JPH05265511A - Control system - Google Patents

Abstract

PURPOSE:To improve processing accuracy designing a control or an information processing based on the theory of concentrated constant system and eliminating approximate operation by extracting the kinds and components of pattern signals, in a mechanism converting the distributed constant signal of a control object into a concentrated constant signal. CONSTITUTION:The information of a distributed constant system is inputted in a distribution constant-concentrated constant conversion mechanism 2 via a sensor. This distributed constant data are converted into concentrated constant by the conversion mechanism 2 and inputted to a concentrated constant type controller 3 constituted of the control theory of the concentrated constant system. The conversion mechanism 2 is composed of a device such as a neurocomputer, for instance. The function preliminarily stores the pattern to be the data of the distributed constant system as a reference waveform and extracts the amount of the reference wave system component to be included in the distributed constant patterns of the outputs of a real control object or an information processing object 2. In a hierarchical type neural net 4, an input is inputted in an input layer, the output of the input layer is inputted in an intermediate layer and the output of an output layer becomes a final output.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a control system for converting a distributed constant type signal to be controlled or processed into a lumped constant type signal and performing control or information processing by using a lumped constant type processing method.

[0002]

2. Description of the Related Art Conventional devices have no established control theory or information processing theory of distributed constant systems for control objects or information processing objects of distributed constant systems. The distributed signal of is divided, and the theory of the lumped constant system is applied to each divided signal, and the control and information processing of the lumped constant system have been approximately performed.

[0003]

In the above-mentioned prior art, the distributed constant system is directly approximated and handled by the lumped constant control system, so that there is a problem in that the capability of control and information processing is limited.

An object of the present invention is to provide a control system capable of performing good control on a controlled object of a distributed constant system.

[0005]

In order to achieve the above object, according to the present invention, a signal of a distributed constant system to be controlled which is a pattern signal is classified into a reference signal component, and the reference signal is included. A component is extracted, and the process corresponding to the component is converted into a lumped constant system process.

[0006]

The converting mechanism for converting the distributed constant signal of the controlled object into the lumped constant signal operates so as to extract the pattern signal as the type of the signal and the component contained therein. Thereby, the control device and the information processing device can be designed based on the control theory of the lumped constant, so that the approximation operation is eliminated,
The processing accuracy can be improved.

[0007]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows a system configuration diagram which is an embodiment of the present invention. The states of the control target and the information processing target 1 are multidimensional distributed constant systems. Information on the distributed constant system is input to the distributed constant-lumped constant conversion mechanism 2 via a sensor (not shown). The distributed constant data is converted into a lumped constant by the distributed constant-lumped constant conversion mechanism 2 and input to the lumped constant type controller 3 constructed by the control theory of the lumped constant system. The lumped constant control device 3 is a device that can be configured by combining, for example, an integrator, an amplifier, a differentiator, or the like according to the multivariable control theory or the classical control theory, or a device to which knowledge processing such as fuzzy control is applied.

The distribution constant-lumped constant conversion mechanism 2 is
For example, it is composed of a device such as a neuro computer. Its function is to store a pattern, which is data of a distributed constant system, as a reference waveform in advance, and
Alternatively, the amount of the reference waveform component included in the distributed constant pattern that is the output of the information processing target 2 is extracted.

For example, an example in which the distribution constant-lumped constant conversion mechanism 2 is composed of a neuro computer will be shown. The neurocomputer is composed of a neural network 4, a learning control mechanism 5, a learning data storage mechanism 6, and a mixing mechanism 7. The neural network 4 is of a multi-layer or hierarchical type.
It is a Rumelhart type network configuration.

The neural network will be described below with reference to FIGS. FIG. 2 shows a neuron model simulating human nerve cells. The number of inputs is plural, and it is assumed here that there are three inputs. The three input values of the neuron are x1, x2, and x3, respectively. Neuron has weight w
1, w2, w3 are multiplied by the respective input values x1, x2, x3, and the sum X thereof is obtained as shown in Equation 1.

[0011]

[Equation 1]

Next, the neuron outputs the output y according to the calculation of the following equation (Equation 2).

[0013]

[Equation 2]

Here, the function f is a logitex function, for example, a sigmoid function expressed by the equation 3 is used.

