i
Is multiplied by the value w _i of the load 37, and the adder 38 adds the product of the input signal x _i input to the neuron 36 and the value w _i of the load. The addition result of the adder 38 is input to the function converter 39. The function converter 39 performs a function conversion capable of continuous differentiation such as a SIGMOID function, and the data after the function conversion is output from the neuron 36.

FIG. 5 (c) is a simplified representation of the neuron 36 in FIG. 5 (b). The signal flows from the input to the output.

Fig. 5 (d) shows Rume constructed using Nieuron 36
1 shows a configuration of an lhart-type neural network 28.

In FIG. 5 (d), the output of the waveform operation mechanism 26 is input via an input switching switch 29 to an input layer 40 composed of a plurality of nylons 36. The output of the neuron 36 forming the input layer 40 is input to the intermediate layer 41 composed of a plurality of neurons, and the output of the neuron 36 of the intermediate layer 41 is input to the neuron 36 of the output layer 42. The output of the output layer 42 is
The output is output to the control operation mechanism 17 via the output switching switch 30. The neural network 28 has an input layer 40 and a middle layer 4
1, composed of an output layer 42, the output of the neuron 36 of each layer has no feedback in its own layer, and information is transmitted in a feedforward manner from the input layer 40 to the output layer 42. I have.

 The number of the intermediate layers 41 may be plural.

Next, with reference to FIG. 6, a method of determining the load w _i 37 of the neuron 36 constituting the natural network 28 will be described.

The value of the load w _i 37 of Niyuron when the Niyurarunetsuto 28 has not been learned state is undefined, usually is given by a random number. As a result, the output of the output layer 42 has a value irrelevant to the waveform input to the input layer 40. As a method of determining the value of the load w _i , for example, “Journal of the Institute of Electronics, Information and Communication Engineers”, Vol. 71, 1
No. 1, pp. 1241-1247, is described in detail in the second section, "Error Back Propagation Learning Method," in the paper "Possibility of Nyrone."

 The operation will be briefly described with reference to FIG.

The output of an input signal generating mechanism 34 for outputting a waveform to be stored in the neural network 28 is connected to an input switching switch 29.
Is input to the input layer 40 via the interface, and at that time, the data desired to be output from the output layer 42 is generated from the teacher signal generating mechanism 35,
It is input to the matching section of the learning control mechanism 32 with the output of the neural network. The butting unit changes the load on the output layer 42 so that the difference between the teacher signal and the output of the neural net 28 goes in a decreasing direction. Intermediate layer after changing load of output layer 42
Change the load of 41. In this manner, since the flow of the information flowing from the input to the output of the neural network is changed from the direction opposite to the direction of the information by the value of the load w _i , it is called back propagation learning.

FIG. 6 (b) shows a specific learning example. The shape pattern (reference shape pattern) 43-a previously stored in the input signal generating mechanism 34 is input to the input layer 40, and the output 44 corresponding to the input (operated actuator) is set to "1". , So that the other outputs are "0",
The output 46 is "1" and the others are "0".

When not learned, the output 44 does not become "1".
And the output layer 42, the intermediate layer 41,
The load of the input layer 40 is changed.

FIG. 6C shows a shape pattern (reference shape pattern) 43.
The relationship between the outputs of the teacher signal generating mechanism 35 when -b is input is shown.

FIG. 7 shows a configuration for storing a specific shape pattern in the input signal generating mechanism 34.

