JPH04167908A - Device for setting rolling mill - Google Patents

Abstract

PURPOSE:To decide and set the optimum setting value at the initial stage in operating terminal for controlling flatness by executing learning of synapse load factor in neural network at any time while using the optimum setting value in the operating terminal obtd. from the actual rolling data. CONSTITUTION:Inputted data of influenced factor to the flatness of the rolled material 1 of steel kind of the rolled material 1, plate width, inlet thickness, outlet thickness, predicted rolled load, accumulated rolled length after changing work roll, etc., are converted into the suitable form to the inputted data in the neural network 9. Based on this inputted data, the setting value at the initial stage in the operating terminal for controlling the flatness in the rolling mill, is operated and the operated output is converted into the suitable form to setting of the operating terminal for controlling the flatness. Further, the optimum setting value in the operating terminal for controlling the flatness, which the actual flatness value in the rolled material with the rolling mill after starting rolling of the rolling material 1 becomes the aimed flatness is operated and this optimum setting value and the inputted data are used as the teacher data to learn the synapse load factor in the neural network 9.

Description

Detailed Description of the Invention [Purpose of the Invention (Industrial Field of Application) The present invention relates to a rolling mill for rolling rolled materials such as steel plates, and in particular to a flatness control operating end for controlling the flatness of the rolled material. The present invention relates to a setting device for a rolling mill that can easily determine and set initial setting values based on rolling results.

(Prior Art) In recent years, rolling mills suitable for controlling the flatness of rolled materials have been successively proposed and put into practical use.

One of them is, for example, a six-layer rolling mill shown in FIG. In FIG. 5, 1 is a rolled material, 2 is a backup roll, 3 is a work roll, and 4 is an intermediate roll.

Further, the operating ends for flatness control in this six-layer rolling mill include the work roll bender 5, the intermediate roll bender 6, the shift d1 of the intermediate roll 4 in the roll axis direction, and the roll leveling mechanism.

In such a six-layer rolling mill, in order to obtain good flatness, it is necessary to skillfully operate and control these operating ends during rolling of the rolled material 1, but also to control these operating ends at the start of rolling. You can also set the initial setting value of
This is very important.

By the way, as a prior art of such an initial setting method, there is, for example, "Japanese Patent Application Laid-Open No. 55-68110". (Flatness) Calculate the correction coefficient of the parameter model, use this model correction coefficient to determine the shape parameter of the next rolled material, and use the shape control operation end for the next rolled material so that this shape parameter approaches the target value. This is a method of determining and setting the intermediate roll position and roll bending pressure.

On the other hand, as other conventional techniques, for example,
-34883. This is the shape (flatness) determined from the actual rolling values and rolling conditions at the end of rolling of the preceding rolled material.
Calculate the shape parameters to evaluate the shape parameters, calculate the intermediate roll position and the optimal value of the roll bending pressure so that the shape parameters meet the target values, and apply these values to the relationship between steel type - plate width - plate thickness or rolling reduction rate. This is a method of storing the values separately and indexing and setting the stored values for the next rolled material. All of these methods are characterized in that the actual rolling values of the preceding rolled material are reflected in the determination of the initial setting values of the next rolled material. Further, in the latter case, the method of reflection is not direct, but is performed via a table of intermediate roll positions and bending pressure setting values.

By the way, it is very difficult to theoretically determine the initial setting value of each operating end to obtain the target flatness.

Therefore, as described above, methods of estimating from actual rolling values of previously rolled materials or indexing from a set value table are often used.

However, if the rolling conditions of the previously rolled material and the next rolled material are significantly different, it is very difficult to estimate the optimal initial setting values from the actual rolling values of the previously rolled material. Furthermore, when indexing the setting value table, the configuration of the setting value table affects the accuracy of the initial settings. For example, in "Special Publication No. 34883/1983", it is specified that the table classification should be steel type - plate width - plate thickness or rolling reduction, but whether or not these elements are appropriate, The question is how to classify each category. In other words, if the classification is rough, it becomes difficult to set the optimum value.

On the other hand, if the table is divided into smaller sections, the table size becomes larger and maintenance thereof requires a great deal of effort.

(Problems to be Solved by the Invention) As described above, in the conventional setting method of a rolling mill, it is not only difficult to determine the optimum initial setting value of the operating end for controlling the degree of shock, but also There was a problem in that it required a lot of effort to maintain because it required creating models and tables.

An object of the present invention is to provide a rolling method that allows easy determination and setting of optimal initial setting values for a flatness control operating end based on rolling results without the need to create a mathematical model or create a table. The purpose of this invention is to provide a machine setting device.