[0015]

[Equation 3]

A neurocomputer can be constructed by combining such neurons, and FIG. 3 shows a hierarchical type Rumelhart having one intermediate layer in one of these neurocomputers.
The structure of a neural network is shown. The input is input to the input layer, the output of the input layer is input to the intermediate layer, the output of the intermediate layer is the input of the output layer, and the output of the output layer is the final output of the neural network.

It has been confirmed that the neural network has a learning ability. For example, the algorithm is based on Kaoru Nakano's "Introduction and Training Neurocomputer", September 1989, page 38, published by Technical Review Company. There is a backpropagation method described on page 47. With reference to FIG. 2, neuron weight values w1, w2, and w3 can be determined by this back propagation method.

Next, the learning mechanism 5 in the distributed constant-lumped constant conversion mechanism 2 of FIG. 1 will be described.

FIG. 4 shows a learning mechanism 5 according to the present invention. The difference from the learning algorithm in the above document is that a plurality of data in the learning data storage mechanism 6 in the distribution constant-lumped constant conversion mechanism 2 which is a teacher signal is mixed at a certain mixing ratio to create a new teacher signal. is there. In addition, general Rumelhart
The configuration of the model neurocomputer is surrounded by the broken line in FIG.

In the conventional method, a combination of inputs and outputs to be learned by a general Rumelhart-type neurocomputer is stored as a plurality of sets of reference waveforms. The learning of a general Rumelhart-type neurocomputer simply inputs the input of the above-mentioned set of reference waveforms into the input layer of a general Rumelhart-type neurocomputer, and outputs the neural net 4 at that time and the above-mentioned reference waveform. In the direction in which the output difference decreases, the steepest gradient method is used for the neuron weight values w1, w
Change 2, w3.

Since the output of each neuron model is represented by a non-linear continuous sigmoid function, the neural network is basically a system with strong threshold processing.
Therefore, the teacher signal, which is the output of the reference waveform, is given as a digital value of 0/1 and has been used to classify the input signal into which reference waveform. That is, it is judged by the dominant feature included in the input signal, and the output is learned to be either 1 or 0.

On the other hand, in the present invention, an analog pattern is given to the teacher signal, and the component analysis is performed to extract the ratio of the reference waveform component included in the input. Since it is a component, it is desirable that a continuous analog value is output instead of a digital value such as 0 or 1. The mixing mechanism 7 shown in FIG. 4 is added to realize this function.

FIG. 5 shows an outline of the operation of the mixing mechanism 7. The input teacher signal 8 is input from the learning data storage mechanism 6, and the coefficient generated from the coefficient generation mechanism 14 is input to the first input signal 8 (greater than 0 and less than 1, but in this case, normalized). Is multiplied by a coefficient in the multiplication mechanism 10. Furthermore, the coefficient generating mechanism 1 is added to the second input teacher signal 9.
A coefficient 11 generated from 4 (greater than 0 and less than 1, but normalized in this case) 11
Multiply by. The outputs of the multiplication mechanisms 10 and 11 are added by the addition mechanism 12 to generate a new input teacher signal 13.

The output teacher signal 15 corresponding to the new input teacher signal 13 is composed of the coefficient of the coefficient generating mechanism 14. That is, when 100% of the input teacher signal 8 is input, the value of the output teacher signal corresponding to the output of the excited neural network takes the coefficient value of the multiplication mechanism 10. Similarly, the value of the output teacher signal corresponding to the output of the neurocomputer excited by the input teacher signal 9 becomes the coefficient of the multiplication mechanism 11.

A simulation was performed using the shape basic pattern 1 shown in FIG. 6 and the shape basic pattern 2 shown in FIG. Here, pattern 1 has a maximum value at point b in the horizontal axis direction, and pattern 2 has a maximum value at point d in the horizontal axis direction. The points a to e in the horizontal axis direction are divided into 36, and the height of each divided point is used as an input to the neural network 4. The structure of the neural network 4 includes 36 neurons in the input layer,
It consists of 18 neurons in the middle layer and 10 neurons in the output layer.