A command setting mechanism 49 for performing an operator operation or a programmed operation generates an operation command to the actuator 13. In accordance with this command, the actuator 13 drives the saddle 24 as shown in FIG. As a result, the rolling mill 1 changes the shape of the material 4 to be rolled. The shape data (the difference between the detected shape pattern and the target shape pattern) of the shape detector 12 that detects the shape of the material 4 to be rolled is sent from the shape detection preprocessing mechanism 15 to the pattern recognition mechanism 16 (neuro
27) is input to the waveform shaping processing mechanism 50. The waveform shaping processing mechanism 50 removes noise and the like, and can control the shape (difference between the target shape pattern and the detected shape pattern) of the material 4 to be rolled shown in FIG.
Output to Upon receiving the shape pattern shown in FIG. 7 (b), the waveform acquisition operation mechanism 51 can flatten the actuator 13 from the shape pattern (b) when the actuator 13 is operated, ie, the controllable waveform (c) shown in FIG. By operating the actuator, a reference shape pattern that can correct the shape of the material to be rolled is obtained and stored in the storage mechanism 52 of the input signal generation mechanism. This waveform (c) is one of the shape patterns (reference shape patterns) used when learning the neural net 28.

Thus, the shape waveform when the saddle 24 of the AS-U roll 21 is operated by the actuator 13-1 is stored. Since the response waveform of these waveforms changes in proportion to the operation amount of the saddle 24, the response waveform is normalized by the maximum value of the operation amount, the unit operation amount is used, the corresponding waveform pattern is set to “1”, and the other actuators are set. The output corresponding to the pattern generated by the operation is stored so as to be “0”.

On the other hand, the lateral waveform storage for moving the first intermediate roll 18 in the axial direction is performed as follows. This will be described with reference to FIG.

FIG. 8 (a) shows the shape waveform A, the controllable waveform a, and the position of the first intermediate roll 19 of the rolled material 4 when the value is "1", which is the value normalized by the maximum value of the lateral operation amount. The pattern of the controllable waveform a is input to the neural network 28, and a teacher signal is generated so that the second output 53 from the second output from the top to the output 53 becomes "1" and the other outputs become "0". The second output 54 from the top of the mechanism is set to “1” and the other output is set to “0”, and learning is performed.

Next, the normalized lateral operation 0.5 is applied to the rolling mill 1
Add to FIG. 8B shows the relationship between the position of the first intermediate roll 19, the shape pattern of the material 4 to be rolled, and the controllable waveform (reference shape pattern). As in the case of FIG. 8A, the controllable waveform (reference shape pattern) is learned by the neural network 28, but FIG.
The difference is that learning is performed so that the value of the output 54 of the teacher signal generating mechanism becomes the lateral operation amount 0.5.

Similarly, FIG. 8 (c) shows the relationship when the lateral operation amount is “0”.

In FIG. 8, the lateral operation amount is 1,0.5,0.0.
However, from the relationship with control accuracy, it is desirable to learn many points when accuracy is required and to learn a small number of cases when accuracy is not required.

This is interpolated (interpolated) by Niuro Computer 27.
It can be realized by having the role of.

Next, a shape pattern and a control method will be described with reference to FIG.

First, the shape pattern Δε is a completely flat material 4 to be rolled.
And the deviation between the output of the shape detector 12 (detected shape pattern) and the flat material to be rolled. By the way, when the output of the shape detector 12 of the rolling mill 1 is controlled to be flat, when the rolling is actually finished and the tension is applied to the material 4 to be rolled, the end portions are thin and the center portion is thick. For this reason, when the material 4 to be rolled becomes flat, the shape detector 12 detects the material 4 to be rolled with a thick end and a thin center, as shown in the target shape pattern ε _{ref in} FIG. Therefore, the rolled material 4
Needs to be controlled so that the shape of the target pattern becomes the target shape pattern ε _ref . When a control pattern Δε is obtained by subtracting the target shape pattern ε _ref from the detected shape pattern ε, and the control pattern Δε is controlled to be zero, the shape detector 12
Is the target shape pattern ε _ref . Thereby, the material 4 to be rolled when the rolling is completed becomes flat.

The shape pattern of FIG. 9 is given as an example based on this idea.

In FIG. 9, for example, in item No. 3, the detected shape pattern of A is the control pattern of a, and the control method at this time indicates that the No. 3 saddle of the AS-U roll is moved upward. .

