JP3308630B2 - Plant control apparatus and method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a rolling mill shape control device and method for performing control in accordance with detected shape data.

[0002]

2. Description of the Related Art An example of applying a neural network to shape control of a rolling mill is disclosed in Japanese Patent Application Laid-Open No. 4-127908, entitled "Method and Control Apparatus for Rolling Mill." In this case, the neural network captures the detected widthwise shape distribution of the rolled material as a shape pattern, and extracts a widthwise shape pattern of a typical rolled material as a correctable shape pattern from the detected shape pattern. have. However, the learning of the neural network performed there is, for example, as described in Chapter 2 “Backpropagation” of the first volume of the Neurocomputer (Supervised by Hajime Nakano). In this method, the number of units is set, the initial value of the weighting factor between neurons, which is a parameter inside the network, is set by a random number, and the weighting factor is corrected by a learning rule based on the steepest gradient method or the like. Therefore, in order to determine the neural network, it was necessary to repeatedly perform trial and error to change the number of hidden neurons and to change the initial value of the weight coefficient of the neural network.

[0003]

Since the neural network has many local optimum points that satisfy the relationship between the learning data, the correction of the weighting factor based on the steepest slope method requires the initial value of the weighting factor to be changed. Since it converges on the load coefficient of the dependent local optimum point, it does not always converge on the optimum load coefficient. However, the shape pattern in the shape control of the rolling mill has a feature that the shape pattern can be approximately expressed by a linear sum of a plurality of basic element patterns because the size itself of the shape pattern is very small. However, in the above-mentioned prior art, since such knowledge on the relationship between such a shape pattern and a basic element pattern is not effectively used, especially when there is little data for learning, the detected rolled material shape is There was a problem that a basic element pattern could not be successfully extracted from a pattern.

An object of the present invention is to realize a relation between data by correcting a weighting factor of a neural network by a learning rule based on a steepest gradient method such as back propagation, which is typical of a typical rolled material set in advance. By using the approximate linear relationship between the basic element pattern expressing a simple shape pattern and the detected shape pattern, an optimal neural network that can extract the basic element pattern from the shape pattern even when the learning data is small is realized. It is an object of the present invention to provide an apparatus and a method.

[0005]

In order to achieve the above object, the present invention provides a basic element pattern storage section comprising a typical shape pattern in the width direction of a rolled material, and regression of a relationship between the shape pattern and the basic element pattern. A function approximation part that is approximated by analysis, a weight coefficient setting part that determines the initial value of the weight coefficient of the neural network using the parameters obtained by the function approximation part, the detected shape pattern and the basic element pattern included in the shape pattern Using the results of the learning data storage unit and the function approximation unit consisting of the component ratios, the approximation degree determination unit that selects data from the learning data storage unit, and correcting the weighting factor of the neural network by the approximation degree determination unit And a correction learning data storage unit storing the above data.

[0006]

The linear relation between the detected width-direction shape pattern of the rolled material and the basic element pattern is approximated by the function approximation unit using the data of the basic element pattern storage unit comprising the typical shape pattern of the rolled material. The weight coefficient setting unit sets the parameter of the approximation function represented by the function approximation unit as an initial value of the weight coefficient of the neural network. Thereby, the neural network is learned using the shape pattern of the correction learning data storage unit selected by the approximation degree determination unit.

[0007]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below in which the embodiment of the present invention is applied to the control of the thickness distribution (shape) in the width direction of a rolled material of the multi-roll mill shown in FIG. The thickness distribution of the rolled material in the width direction is detected at multiple points by a shape detector 12 including a plurality of sensors, and the envelope of the detected amount detected at the multiple points is defined as the shape pattern of the rolled material (hereinafter referred to as a shape pattern). ). At this time, the output value of each sensor is input to each unit of the input layer of the neural network 7. And
The neural network 7 obtains the component amount of each basic element pattern (basic shape pattern specific to the target for which the operation rule is determined) included in the input shape pattern of the rolled material. Control by obtaining the amount of operation to The neural network 7 in this control configuration has a role of extracting a preset basic element pattern component from the detected shape pattern of the rolled material. It is assumed that the shape pattern of the rolled material detected at this time can be approximated by a linear sum of basic element patterns. Learning of a neural network using the present invention is performed as follows.