Table 1 shows the operation results of the conventional neuro computer. That is, the basic pattern 1 in FIG. 6 is used as an input teacher signal, and as the output teacher signal, the first output neuron is 1 and the other output neurons are 0.
Next, the basic pattern 2 shown in FIG. 7 is used as an input teacher signal, the second neuron is used as an output teacher signal, and the other output neurons are learned as 0.

[0027]

[Table 1]

After the learning is completed, for example, as shown in item 2 of Table 1, the waveforms are synthesized with the ratio of the basic pattern 1 being 0.3 and the ratio of the basic pattern 2 being 0.7, and the neural network 4 To enter. As a result, the certainty factor of the basic pattern 1, which is the output value of the first output neuron, is 0.071, and the certainty factor of the basic pattern 2 is 0.996, and the neurons showing the dominant component are classified. Table 1 shows the result of mixing the basic pattern 1 and the basic pattern 2 in the ratio of item number 1 to item number 5. From this result, it can be said that the neural network 4 approaches 1 when the threshold value is exceeded and approaches 0 when the threshold value is exceeded.

On the other hand, according to the present invention, the basic patterns are mixed at a certain ratio for learning. For example, the basic pattern 1 and the basic pattern 2, which are the reference waveforms in Table 1, are (0.
0, 1.0), (0.2, 0.8), (0.4, 0.6),
(0.6, 0.4), (0.8, 0.2), (1.0, 0.0).
6 sets of reference waveforms are mixed with each other at a ratio of 0), and the neural network 4 is made to learn.

When the learning is completed, as shown in item 2 of Table 2, the waveforms are synthesized with the ratio of the basic pattern 1 being 0.3 and the ratio of the basic pattern 2 being 0.7.
Input to the neural network 4. As a result, the first
The output pattern of the output neuron of the basic pattern 1 has a certainty factor of 0.308, and the basic pattern 2 has a certainty factor of 0.692. That is, the components of the basic pattern 1 and the basic pattern 2 can be extracted. Similarly, item Nos. 3 and 4
Although the combination of the basic pattern 1 and the basic pattern 2 has not been learned, it shows that the components included in these patterns can be faithfully extracted.

[0031]

[Table 2]

Further, the same result can be obtained by learning by combining the basic pattern 3 shown in FIG. 8 and three patterns of the basic patterns 1 and 2.

FIG. 9 shows a configuration diagram when the above results are organized and applied to a ZR mill of a rolling mill. ZR mill body 5
Reference numeral 1 is composed of rolling rolls arranged in a tuft of grapes. The rolling roll is driven by the main motor 53 via the speed reducer 52.
Driven by. The rolled material 54 placed in the gap between the upper and lower rolling rolls is supplied from the payoff reel 55 and wound up by the tension reel 56. A tension is generated in the pay-off reel 55 due to a speed difference between the electric motor 53 that drives the reel via the reduction gear 57 and the work roll, and a desired plate thickness is obtained. Similarly, in the tension reel 56, tension is generated by the speed difference between the electric motor 60 that drives the reel and the work roll via the speed reducer 58, and a desired plate thickness is obtained. The output of the shape detection mechanism 62 that contacts the rolled material 54 and detects the tension distribution of the rolled material (which can be converted into a sheet thickness distribution) is input to the shape detection processing mechanism 65. These shape detection processing mechanism 6
5, the distributed constant-lumped constant conversion mechanism 2 and the control calculation mechanism 3 are collectively referred to as a control mechanism 64. The control command output from the control mechanism 64 is A, which is the actuator of the ZR mill.
Generates commands for S-U roll and lateral shift.

The shape detection processing mechanism 65 removes a noise component from the waveform on which noise is superimposed, and compensates for the non-linearity of the shape detection mechanism 62. The output of the shape detection processing mechanism 65 is the distribution constant-lumped constant conversion mechanism 2
Is input to and the reference waveform component is extracted. The control calculation mechanism 3 using the reference waveform component is a control device designed by the lumped constant control theory. In this case, fuzzy control is used.