Similarly, item No. 5 indicates that the normalization operation amount of the upper and lower first intermediate rolls 18 is moved in a direction to increase.

Next, the pattern recognition mechanism 16 and the shape detection pre-processing mechanism 15
FIG. 10 shows an example of the control operation mechanism 17 for controlling using the information obtained in step (1).

The pattern recognition mechanism 16 outputs the components of the shape pattern included in the operation waveform. When the control rule of the control operation mechanism 17 uses fuzzy inference, a fuzzy inference rule is assigned to a plurality of output lines of the pattern recognition mechanism 16. The details of the fuzzy inference are described in "Fuzzy Nyeur Computing" (published by Trikeps Co., Ltd.). If the control operation mechanism 17 performs the fuzzy inference operation, the output of the pattern recognition mechanism 16 is set to the certainty factor.

On the other hand, the output of the shape detection preprocessing mechanism 15 is directly input to the fuzzy antecedent inference mechanism 60, and the degree of conformity with the antecedent of the fuzzy rule is output to the interface 61 as a certainty factor.
The fuzzy antecedent inference mechanism 60 is a rule of qualitative inference by the operator.

The certainty factor which is the output of the pattern recognition mechanism 16 and the fuzzy antecedent inference mechanism 60 is output by the interface section 61 to the rule corresponding to each actuator 13 of the fuzzy conclusion partial synthesizing mechanism 62. An operation command is issued to the actuator 13 by the fuzzy control in which is set as an operation amount.

Next, the ZR mill control operation having the above configuration will be described with reference to FIG.

When the saddle of the AS-U roll 21 is operated, an operation waveform 70 which is a shape pattern of the material 4 to be rolled is obtained. The operation waveform 70 is input to the waveform acquisition operation mechanism 51 (FIG. 7), and a controllable waveform (reference shape pattern) 71 is obtained. The controllable waveform (reference shape pattern) 71 is stored in the euro computer 27 by the learning mechanism 31 (FIG. 5). ZR after learning
Mill 1 goes into operation. Is the difference pattern Δε of detection feature pattern epsilon and target shape pattern epsilon _ref obtained from the shape detector 12 during operation operation waveform 72 is two euro-computer 27
Is input to The neurocomputer 27 collates with the reference shape pattern stored at the time of learning, and outputs a component 73 indicating the similarity ratio to the fuzzy conclusion part synthesizing mechanism 62 of the control arithmetic mechanism 17 as a certainty factor. The fuzzy conclusion part synthesizing mechanism 62 gives an operation command to each actuator 13-1. The actuator 13-1 operates the AS-U roll 21 according to the operation command to control the material 4 to be rolled into a desired shape.

FIG. 12 shows the relationship between the input of the euro computer 27 and its output.

Δε = ε-ε control pattern (difference pattern) represented by _ref [Delta] [epsilon] is represented as a bar graph (a), the output value of 35 pieces of the shape detector 12 installed in the plate width direction is a value of a bar graph. Fig. 12 (b)
Indicates a case where eight reference shape patterns are stored in the Euro Euro computer 27 as the Euro storage patterns 1 to 8. When the control pattern Δε is input to the neuro-computer 27, the value of each output of the neural network 28 is obtained. In FIG. 12 (c), these output values (in the technical field of neuron, generation of an output is called firing of an output layer) is represented by a square area. The larger the area is, the higher the matching degree (similarity ratio) is. These patterns are associated with the actuator 13, and for example, the pattern 1 and the AS-U roll 21 are associated with each other.
The No. 2 roll is made to correspond to the pattern 7 and the lateral shift of the upper roll of the first intermediate roll 18 (FIGS. 12 (d) and 12 (e)).