First, the basic material pattern storage unit 8 extracts and stores a specific thickness distribution pattern as the basic element pattern from the shape pattern of the rolled material detected by the shape detector 12. These basic element patterns are sent to the function approximation unit 2. Since the shape pattern of the rolled material is represented by a linear sum of the basic element patterns, the function approximation unit 2 approximates the relationship between the shape pattern and the basic element pattern by a regression approximation model described later.

The function approximation unit 2 outputs the approximated parameters of the regression approximation model to the weighting factor setting unit 3, and the weighting factor setting unit 3 converts the relationship obtained by the regression approximation model into a weighting factor corresponding to the neural network 7. Convert to

On the other hand, with respect to the shape pattern of the rolled material detected by the shape detector 12, the operator inputs from the input unit 13 the component ratio of each basic element pattern stored in the basic element pattern storage unit 8, The data is stored in the learning data storage unit 1 together with the shape pattern. The approximation degree determination unit 4 estimates the component ratio of each basic element pattern stored in the learning data storage unit 1 using the regression approximation model sent from the function approximation unit 2. When the difference between this component ratio and the component ratio of each basic element pattern set in the learning data storage unit 1 in advance is equal to or larger than the determination reference value, the shape pattern is used as correction learning data for the neural network 7 for correction learning. The data is stored in the data storage unit 5.

The weight coefficient corresponding to the neural network 7 obtained by the weight coefficient setting unit 3 is output to the learning control unit 6. At this time, if the shape pattern of the rolled material is completely represented by the linear sum of the respective basic element patterns, the ratio of the respective basic element patterns can be output. If it cannot be expressed only by the linear sum of the basic element patterns, the learning control unit 6 uses the weighting factor sent from the weighting factor setting unit 3 as an initial value and sets the learning data storage unit 1 and the correction learning data storage unit. 5
Learning by the steepest slope method such as back propagation is performed using the shape pattern of the rolled material sent from the company, and the weight coefficient of the neural network 7 is corrected.

FIG. 6 shows an example of the learning data storage 1. The shape distribution in the width direction of the rolled material from the detector 12 is sent to the learning data storage unit 1 as a shape pattern. Also,
The component ratio of the basic element pattern existing in the detected shape pattern is input from the input unit 13 to the learning data storage unit 1. The input component ratio is sent to the basic element pattern component ratio storage unit 102, and the shape pattern is stored in the shape pattern storage unit 101 in association with the component ratio of the basic element pattern in the shape pattern.

FIG. 2 shows an example of the function approximation unit 2. The basic element pattern is input from the basic element pattern storage unit 8 to the function approximation unit 2. Inside the function approximation unit 2, the basic element pattern is sent to the regression approximation model 202, and the structure of the approximate model is determined from the order when the basic element pattern is displayed as a vector and the number of basic element patterns. The basic element pattern is also input to the model parameter setting unit 201. Inside the model parameter setting unit 201,
The parameters of the regression approximation model 202 are set as described below.

Assuming that the shape pattern of the rolled material is represented by a linear sum of the basic element patterns in the basic element pattern storage unit 8, the output value X of the shape pattern of the rolled material by m sensors and the output value X included in the shape pattern The approximate linear relationship with the ratio Y of each of the n basic element patterns is expressed by a regression model shown in Expression 1.

[0015]

(Equation 1)

Here, X1 to Xm are the shape patterns of the rolled material, and Xi is the output value of the i-th sensor distributed in the width direction. Y1 to Yn are the ratios of the basic element patterns included in the shape pattern, and Yi is the ratio of the i-th basic element pattern. Further, nv _{i1 to} nv _im are the i-th basic element patterns, and i is a natural number from 1 to n.

[0017]

(Equation 2)

[0018]

(Equation 3)

[0019]

## EQU00004 ## w _ik = nv _ik -p _i (Equation 4)

[0020]

(Equation 5)

Here, Pi is the average of the distribution values of each basic element pattern.