[Structure of the Invention] (Means for Solving the Problems) In order to achieve the above object, the present invention provides a setting device for a rolling mill that is provided with an operating end for controlling the flatness of the rolled material and that rolls the rolled material. , A flatness detection means installed on the exit side of the rolling mill to detect the flatness of the rolled material, and information on the steel type, plate width, input thickness, exit thickness, predicted rolling load, and cumulative rolling after work roll rearrangement of the rolled material. input data processing means that takes as input data factors that influence the flatness of the rolled material, such as material length, and converts the input data into a form suitable for input data of the neural network; and based on the input data from the input data processing means. , a neural network that calculates the initial setting value of the operating end for flatness control of the rolling mill, an output data processing means that converts the output from the neural network into a form suitable for setting the operating end for flatness control, and a rolled material. After the start of rolling, the flatness actual value detected by the flatness detecting means becomes the target flatness.The optimum setting value of the flatness control operating end is calculated by the optimum setting value calculating means and the optimum setting value calculating means. a data storage means for storing the optimal setting values and input data to the input data processing means; and a learning term calculation means for learning synapse weight coefficients of the neural network using the data stored in the data storage means as training data. It is composed of:

(Function) Therefore, in the rolling mill setting device according to the present invention, there is no need to create a mathematical model for determining operating end setting values or to create a table for indexing operating end setting values, and it is not necessary to create a table for indexing operating end setting values. You can obtain the initial setting value. In addition, by learning the synaptic load coefficients of the neural network at any time using the optimal setting values of the operating end obtained from the rolling performance data, changes over time in the rolling state are reflected, and the optimal solution can always be obtained. Thereby, a product with good flatness can be rolled.

(Example) In the present invention, factors influencing the flatness of the rolled material, such as the steel type of the rolled material, the plate width, the plate thickness, the predicted rolling load, and the length of the rolled material after work roll rearrangement, are used as input data to create a neural network. The initial setting value of the flatness control operating end is obtained as an output from this neural network.

In other words, it goes without saying that the neural network is trained in advance using training data before application, but in the present invention, the optimal operating end setting value is calculated from the actual rolling values, and this is added to the training data and Update the neural network by performing learning. This is a feature of improving the accuracy of the output of the neural network.

Here, we will briefly explain neural networks.

A neural network, as shown in FIG. 2, is a circuit system consisting of neurons and links connecting the neurons, and each neuron receives multiple input signals x1 to x1 as shown in FIG. One output signal y is output from X9.

The relationship between y and x1 to x is expressed by the following equation (1). That is, the input signal to the neuron X, -
The output is determined by the linear combination phase of X, load coefficients w1 to w4 called synaptic loads, and the threshold value θ.

y−f(Σ This learning function is
This is performed using a desired output signal (hereinafter referred to as a teacher output signal) for a certain input signal.

The back propagation method, which is a typical learning method, will be explained below.

This backpropagation method works in a multilayer neural network by moving backwards from the output layer to the previous stage so as to minimize the double error at the output layer, that is, the square of the difference between the teacher output signal and the neural network output signal. In this method, the synapse weight coefficients are sequentially corrected.

Here, in the m-layer neural network, the ``th output Y+'' of the th layer is (2
) is expressed by the formula.

y,'-f(Σ W IIX H'-'l +θ k) ... (2) An input signal in the input layer is given, and the output of the j-th unit in the output layer (m-th layer) at this time is yj′″, and the teacher output signal is Y+′
'', the loss function r is expressed by equation (3).

r-(1/2) Σ (y+--y+")'...(
3) Also, the update of W that minimizes this loss function r is (4)
It is held in a ceremony.

w 1+ (NE W ) −w , + (OLD
) + ΔW1゜... (4) ΔW, 1-77"δ%: +"-' - (5) δ,
'-f' (Σ W, · k-1+6%) Σ δ k+
l・W.

(6) δ,'''-tY+'')'+-) (7) Here, η is a constant, and f' represents the differential of the function f.

Generally, the more teaching signals are used for learning, the more accurate the output of the neural network will be.

Hereinafter, an embodiment of the present invention based on the above concept will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing an example of the overall configuration of a setting device for a rolling mill according to the present invention. In addition, here, in a six-high rolling mill composed of a pair of backup rolls 2, a work roll 3, and an intermediate roll 4, a work roll bender 5 and an intermediate roll bender, which are operating ends for controlling the flatness of the rolled material 1, are used. 6 and intermediate roll 4 are determined.