As a result, the AS-U roll operation amount (a) (FIG. 12 (f)) and the upper lateral shift amount (b) (FIG. 12 (g)) of the first intermediate roll 18 corresponded to the above area. An amount of manipulated variable is obtained. FIGS. 12 (f) and 12 (g) show the up operation command of the No. 2 roll of the AS-U roll 21 and the first intermediate roll.
18 shows a state in which an instruction to shift the upper roll 18 to the right is issued.

When such control is applied for a certain period of time, a simulation result in which the change for each sampling time is displayed three-dimensionally is displayed as the second
Figure 13 shows.

FIG. 13 (a) shows a control shape pattern (Δε), which is the difference between the detected shape pattern ε and the target shape pattern ε _ref , in FIG.
FIG. 2B shows the detected shape pattern (ε). The detected shape pattern that has deteriorated at time t = 0 changes with time, and the detected shape pattern ε becomes the target shape pattern,
It can be seen that the control shape pattern Δε approaches flat.

〔The invention's effect〕

According to the present invention, a waveform component that can be corrected by control is extracted from the entire shape of the material to be rolled using a neural network, and control can be performed using the extraction result, so that a good shape of the material to be rolled can be obtained. it can.

[Brief description of the drawings]

FIG. 1 is an overall configuration diagram of one embodiment of the present invention, and FIG.
FIG. 3 is a configuration diagram of a rolling mill, FIG. 3 is a diagram showing a rolling mill actuator and its operation waveform, FIG. 4 is a configuration diagram of a control device according to an embodiment of the present invention, FIG. 6 is an explanatory diagram of a learning example, FIG. 7 is an explanatory diagram of a shape pattern storage method, FIG. 8 is an explanatory diagram of a lateral operation of a rolling mill, FIG. 9 is a diagram showing a relationship between a shape pattern and control, FIG. FIG. 11 is a diagram showing the detailed configuration of the control mechanism, FIG. 11 is an explanatory diagram of the operation of one embodiment of the present invention, FIG. 12 is a diagram showing a simulation result of a neurocomputer, and FIG. 13 is a simulation of one embodiment of the present invention. It is a characteristic view showing a result. 1 ... Sendimer Mill, 12 ... Shape Detector, 16 ... Pattern Recognition Mechanism, 17 ... Control Operation Mechanism, 28 ... Neural Network.

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Hiroshi Saito 5-2-1 Omika-cho, Hitachi City, Ibaraki Prefecture Inside the Omika Plant, Hitachi, Ltd. (72) Inventor Yoki Katayama 4026 Kuji-cho, Hitachi City, Ibaraki Shares (72) Inventor Yasuo Morooka 4026 Kuji-cho, Hitachi City, Ibaraki Prefecture Within Hitachi Research Laboratory, Hitachi, Ltd.

Claims (6)