In the rolling mill shape control system having such a configuration, the shaded shape waveform 70 in FIG.
Suppose you typed in. The waveform of each part of the control mechanism 64 at this time is shown in FIG. The output of the shape detection processing mechanism 65 is saturated because the signal level is high like the output 71 of the shape detection processing mechanism 65 shown in FIG. 11, and the peak is cut. The output of the distributed constant-lumped constant conversion mechanism 2 shows the value in the area of a rectangle. That is, the proportion of the component 72 of the reference pattern 1 that has been learned in advance is the rectangle 73
The value is shown by the area of a rectangle filled with a diagonal line inside, and the maximum value is 1. In the case of FIG. 11, the certainty factor including pattern 1 is 0.3, and the certainty factor including pattern 7 is 0.7.
It can be seen that it is.

As shown in FIG. 11, upon receiving the output 72 of the distributed constant-lumped constant conversion mechanism 2, the control calculation mechanism 3
"If pattern 1 is used, move AS-U roll No. 2 in the upward direction. If pattern 7 is used, the upper first intermediate roll (tapered at the right end of the roll) is shifted to the right (called the upper lateral shift). The control commands 74 and 75 are generated according to the fuzzy control rule of "Yes.", And the commands are sent to the actuator.

It can be seen that when such control commands are sequentially sent to the actuator, the shape of the non-rolled material changes and becomes flat as shown in FIG.

Next, a second embodiment of the present invention will be described. The conversion mechanism shown in FIG. 12 will be described below as a method of classifying a waveform that is a distributed constant system into reference patterns.

The output waveform g (x) of the shape detecting mechanism 63 is converted into the waveform component a _i.

[0040]

[Equation 4]

Is represented by However,

[0042]

[Equation 5]

Is the i-th reference pattern and n (n is a positive number)
Let's make it a kind.

Then, the both sides of the equation 4 are multiplied by the reference pattern to obtain the equation 6.

[0045]

[Equation 6]

Here, the equation 3 is transformed to obtain the equation 7.

[0047]

[Equation 7]

However,

[0049]

[Equation 8]

[0050]

[Equation 9]

[0051]

[Equation 10]

It is That is, F is a square matrix with n rows and n columns,
A is an n-row, 1-column vector. Equation 7 is a simultaneous equation,
Solving for a _i gives the components of f _i (x).

[0053]

[Equation 11]

When f _i (x) and f _j (x) are orthogonal, f _i (x) f _j (x) = 0 (where i ≠ j),

[0055]

[Equation 12]

F is a diagonal matrix and is easy to solve.

FIG. 1 shows a functional block for executing the above result.
2 shows. The output 71 of the shape detection mechanism 65 is a multiplication mechanism with the reference waveform 100 stored in the learning data storage mechanism 6.
It is multiplied by 101 and input to the integration mechanism 102. The output of the integration mechanism 102 corresponds to the expression (8).

The simultaneous equation coefficient generator 103 and the simultaneous equation solver 104 are composed of a microcomputer 105, and the simultaneous equation coefficient generator 103 has a reference waveform 100 in advance.
Data is passed to the simultaneous equation solver 104 using the coefficient obtained by the calculation of Equation 9. Each time the output of the integration mechanism 102 is input, the simultaneous equation solver 104 multiplies the inverse matrix generated from the coefficients of the simultaneous equations to generate the waveform component a _i (i = 1).
~ N) 106 is output. These waveform components 106 are input to the control calculation mechanism 3. A simulation was conducted to confirm the effect of this method, and the results are shown below.

Table 3 shows the inputs in another embodiment of the present invention. The position is 0 at a position corresponding to point a in FIG. 6, and e
The place corresponding to the point is 6, and the space between them is divided into 36 equal parts. This indicates the location of the detection point of the shape sensor that detects the shape of the waveform. The random number is a uniform random number generated assuming that noise is placed on the sensor, and the value is shown. As the waveforms A, B, and C, the waveforms of normal distribution represented by the equations 13, 14, and 15, respectively, were used. (However, x represents a position.)

[0060]

[Table 3]

[0061]

[Equation 13]

[0062]

[Equation 14]

[0063]

[Equation 15]

The input waveform is generated by mixing the waveform component A of 0.7, the waveform component B of 0.6, the waveform component C of 0.05, and the random number component of 0.1.

The equation formed by the simultaneous equation solver 104 is expressed by equation 16.

[0066]

[Equation 16]

The output of the simultaneous equation solver 104 is
The answer is

[0068]

[Equation 17]

It becomes At this time, the error components of the waveform A, the waveform B, and the waveform C are 11%, 11%, and 154%, respectively.