As described above, inside the regression approximation model 202, the size of the model is determined by the distribution amount of the basic element pattern in the basic element pattern storage unit 1 and the number of the basic element patterns, and the model structure as shown in Equation 1 is obtained. Be built. The model parameter setting unit 201 sets parameters w _ik and wb _ij inside the regression approximation model 202.

Here, the meaning of Equation 1 will be described. First, _wik
The k-th distribution value of the i-th basic pattern (strictly speaking,
It represents the deviation) from the average value P _i. Therefore W · X
Operation, it means that by calculating the correlation between the shape pattern X and the basic element patterns nv _i of the rolling material. Also, Wb
Represents the correlation between the basic element patterns. Therefore, the ratio Wb · W · x is determined based on the correlation result between the shape pattern of the rolled material and the basic element pattern.

FIG. 3 shows a detailed view of the load coefficient setting section 3.
The load coefficient setting unit 3 includes a conversion function selection unit 301 and a parameter conversion unit 302.
1 stores in advance a sigmoid function expressed by the following equation 6, which is a non-linear conversion function used in the internal operation of the neural network 7.

[0025]

(Equation 6)

Here, u is an input, y is an output, and km and kt are parameters.

The parameters of the regression approximation model are input from the function approximation unit 2 to the parameter conversion unit 302. Further, the parameters (km, kt) of the sigmoid function used in the neural network 7 from the learning control unit 6 are input to the conversion function setting unit 301, and the inverse function of the sigmoid function is determined. The inverse function and the parameters (wb _ij , w _jk ) of the regression approximation model obtained by the function approximation unit 2 are input to the parameter conversion unit 302. Then, the parameters (wb _ij , w _jk ) of the regression approximation model are converted based on the inverse function from the conversion function selection unit 301 and output to the learning control unit 6. The regression approximation model and the neural network are considered in this way.

The parameter conversion is performed as follows. In the function approximation unit 2, the relationship between input and output can be represented by the combination of two linear mappings by the basic element pattern, as shown in Expression 1. At this time, it can be seen that the structure of the neural network 7 also consists of two conversions: a product-sum operation between the input layer and the intermediate layer and a product-sum operation between the intermediate layer and the output layer. Therefore, the conversion by the parameter w in the regression model is made to correspond to the conversion between the input layer and the intermediate layer, and the conversion by the parameter wb in the regression model is made to correspond to the conversion between the intermediate layer and the output layer. That is, w _ik corresponds to the weight coefficient between the input layer and the intermediate layer of the neural network, and wb _ij corresponds to the weight coefficient between the intermediate layer and the output layer.

Further, the structure of the neural network 7 is also determined. That is, the number of units in the input layer is the same as the number of sensors distributed in the width direction of the rolled material (= m), and the number of units in the intermediate layer and the output layer is the number of basic element patterns (=
n).

Next, the process of converting the parameters of the regression approximation model, which is performed inside the parameter conversion unit 302, into the weighting factors of the neural network will be described. The sigmoid function used for the neural network 7 is
It can be approximately expressed as in Equation 7 near zero of u. The parameter KL is a gradient obtained by approximating the sigmoid function of Equation 6 in the vicinity of zero of u by a least square method using a linear expression.

[0031]

Y = KL × u (Equation 7) Therefore, considering the input / output relationship between the intermediate layer and the output layer of the neural network 7, the following Expression 8 is obtained.

[0032]

(Equation 8)

The input pattern where u _k is inputted to the input layer, Wn1 _jk intermediate layer weighting factor for that input, y
_cj is the output of the intermediate layer, u _k is inputted to the output layer input,
Wn2 _ij output layer load factor for the input, y _oi is the output of the output layer. Equation 8 is the same as the structure of the regression approximation model. Therefore, the weight coefficient W of the neural network
n1 _jk , Wn2 _ij and the parameters (w
_jk , wb _ij ) can be expressed as in the following Expression 9.