That is, the rolling mill setting device according to the present embodiment includes a flatness meter 7 as flatness detection means, an input signal processing section 8, a neural network 9, an output signal processing section 10, and an operating end control device 11. , the optimum setting value calculation device 12, the teacher data storage section 13, the comparator 14, and the learning term calculation section 15.
It consists of.

Here, the flatness meter 7 is provided on the exit side of the rolling mill to detect the flatness of the rolled material 1. In addition, the input signal processing unit 8 also processes the steel type of the rolled material 1, the plate width B, the thickness H1, the thickness h,
Factors that influence the flatness of the rolled material 1, such as the predicted rolling load P and the cumulative rolled material length after work roll rearrangement, are used as input data, and the input data is sent to the neural network 9.
It converts the input data into a format suitable for the input data. Further, the neural network 9 calculates the initial setting value of the flatness control operating end of the rolling mill based on the input data from the input signal processing section 8. Furthermore, the output signal processing unit 10 receives the output from the neural network 9 and converts it into a form suitable for setting the flatness control operating end.

On the other hand, the operating end control device 11 controls the flatness control operating end based on output data from the output signal processing section 10. Further, the optimum setting value calculation device 12 determines the optimum setting value of the operating end for flatness control so that the actual flatness value detected by the flatness meter 7 becomes the target flatness after the rolling load detector 17 starts rolling the rolled material 1. This is to calculate the set value.

Further, the teacher data storage section 13 stores the optimum setting values calculated by the optimum setting value calculation device 12 and the input data to the input signal processing section 8.

Further, the comparator 14 uses the data stored in the teacher data storage section 13 as teacher data, and compares this with the output data from the output signal processing section 10. moreover,
The learning term calculation unit 15 is for learning the neural network weight coefficients of the neural network 9 based on the comparison data from the comparator 14.

Next, the operation of the rolling mill setting device according to this embodiment configured as described above will be explained.

In FIG. 1, the input signal processing section 8 receives as its input signals the steel type, the plate width B of the rolled material 1, the input thickness H1 the output thickness h1, the predicted rolling load P, and the cumulative rolling load after work roll rearrangement. The material length is input.

In the input signal processing unit 8, these input signals are converted into a form that can be easily handled as input signals for the neural network 9, and are input to the neural network 9. for example,
Regarding steel types, a large number of steel types are classified into several categories, for example, 10 categories, and a signal of "1" is output only for the corresponding category, and '0'' is output for the others.

In addition, in the case of an analog quantity such as plate width B, it is classified into necessary and sufficient categories, and as described above, only the signal of the corresponding category is outputted as "1", and the others are outputted as "0". Ru. For example, the plate width range of the rolled material 1 is 600 mm to 2000 mm.
In the case of u+, 100 square divisions are sufficient for determining the initial setting value, so only the 100 mm digit and 1000 mm digit are valid, and the 100 square digit 10 divisions and 1000 square division are valid.
It is classified into +1m digit 2 classification. That is, for example, if the plate width B is 1300 l, the signals of 1 in the digit of 1000 l and the signals in the digit of 100+u are outputted as "11", and the others are outputted as "0°".

Further, similar conversion is performed on the predicted rolling load P and the cumulative length of the rolled material, which are provided as input signals to the neural network 9.

Next, the neural network 9 calculates the initial setting value of the flatness control operation end based on these input signals, and the result is input to the output signal processing section 10. Further, in the output signal processing section 10, the neural network 9
The output from the input signal processing section 8 is converted in a manner opposite to that of the input signal processing section 8, and the work roll bender setting value Fw, intermediate roll bender setting value F1, and intermediate roll position setting value d are converted to the operating end control device 11. are input into each.

Further, in the operating end control device 11, each operating end is set to the set value from the output signal processing section 10 before starting rolling of the rolled material 1.

The above is the explanation for determining the initial setting value.

Next, learning of synapse weight coefficients by the neural network 9 will be explained.

The synaptic weight coefficients are learned using previously accumulated training data before online application.

However, not all rolling ranges can be learned before online. Naturally, it is also possible to extend or change the rolling conditions.

Therefore, in this embodiment, the synapse load coefficient is updated by calculating the optimal setting value for each operating end from the rolling performance data, adding this to the teacher data, and executing learning as needed.

That is, after the rolling of the rolled material 1 is started, at the timing when the initial setting target position on the rolled material 1 is rolled, the actual values FwA of the work roll bender setting value, intermediate roll bender setting value, and intermediate roll position setting value at that time are determined. ,F+^,d
^ is from the operating end control device 11 to the optimum setting value calculation device 12
is input.