FIG. 13 illustrates the concept of the distributed constant-lumped constant conversion mechanism 2 described above. That is, data that represents a pattern using a plurality of sensors, such as the strip thickness distribution of a rolling mill, and data from different types of sensors including a speedometer and an audiometer are collected and abstracted. It is converted into scalar data.

FIG. 14 shows a block of a conventional control system, and the output of the controlled object is distribution constant / vector information of distributed data. These pieces of information are detected by the sensor and fed back as scalar information. As the controller at this time, various controllers that are said in control theory can be used.

On the other hand, the distribution constant-lumped constant conversion mechanism 2 of the present invention abstracts the information of the distribution constant and inputs it to the controller 3 as one-dimensional scalar information having an abstract meaning.

Conventionally, there was a limit to the improvement of accuracy because the control system was constructed by the approximate calculation, but such a configuration enables precise control.

[0074]

According to the present invention, it is possible to classify data distributed in a distribution constant into a reference waveform stored in advance, and to detect the amount of the component, which enables more accurate control. Also, the control law can be designed based on the design theory of the lumped parameter system.

[Brief description of drawings]

FIG. 1 is a configuration diagram of an entire system that is an embodiment of the present invention.

FIG. 2 is an explanatory diagram of a neuron model constituting the distributed constant system-lumped constant system conversion mechanism 2 of FIG.

FIG. 3 is a configuration diagram of a hierarchical neural network.

FIG. 4 is a configuration diagram of a distributed parameter system-lumped parameter system conversion mechanism 2 of FIG.

5 is an operation explanatory view of the mixing mechanism 7 of FIG.

FIG. 6 is a diagram illustrating an example of a shape reference pattern.

FIG. 7 is a diagram for explaining another example of the shape reference pattern.

FIG. 8 is a diagram illustrating another example of the shape reference pattern.

FIG. 9 is a configuration diagram of a control system that is an embodiment of the present invention.

FIG. 10 is a diagram showing a simulation result showing a time change of a control system.

FIG. 11 is a diagram showing a response waveform of each part of the control system in the simulation.

FIG. 12 is a diagram showing another configuration example of the distributed parameter system-lumped parameter system conversion mechanism 2.

FIG. 13 is a diagram illustrating a schematic configuration of a distributed constant system-lumped constant system conversion mechanism.

FIG. 14 is a diagram illustrating a schematic configuration of a conventional control system.

[Explanation of symbols]

1 ... Control object, 2 ... Distributed constant system-lumped constant system conversion mechanism,
3 ... Control device, 51 ... Cluster mill, 65 ... Shape detection mechanism, 104 ... Simultaneous equation solver.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Junzo Kawakami 4026 Kuji-cho, Hitachi City, Hitachi, Ibaraki Prefecture Hitachi Research Laboratory, Hitachi, Ltd. (72) Masaaki Nakajima 5-2-1 Omika-cho, Hitachi City, Ibaraki Prefecture Incorporated company Hitachi Ltd. Omika Plant (72) Inventor Satoshi Hattori 52-1 Omika-cho, Hitachi City, Ibaraki Prefecture Incorporated Hitachi Ltd. Omika Plant (72) Executive executor Masakazu Omika-cho, Ibaraki Prefecture Hitachi City 2-1-1, Hitachi Ltd. Omika factory

Claims (5)

[Claims] 1. A distributed constant system comprising a controlled object of a distributed constant system, a detection mechanism for detecting a state of the controlled object, and a control mechanism for controlling the controlled object using a signal of the detection mechanism. A control system having a conversion mechanism for abstracting a detection signal representing the state of the. 2. The control system according to claim 1, wherein the conversion mechanism converts a distributed constant system signal into a lumped constant system signal. 3. The control system according to claim 1, wherein the conversion mechanism learns in advance a relationship between a distribution signal which is an input of the conversion mechanism and an output signal corresponding to the distribution signal. 4. The control system according to claim 3, wherein in the process of learning the relationship between the plurality of inputs and outputs, a new input / output relationship is generated by linearly combining the relationships between the plurality of inputs and outputs. 5. The control system according to claim 1, wherein the control target is a metal manufacturing plant.