[0034]

W _jk = KL × Wn1 _jk wb _ij = KL × Wn2 _ij (Equation 9) Therefore, inside the parameter conversion unit 302, the neural network 7 includes the neural network 7 together with the number of units of the input layer, the intermediate layer and the output layer. The weight coefficient corresponding to the network 7 is calculated by the following equation 10, and is output to the learning control unit 6.
Then, the learning control unit 6 calculates the weighting factor obtained by this calculation together with the number of units of the input layer, the intermediate layer, and the output layer of the neural network 7.
Is set in the neural network 7 as the initial value of the weighting factor. Note that the transformation of Expression 10 is an approximate expression of a sigmoid function, and the inverse transformation of Expression 7.

[0035]

(Equation 10)

FIG. 4 is a detailed diagram of the approximation degree judging section 4. The approximation degree judgment unit 4 includes a regression approximation model 401 having the same approximation model as the regression approximation model 202 which is a component of the function approximation unit 2, an approximation degree calculation unit 402, and a judgment unit 403. The regression approximation model parameters from the function approximation unit 2 are as follows:
A linear approximation relationship between the basic element pattern and the shape pattern input to the regression approximation model 401 and obtained by the function approximation unit 2 is realized. When the regression approximation model 401 is realized, a shape pattern (input data) is input from the learning data storage unit 1 to the regression approximation model 401, and an estimated value of a component ratio of each basic element pattern included in the shape pattern of the rolled material Is calculated. The estimated value calculated by the regression approximation model 401 and the initial value of the component ratio of each basic element pattern preset in the learning data storage unit 1 by an expert or the like are input to the approximation degree calculation unit 402, The error between the estimated value and the initial value is calculated. This estimation error is calculated from the sum of squares of the estimation error of the component ratio for each basic element pattern.

The approximation degree calculating section 402 outputs the estimation error to the determining section 403. When the estimation error sent from the approximation degree calculation unit 402 exceeds a preset allowable estimation error, the determination unit 403 determines that the regression model cannot be sufficiently approximated, and stores the shape pattern in the correction learning data storage unit. Send to 5. Data that could not be sufficiently approximated by the regression model is stored in the correction learning data storage unit 5 in this way.

FIG. 5 is a detailed diagram showing the relationship between the learning control unit 6 and the neural network 7. The learning control unit 6
It comprises a data storage unit 601, a weight coefficient storage unit 602, a learning execution unit 603, and a switching unit 604. The shape pattern output from the learning data storage unit 1 and the correction learning data storage unit 5 and the component ratio of each basic element pattern included therein are stored in the data storage unit 601. The load coefficient obtained by the load coefficient setting unit 3 is stored as a load coefficient initial value in the load coefficient storage unit 602 inside the learning control unit 6. In addition, the load coefficient setting unit 3 outputs the number of units of the input layer, the intermediate layer, and the output layer according to the calculated load coefficient initial value to the learning execution unit 603. The learning execution unit 603 outputs the number of units in the hidden layer of the neural network 7 and the parameters (km, kt) of the sigmoid function used in the internal calculation to the neural network 7 to determine the neural network 7.

Next, a method of learning the neural network 7 will be described. When the neural network 7 is determined, the switching unit 604 connects the weight coefficient storage unit 602 and the neural network 7, and switches the weight coefficient storage unit 602.
Is output to the neural network 7 as an initial value. Next, the switching unit 604 is switched, the learning execution unit 603 and the neural network 7 are connected, and learning is started.

The learning execution unit 603 includes a data storage unit 601
And the neural network 7
, And corrects the load coefficient sent from the load coefficient storage unit 602 according to a learning rule such as back propagation. The corrected load coefficient is applied to the switching unit 6
The signal is sent to the neural network 7 via the network 04 and changes the weight coefficient of the neural network 7. When learning is performed using the data sent from both the learning data storage unit 1 and the correction learning data storage unit 5 for correcting the neural network 7, the data in the correction learning data storage unit 5 is given priority. Use it for learning.
This is because the initial value of the neural network 7 has a linear relationship, and therefore has nonlinearity and may not be sufficiently approximated by a linear model. Therefore, the data from the correction learning data storage unit 5 is taught to the neural network 7 more frequently.