At the same time, the rolling load actual value pA detected by the rolling load detector 16 is input to the optimum setting value calculation device 12. Further, when the initial setting target position reaches the flatness needle 7 provided on the exit side of the rolling mill, the output from the flatness needle 7 is inputted to the optimum set value calculation device 12 as the actual flatness value βIA. .

Next, the optimum setting value calculation device 12 calculates the optimum setting value of each operating end, for example, by the following method.

Figure 4 shows the actual flatness value β1^ and the target flatness β, RE
It is a figure which shows an example of the relationship with P. In Figure 4, 1
represents the position in the board width direction, which usually coincides with the position detected by the flatness needle 7.

Here, the flatness deviation ε1 is expressed by equation (8).

ε・−β・A−β・REF ・(8) Then, the work roll bender setting value correction amount ΔFW, the intermediate roll bender setting value correction amount ΔFl, and the intermediate roll that minimize the evaluation function J shown in equation (9) A position correction amount Δd is determined.

J-(1/2)fεI-(gW,・ΔFw+ g z・
ΔF 1+ g a+・Δd))2 Here, gwl, g
liSg and + are coefficients of influence on the flatness of the operating end, and can be determined using a simple formula or the like.

In this case, ΔFW minimizes equation (9).

Methods for calculating ΔF+ and Δd include the least squares method, which is a known technique.

From the above, the optimal setting value Fw0, FIOldo for each operating end
is found by .

Next, the optimum setting value calculating device 12 outputs the F VMOs Floq d O and the actual rolling load value P^ obtained by the above equation (10) to the teacher data storage unit 13 and stores them therein. At the same time, the input data for the rolled material 1, such as the steel type, plate width B, input thickness H1, output thickness h1, and cumulative rolled material length, are also input to the teacher data storage section 13 and stored in pairs with the above data. Learning is then performed when a certain amount of new training data is added.

On the other hand, at this time, the steel type, plate width B1 thickness H1 thickness h, actual rolling load value P^, and cumulative rolled material length stored in the teacher data storage section 13 are input to the input signal processing section 8 as teacher input data. is input. Also, neural network 9
Then, a calculation is performed based on this input signal, and the result of the calculation is input to the output signal processing section 10. Furthermore, the work roll bender setting value FW, intermediate roll bender setting value Fl, and intermediate roll position setting value d are input from the output signal processing section 10 to the comparator 14.

On the other hand, the comparator 14 calculates the difference between the teacher input data outputted from the teacher data storage section 13 and Fwo-F 1o1d□, which is the teacher output data stored in pairs, and sends it to the learning term calculation section 15. is input. Further, in the learning term calculation unit 15, the synapse weight coefficient is updated by the learning method described above. This learning calculation is performed using all the teacher data stored in the teacher data storage section 13.

As described above, in this embodiment, the work roll bender 5, the intermediate roll bender 6, the shift d1 in the roll axis direction of the intermediate roll 4, and the roll leveling mechanism are used to control the flatness of the rolled material 1. A flatness needle 7, which is provided on the exit side of the rolling mill and is a flatness detection means for detecting the flatness of the rolled material 1, is equipped with an operating end for setting a six-fold rolling mill for rolling the rolled material 1. Input data includes factors that influence the flatness of the rolled material 1, such as the steel type of the rolled material 1, the plate width B, the input thickness H1, the outgoing thickness h, the predicted rolling load P, and the cumulative length of the rolled material after work roll rearrangement. , an input signal processing unit 8 that converts the input data into a form suitable for the input data of the neural network, and an initial setting of the operating end for flatness control of the rolling mill based on the data input from the input signal processing unit 8. The neural network 9 that calculates values and the output data from the neural network 9 are in a form suitable for setting the Heiwa i control operation end! ! ! Optimum settings for the output signal processing unit 10 that converts into FW, Fl, and d, and the operating end for flatness control such that the actual flatness value βIA detected by the flatness needle 7 after the start of rolling of the rolled material 1 becomes the target flatness. value Fwo,
The optimum setting value calculating device 12 calculates F + o, do, PA, and the optimum setting values FWO, F! calculated by the optimum setting value calculating device 12. 01d0. PA, and a teacher data storage unit 13 that stores input data to the input signal processing unit 8 (steel type, plate width B, input thickness H1, output thickness h1, predicted rolling load P, cumulative rolled material length), and a teacher data storage unit. 13
The learning term calculation unit 15 uses the data stored in the neural network 9 as training data, and learns the synapse weight coefficient of the neural network 9 based on the data obtained by comparing the training data with the output data from the output signal processing unit 10. It has been prepared and configured.