(57) [Claims] The present invention further comprises a plurality of actuators for executing the operation of a multi-roll Sendzimer mill, wherein the shape of a material to be rolled is detected by a shape detector arranged on the exit side of the Sendzimer mill, and the actuator is operated to activate the rolled material. In a shape control method of a Sendzimir mill for controlling a shape of a material, a neural network for obtaining an operation amount of the actuator based on a difference between a detected shape pattern detected by the shape detector and a target shape pattern, and a similarity ratio with a reference shape pattern described later. Provided, at the time of learning of the neural network, to obtain a reference shape pattern that can make the material to be rolled into the target shape pattern by operating the actuator from the difference between the detected shape pattern and the target pattern, input to the neural network, The neural network Over clause and the reference shape pattern,
A method for controlling a shape of a Sendzimer mill, wherein learning is performed on a relationship between the actuator and the amount of operation of the actuator to form the reference shape pattern. 2. The apparatus according to claim 1, further comprising a plurality of actuators for operating a multi-roll Sendzimer mill, detecting a shape of the material to be rolled by a shape detector arranged on an output side of the Sendzimer mill, and operating the actuator to operate the rolled material. In the shape control method of a Sendzimir mill for controlling the shape of a material, a difference between a detected shape pattern detected by the shape detector and a target shape pattern, and a neural network for calculating a similarity ratio with a reference shape pattern described later, and the neural network A fuzzy inference mechanism for inputting the obtained similarity ratio and obtaining the operation amount of the actuator according to the degree of the similarity ratio, and operating the actuator from the difference between the detected shape pattern and the target shape pattern when learning the neural network; By rolling the material to be rolled A reference shape pattern that can be a mark shape pattern is obtained and input to the neural network, and the neural network learns the relationship between the reference shape pattern and the amount of operation of the actuator for forming the reference shape pattern. And a method for controlling the shape of a Sendzimer mill. 3. A plurality of actuators for performing a vertical operation of a saddle of a Sendzimer mill and a shift operation of an intermediate roll, wherein the shape of a material to be rolled is detected by detecting a shape of a material to be rolled by a shape detector arranged on an output side of the Sendzimer mill. In the shape control method of the Sendzimir mill that operates an actuator to control the shape of the material to be rolled, a difference between a detected shape pattern detected by the shape detector and a target shape pattern,
A neural network for obtaining the operation amount of the actuator based on a similarity ratio with a reference shape pattern described later is provided, and the learning material learns the neural network. A reference shape pattern that can be a target shape pattern is obtained and input to the neural network, and the neural network learns the relationship between the reference shape pattern and the amount of operation of the actuator to form the reference shape pattern. And a method for controlling the shape of a Sendzimer mill. 4. A multi-roll Sendzimer mill for rolling a material to be rolled, a plurality of actuators for operating the Sendzimer mill, a shape detecting means for detecting the shape of the material to be rolled by the Sendzimer mill, and the shape detecting means. By inputting the difference between the detected shape pattern detected in the above and the target shape pattern, the actuator is actuated to obtain a reference shape pattern capable of turning the material to be rolled into the target shape pattern, and the reference shape pattern and the reference shape Learning the relationship between the amount of operation of the actuator to form a pattern, the difference between the detected shape pattern and the target shape pattern actually detected during the actual operation of the Sendzimer mill, the reference shape pattern Neuro that calculates the operation amount of the actuator by a similarity ratio with Shape control apparatus for Senjimamiru equipped a computer. 5. A multi-roll Sendzimer mill for rolling a material to be rolled, a plurality of actuators for operating said Sendzimer mill, shape detecting means for detecting the shape of the material to be rolled by said Sendzimer mill, and said shape detection. Learning control means for inputting a difference between the detected shape pattern detected by the means and the target shape pattern, and operating the actuator to obtain a reference shape pattern capable of turning the material to be rolled into the target shape pattern; and Learning about the relationship between the pattern and the amount of operation of the actuator to form the reference shape pattern, the difference between the detected shape pattern and the target shape pattern during the actual operation of the Sendzimer mill, and the reference shape The amount of operation of the actuator is calculated based on the similarity ratio with the pattern. Shape control apparatus for Senjimamiru equipped and Rarunetto. 6. A multi-roll Sendjimer mill for rolling a material to be rolled, a plurality of actuators for operating said Sendzimer mill, a shape detecting means for detecting the shape of the material to be rolled by said Sendzimer mill, and said shape detecting means. By inputting the difference between the detected shape pattern detected in the above and the target shape pattern, the actuator is actuated to obtain a reference shape pattern capable of turning the material to be rolled into the target shape pattern, and the reference shape pattern and the reference shape Learning the relationship between the amount of operation of the actuator to form a pattern, the difference between the detected shape pattern and the target shape pattern actually detected during the actual operation of the Sendzimer mill, the reference shape pattern A neurocomputer for determining the similarity to Enter the similarity rate calculated in over data, wherein the degree of similarity ratio, shape control apparatus Senjimamiru which and a control arithmetic unit having a fuzzy inference engine for determining an operation amount of the actuator.