FIG. 7 shows the result of comparison between the neural network designed according to the present invention and a conventional neural network in which the initial values of the weighting factors are set by random numbers. In the simulation, the basic element patterns f ₁ , f ₂ , and f ₃ were synthesized at a random ratio (ratio = α: β: γ), and a noise was added to form a shape pattern. Note that the component ratio estimation error is the sum of squares of the difference between the component ratio of each basic element pattern obtained by each neural network and the component ratio when each basic element pattern is actually synthesized. The maximum value of the ratio of noise to the size of the shape pattern is defined as the noise level.

From the results shown in FIG. 7, the error of the component ratio estimation can be reduced in the entire noise level in the present invention than in the conventional neural network. This indicates that the components of each basic element pattern can be more accurately obtained from the detected shape pattern. Further, the present invention has an effect that the component ratio estimation error hardly changes with respect to the magnitude of the ratio of noise included in the detection pattern, and is strong against noise.

FIG. 8 shows a comparison of the weight coefficient between the input layer and the intermediate layer of the neural network between the conventional example and the present invention. In FIG. 8, the black squares are the load coefficient values set from the regression approximation model, and the basic element patterns f ₁ ,
A numerical value corresponding to each distribution value is set from the width direction shape pattern expressing f ₂ and f ₃ as a linear sum. ○ is a load coefficient obtained by learning with the load coefficient set from the regression approximation model as an initial value. In this case, the weighting factor is modified so as to be less sensitive to noise, taking into account the non-linear sum of the basic element patterns. In addition, it can be seen that the values converge at a value similar to the load coefficient set from the regression approximation model, and the basic linear characteristics remain. This means that the calculation between the input layer and the intermediate layer can be performed while maintaining the correlation between the input shape pattern and the basic element pattern.

On the other hand, △ is a conventional neural network, which is a weighting factor whose initial value is set by a random number. When there is not only a weighting factor similar to that of the regression approximation model but also a completely different weighting factor. There is a problem that it is hard to catch the standard of the.

As described above, since the present invention makes it easier to understand the state inside the neural network, there is also an effect that the problem of the neural network can be analyzed when erroneous recognition is made.

The flowchart of FIG. 10 is a method for designing a neural network according to the present invention described above. First, in step 1, a width direction shape pattern of a rolled material is detected. Basic element patterns characteristic of these detected shape patterns are input in step 2.1. In step 3.1, a regression approximation model is created using the basic element pattern input in step 2.1. step
In 4.1, the parameters of the regression approximation model created in step 3.1 are converted to the initial weighting factors of the neural network. In step 5.1, the number of hidden units in the neural network and the parameters of the sigmoid function are determined, and the weighting factor obtained in step 4.1 is set as the initial value of the neural network.

On the other hand, the shape pattern detected in step 1 is collected as learning data in step 2.2,
Input what component ratio each basic element pattern input in ep2.1 is included in the shape pattern. Then, st
In ep3.2, the component ratio of the basic element pattern included in the shape pattern collected in step 2.2 is estimated using the regression approximation model created in step 3.1. Then, step
In 4.2, the result estimated in step 3.2 is compared with the component ratio input in step 2.2, and a shape pattern having a large difference is registered as correction learning data. In step 5.2, a combination of the shape pattern determined in step 4.2 and the component ratio of the basic element pattern included in the pattern is used as learning data for correcting the weighting factor of the neural network 7 as a correction learning data storage unit. Set to 5. st
In ep6, using the neural network determined in step 5.1, the learning data for correction (shape data) collected in step 5.2, and the learning data collected in step 2.2,
Modify the weighting factor of the neural network by a learning rule such as back propagation. At this time, the method of teaching data is to repeatedly teach the pattern of the correction learning data storage unit 5 to the neural network 7 to correct the load coefficient. Then, after repeatedly learning the pattern of the correction learning data storage unit 5 several times, the learning data storage unit 1 has a smaller number of times than the number of times of learning of the correction learning data storage unit 5.
To learn the pattern. Thereafter, the pattern of the correction learning data storage unit 5 is repeatedly learned again. By repeating the above operation, the shape pattern that could not be sufficiently approximated (the pattern of the correction learning data storage unit 5) is weighted and learned.

FIG. 9 is a diagram showing a method of changing the neural network 7 when adding one new basic element pattern to a shape pattern. When a basic element pattern is newly added, first, a unit 700-2 for outputting the component ratio of the basic element pattern is added to the output layer. Further, in order to output the correlation between the basic element pattern calculated between the input layer and the intermediate layer and the input shape pattern in the intermediate layer, one unit 700-1 is added to the intermediate layer. When adding a large number of basic patterns, this operation may be repeated.

[0049]

According to the present invention, the correlation between the detected shape pattern and the basic element pattern is approximated in advance by using a regression model, and the initial value of the weighting factor of the neural network is set based on the result. Therefore, since the approximate relationship between the shape pattern and the basic element pattern can be realized in the neural network, it is possible to converge to the neural network that realizes the basic element pattern component ratio included in the detected shape pattern more optimally. it can. In addition, there is an effect that the state inside the neural network can be easily understood, and a problem of the neural network can be analyzed when erroneous recognition is performed.

[Brief description of the drawings]

FIG. 1 is a configuration diagram illustrating an embodiment of the present invention.

FIG. 2 is a diagram illustrating a function approximation unit.

FIG. 3 is a diagram illustrating a load coefficient setting unit.

FIG. 4 is a diagram illustrating an approximation degree determination unit.

FIG. 5 is a diagram showing a learning control unit and a neural network.

FIG. 6 is a diagram illustrating a learning storage unit.

FIG. 7 is a simulation result using the present invention.

FIG. 8 is an explanatory diagram showing weight coefficients of a neural network when the present invention is used.

FIG. 9 is an explanatory diagram of a neural network structure when the present invention is used.

FIG. 10 is a flowchart of the neural network design method of the present invention.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Learning storage part, 2 ... Function approximation part, 3 ... Weight coefficient setting part, 4 ... Approximation degree determination part, 5 ... Correction learning data storage part,
6: learning control unit, 7: neural network, 8: basic element pattern storage unit.

──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Yuki Katayama 5-2-1 Omika-cho, Hitachi City, Ibaraki Prefecture Within Hitachi Information & Control Systems Co., Ltd. (72) Inventor Masaaki Nakajima 5-chome Omika-cho, Hitachi City, Ibaraki Prefecture No. 1 Hitachi, Ltd. Omika Plant (72) Inventor Satoshi Hattori 5-2-1, Omika-cho, Hitachi City, Ibaraki Prefecture Hitachi, Ltd. Omika Plant (56) References JP-A-4-90004 (JP, A) JP-A-4-127908 (JP, A) JP-A-4-313410 (JP, A) JP-A-5-42314 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) ) B21B 37/00-37/78 G05B 13/02 G06F 15/18 G06G 7/60

Claims (4)

(57) [Claims] 1. A shape detector for detecting sheet thickness shape data of a rolled material, a basic element pattern storage means for storing a basic element pattern of the shape data, and a polynomial of a component ratio of the basic element pattern to the shape data. Function approximation means for approximating, weighting factor setting means for calculating an initial value of the weighting factor of the neural network from the result of the function approximation means, learning data storage means for storing the shape data and the component ratio, A result of the function approximation means, an approximation degree determination means for determining an approximation degree from a comparison between the data from the learning data storage means, and correction data when the determination from the approximation degree determination means is outside an allowable error. Correction learning data storage means for storing the shape data and a component ratio with a basic element pattern included in the shape data; A plant control device comprising: learning control means for learning a neural network from an output from a load coefficient setting means and outputs from the learning data storage means and the correction learning data storage means. . 2. A function approximation of a component ratio of a basic element pattern to the shape data from the thickness data of a rolled material detected by a shape detector, and an initial value of a weighting factor of a neural network is calculated from the approximation result. And inputting the component ratio of the basic element pattern included in the shape data and the shape data as learning data, and determining the degree of approximation from a comparison between the component ratio of the function approximation result and the component ratio of the learning data. A plant control method for controlling a plant by learning a neural network using the shape data as correction learning data when the result of the determination is out of an allowable error. 3. The plant control apparatus according to claim 1, wherein regression analysis means is used as said function approximation means. 4. The plant control apparatus according to claim 1, wherein the determination from the approximation degree determination means is a square of the difference between the component ratios.