Therefore, there is no need to create a mathematical model to determine operating end setting values or create a table to index operating end setting values as in the past, and the optimum initial setting values can be obtained based on rolling results. becomes possible. In addition, the learning of the synaptic load coefficient of the neural network 9 is performed using the optimum setting value FW of the operating end for flatness control obtained from the rolling performance data.
O+ F l01d0. Since the PA is used to make the rolling process happen at any time, changes over time in the rolling state of the rolled material 1 can be reflected, making it possible to always obtain the optimum solution.

This makes it possible to roll a product with extremely good flatness.

It should be noted that the present invention is not limited to the above-mentioned embodiments, but can be similarly implemented in the following manner.

(a) In the above example, the case was described in which the steel type, plate width B1 thickness in place 1 thickness h1 expected rolling load P1 cumulative length of rolled material were used as input data. It varies depending on the type etc.

For example, in a rolling mill in which the flatness target value is changed in various ways, the target flatness pattern is added as input data.

Moreover, when the intermediate roll position d is determined in advance by another algorithm, the intermediate roll position d also becomes input data.

Furthermore, in a rolling mill that uses work rolls with different surface roughnesses, data representing the surface roughness of the rolls is also input data. However, these do not affect the essence of the present invention.

(b) In the above embodiment, the case was described in which only the rolling load was used as the actual data to be stored as the teacher data, but the sheet width B1, the input thickness, the outgoing thickness, h1, and the cumulative rolled material length are also measurable values. These performance values can also be stored in the teacher data storage section 13 as teacher data.

(c) In the above embodiment, the present invention was applied to a 6-height rolling mill, but the present invention can also be applied to other rolling mills.

[Effects of the Invention] As explained above, according to the present invention, the flatness of the rolled material, such as the steel type, plate width, entry thickness, exit thickness, predicted rolling load, cumulative rolled material length after work roll rearrangement, etc. Convert the input data into a form suitable for the input data of the neural network, calculate the initial setting value of the operating end for flatness control of the rolling mill based on this input data, and calculate the calculation output in parentheses. Convert to a form suitable for setting the operating end for flatness control, and optimally set the operating end for flatness control so that the actual flatness value of the rolled material on the exit side of the rolling mill becomes the target flatness after rolling of the rolled material starts. The synapse load coefficients of the neural network are learned by calculating the values and using these optimal setting values and input data as training data, so it is possible to learn the synaptic load coefficients of the neural network based on the rolling results without the need to create a mathematical model or create a table. It is possible to provide a setting device for a rolling mill that can easily determine and set the optimum initial setting value of the flatness control operating end.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of a setting device for a rolling mill according to the present invention; FIG. 2 is a diagram showing an example of the configuration of a neural network; FIG. 3 is a diagram showing an example of the configuration of neurons in the neural network; FIG. 4 is a diagram showing an example of a flatness pattern, and FIG. 5 is a schematic diagram showing an example of the configuration of a six-high rolling mill. DESCRIPTION OF SYMBOLS 1...Rolled material, 2...Backup roll, 3...Work roll, 4...Intermediate roll, 5...Work roll bender, 6...Intermediate roll bender, 7...Flat deflection, 8... - Input signal processing unit, 9... Neural network, 10... Output signal processing unit, 11... Operating end control device, 12... Optimal setting value calculation device, 13. Teacher data storage unit, 14... Comparator, 15...Learning term calculation unit, 16...Rolling load detector. Applicant's agent Patent attorney Takehiko Suzuhan Input signals Figure 2 Figure 3 Figure 1; Figure 4

Claims (1)

[Scope of Claims] A setting device for a rolling mill comprising an operating end for controlling the flatness of the rolled material and configured to roll the rolled material, the setting device being provided on the outlet side of the rolling mill and configured to control the flatness of the rolled material. flatness detection means for detecting the flatness of the rolled material, the steel type, plate width, entry thickness, exit thickness, predicted rolling load,
an input data processing means that takes as input data an influence factor on the flatness of the rolled material, such as the cumulative rolled material length after work roll rearrangement, and converts the input data into a form suitable for input data of a neural network; Based on the input data from the processing means,
a neural network that calculates an initial setting value of the flatness control operating end of the rolling mill; and an output data processing means that converts the output from the neural network into a form suitable for setting the flatness control operating end. an optimum setting value calculation means for calculating an optimum setting value of the flatness control operation end such that the actual flatness value detected by the flatness detection means after the start of rolling of the rolled material becomes the target flatness; and the optimum setting. The optimal setting value calculated by the value calculation means,
and data storage means for storing input data to the input data processing means; learning term calculation means for learning synapse weight coefficients of the neural network using the stored data stored in the data storage means as training data; A setting device for a rolling mill, comprising: