JPH0675936A - Method and device for sampling teacher signal in multivariable system - Google Patents

Abstract

PURPOSE:To provide method/device for sampling a teacher signal by which time required for a calculation processing is appropriate and the convergence of an input/ output relation is satisfactory by sampling a part with much quantity of information in the output signal waveform of an unknown system and setting the number of the teacher signals which are to be learned by a multilayered perceptron to be appropriate. CONSTITUTION:An input part 21 samples a time sequential information signal xj which the unknown system 3 outputs at the interval of prescribed time. An output signal xj is inputted to a differential part 22, and the differential vector xj of the output signal xj is operated. An inner product operation part 24 calculates the cosine values of the differential vector xj of the output signal xj and a differential vector xj-1 at a previous sampling point. A comparison part 25 operates an angle thetaj which the two differential vectors make based on the cosine values. When the angle thetaj is compared with a reference value theta0 and the angle thetaj is equal or larger than the reference value theta0 the sample value xj is adopted as the teacher signal. When it is smaller, it is not adopted as the teacher signal. The teacher signal of the time sequential information which is thus selected is used for the learning of the multilayered perceptron.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for sampling a teacher signal in a multivariable system such as a neural network, and more particularly to a method and apparatus for sampling a teacher signal in a multivariable system using a perceptron.

[0002]

2. Description of the Related Art As a method of identifying the input / output relationship of a system in which the input / output functional relationship is unknown, a method of learning by an error backpropagation method using a multilayer perceptron has been conventionally known.

The multi-layer perceptron is a network in which a large number of unit units each having multiple inputs and one output are connected in layers by links for transmitting signals. A coefficient called a weight that represents the strength of the connection is defined for each link.

Each unit receives as input the output of several other units coupled to the input by the link times the weight of the link. The inputs received are added and the sum of the inputs is calculated. The output of the unit is determined according to the sum of this input and the input / output function of the unit. This output is transmitted to the next unit unit through the output side link.

In the multi-layer perceptron, there is no link or feedback link between unit units in the same layer, and the signal flow is from the input layer side to the intermediate layer → ... → The output layer side to the output layer side. One way. The multi-layer perceptron is used for the purpose of estimating the functional relationship between the input and output of an unknown system by the above operation.

Learning to adjust the weights of the multi-layer perceptron as described above by using a signal called a teacher signal so that the input / output relationship of an unknown system can be estimated is called learning. The teacher signal used for this learning samples and uses actual input / output data of an unknown system for estimating the input / output relationship. Here, a case where the output information obtained from this unknown system is time series information will be described.

A method of sampling the time series information input to the multilayer perceptron 10 will be described below. FIG. 7 is a diagram showing a case where the time-series information A is sampled with a short sampling interval. FIG. 8 is a diagram showing a case where the time-series information A is sampled with a long sampling interval. Here, the sampling interval is constant in each of FIGS. 7 and 8.

Time-series information A output by an unknown system
_Are shown in FIGS. 7 and 8 (..., t _j-1 , t _j , t).
_{j + 1} ,. ． ． ) (J = integer), the data is sequentially sampled (sampled) and input as sampling data (output sample) to the computer from a measuring device (not shown) connected to the computer. This output sample is
It is stored in the computer together with the data (input sample) input to the corresponding unknown system.

The stored sampling data is used as a teacher signal (input / output sample) when the multilayer perceptron 10 as described above is trained. Actually, the time series information output by the unknown system is the unit unit 110 of the input layer 11 of the multilayer perceptron 10.
Are stored in the computer in the same manner as the time series information A, and are used as input / output samples of the multilayer perceptron 10 together with input samples to the corresponding unknown system.

[0010]

As described above, in the conventional teaching signal sampling method and apparatus in a multivariable system (multilayer perceptron),
Since the teacher signal (input / output sample) was configured to sample the output signal waveform of the unknown system at a preset fixed time interval, improve the accuracy of identifying the input / output relation of the unknown system. If the sampling interval is shortened as, the number of teacher signals to be learned by the multi-layer perceptron will significantly increase, the calculation process will take a long time, and so-called over-learning will occur. There is a problem that it is difficult to do. On the other hand, when the sampling time interval is lengthened, the time required for the calculation process is shortened. However, in this case, there is a high possibility that wrong learning will be performed.

The present invention has been made in view of the above-mentioned problems, and mainly samples a portion (a portion having a large amount of information) of a signal change in an output signal waveform of an unknown system. However, it is possible to make the number of teacher signals to be learned by the multilayer perceptron appropriate,
As a result, it is an object of the present invention to provide a teacher signal sampling method and its apparatus in a multivariable system in which the time required for the calculation processing is appropriate and the input / output relation of the multilayer perceptron is well converged.

[0012]

In order to achieve the above object, a teaching signal sampling method and apparatus in a multivariable system according to the present invention performs a first operation on one independent signal input. , Outputting a plurality of operation processing results, performing a second operation on these operation results, outputting a plurality of operation processing results, performing a third operation on the second operation results, Using a multilayer perceptron for outputting the calculation processing result, the time series signal is sampled at sampling points having a certain time interval, and the differential vector at the sampling point of the time series information is sequentially calculated from the sampling value, Temporarily storing, the differential vector of the time series information at the sampling points before the sampling point, and the differential vector of the time series information at the sampling points. Based on, to determine the significance of the sampled values of the time-series information at the sampling point,
In the determination of the significance, the sampling value determined to be significant is used as the teacher signal of the multilayer perceptron.

Further, the significance of the sampling value of the time series information is determined by creating a differential vector of the time series information at the sampling point and a differential vector of the time series information at sampling points before the sampling point. It is characterized in that the angle is calculated based on the value of the inner product of these two differential vectors and the angle is equal to or greater than a certain value.

Further, an input processing means having a plurality of input nodes for performing a first calculation on one independent signal input and outputting a plurality of calculation processing results, and a plurality of input nodes in the input processing means. From a plurality of intermediate nodes in the intermediate processing means having a plurality of at least one intermediate node for outputting a plurality of arithmetic processing results and outputting a plurality of arithmetic processing results A time series output from an unknown system having a multilayer perceptron having an output processing means having a plurality of output nodes for inputting a plurality of arithmetic processing results and performing a third arithmetic operation and outputting one output signal The signal is sampled at sampling points having a certain time interval, the differential vector of the time-series information at the sampling points is sequentially calculated from the sampling values, and temporarily stored. , The significance vector of the sampling value of the time-series information at the sampling point based on the differential vector of the time-series information at the sampling point before the sampling point and the differential vector of the time-series information at the sampling point. The sampling value determined to be significant in the determination of the significance is used as a teacher signal of the multi-layer perceptron to identify the input / output relationship of the unknown system.

Further, the significance of the sampling value of the time series information is determined by making a differential vector of the time series information at the sampling point and a differential vector of the time series information at sampling points before the sampling point. It is characterized in that the angle is calculated based on the value of the inner product of these two differential vectors and the angle is equal to or greater than a certain value.

Further, an input processing means having a plurality of input nodes for performing a first calculation on one independent signal input and outputting a plurality of calculation processing results, and a plurality of input nodes in the input processing means. From a plurality of intermediate nodes in the intermediate processing means having a plurality of at least one intermediate node for outputting a plurality of arithmetic processing results and outputting a plurality of arithmetic processing results A multilayer perceptron having an output processing means having a plurality of output nodes for outputting a single output signal by inputting a plurality of arithmetic processing results and outputting a single output signal is provided. Means for sampling at the sampling point, means for sequentially calculating a differential vector of the time-series information at the sampling point from the sampling value, and temporarily storing the differential vector; It said differential vector of the time-series information of the ring points previous sampling point, based on the differential vector of the time-series information at the sampling point,
It has means for determining the significance of the sampling value of the time-series information at the sampling point, and means for using the sampling value determined as significant in the significance determination as a teacher signal of the multilayer perceptron.

The means for determining the significance of the sampling value of the time-series information includes a differential vector of the time-series information at the sampling point and a differential vector of the time-series information at sampling points before the sampling point. Is calculated based on the value of the inner product of the two differential vectors to detect that the angle is equal to or greater than a certain value.

[0018]

The output signal of the unknown system is sampled at constant time intervals, and the differential value vector of the output signal waveform at each sampling point is calculated. From this differential value vector and the differential value vector of the signal waveform at the sampling point immediately before each sampling point, the change point of the output signal waveform of the unknown system is extracted, and only this change point is used as a teacher signal in the computer. By storing, only the point having a large amount of information in the output signal of the unknown system can be input to the multilayer perceptron. As a result, the number of teacher signals becomes appropriate and the sled time becomes appropriate as compared with the case of simply sampling at equal time intervals. Further, since each teacher signal has a large amount of information (significant), the input / output relationship identified by the multilayer perceptron is easily converged, and its accuracy is increased.

[0019]

EXAMPLES Examples will be described below. FIG. 1 is a diagram showing a structure of a multilayer perceptron 10 applied to a teacher signal sampling method and its apparatus in a multivariable system of the present invention. In FIG. 1, a multilayer perceptron 1
0 is composed of the input layer 11, the intermediate layer 13, and the output layer 15, and the layers have the coupling relations (the coupling between the input layer and the intermediate layer 12 and the coupling between the intermediate layer and the output layer 14) as shown in FIG. It is a three-layer perceptron having. Input layer 11, middle layer 1
3 and the output layer 15 are k, m, and n unit units (nodes) 110a to 110k and 130a, respectively.
˜130 m, 150 a to 150 n.

Here, the number of unit units 110 included in the input layer 11 is the number of input variables of the unknown function for identification, and the number of unit units 150 included in the output layer 15 is the number of output variables. . The initial number of the unit units 130 included in the mid layer 13 is arbitrary. Also, the unit units 110a to 110k, 130a to 130m, 150a to
150n is on a computer (not shown),
It is composed of software.

The operation of the multilayer perceptron 10 will be described below. The input layer 11 and the intermediate layer 13 have a coupling relationship as indicated by an arrow of the coupling 12 of the input layer and the intermediate layer,
The intermediate layer 13 and the output layer 15 have a coupling relationship as indicated by an arrow of the coupling 14 between the intermediate layer and the output layer.

The operation of the unit unit 130i (i = 1, 2, ..., M) included in the intermediate layer 13 will be described. The unit unit 130i includes the unit unit 110j and the unit unit 130 in the output Y _j of the unit unit 110j (j = 1, 2, ..., K) included in the input layer 11.
A numerical value representing the coupling strength of i, the sum of the values obtained by multiplying the weights W _ij are input, and a certain nonlinear function f is applied to the input.
The value Y _i subjected to is output. That is, the unit unit 13
The output Y _j 0 of the output Y _i and basic unit 110,

[Equation 1] The relationship is established.

Unit unit 130 included in the intermediate layer 13
The same relationship holds between i and the unit unit 150 _h included in the output layer 15. That is, the unit unit 150
Output Y _{h of h} (h = 1, 2, ... N) Unit unit 1
The output Y _i of 30 is

[Equation 2] The relationship is established. Here, the function g is a non-linear function corresponding to the above-mentioned function f, and usually g = f. Also, W
_hi is a numerical value representing the strength of the bond corresponding to the weight W _ij .

When there are a sufficient number of samples (teacher signals) of combinations of input values and output values of an unknown system,
A sample of the input value of the unknown system is input to the multilayer perceptron 10 described above, and the output of the multilayer perceptron 10 with respect to this input value is compared with the sample of the output value of the corresponding unknown system.
The weights W _hi and W _ij are adjusted so that the output of 0 and the sample value of the output value of the unknown system match (hereinafter, this is referred to as learning). By repeating this operation a sufficient number of times, the multilayer perceptron 10
Identifies the input / output relationship of an unknown system.

In the form of a multilayer perceptron 10,
For perceptrons with three or more layers, if the number of unit units in the middle layer (layers other than the input layer and output layer in a multilayer perceptron of three or more layers) is sufficiently large, the input / output of any unknown system It is known that the functional relation of can be approximated with arbitrary accuracy.

As a method for causing the multilayer perceptron 10 to perform the learning, a method called an error back propagation method is known to be effective. Backpropagation method is a method that minimizes the sum of squares of the errors in the output Z _p of the p-th output sample Z _p 'and multilayer perceptron 10 of the unknown system. The sum of the squared errors is a loss function E. The loss function E is

[Equation 3] However, in (Equation 3), w is the number of input samples.
It can be expressed as.

[0027] Each weight W _hi, each weight in each gradient (gradient) direction about W _ij W _hi, to modify the W _ij. This method corresponds to the steepest descent method. That is,
Weights W _hi and W _hi before learning and weights W _hi 'and W _ij ' after learning are

[Equation 4]

[Equation 5] and,

[Equation 6]

[Equation 7] It becomes a relationship like.

FIG. 2 is a diagram showing the configuration of the sampling device 1 for implementing the present invention. In FIG. 2, computer 2
Is a CPU, a memory, which realizes the sampling device 1.
This is a general digital computer including an I / O circuit and the like (not shown). Hereinafter, for convenience of description, the operation of the hardware of the sampling apparatus 1 will be omitted as appropriate. The unknown system 3 is a system whose input / output relation is identified. For convenience of explanation, in this embodiment, the case where the unknown system 3 is a one-input, one-output system will be described. Therefore, the multilayer perceptron 10 is also a one-input, one-output multilayer perceptron. The sampling section 20 is a section that gives an input signal to the unknown system 3 and samples the output signal thereof.

FIG. 3 is a diagram showing the configuration of the sampling section 20. In FIG. 3, the input unit 21 is a unit that samples the output signal of the unknown system 3 at constant time intervals. The differentiating unit 22 is a unit that differentiates the output signal of the unknown system 3 sampled by the input unit 21 and calculates a differential vector. The storage unit 23 is a unit that stores the differential vector obtained from the differentiating unit 22 up to the next sampling point.

The inner product calculating unit 24 calculates the inner product of the differential vector obtained by the differentiating unit 22 and the differential vector at the previous sampling point stored in the storage unit 23, and the cosine of these two vectors is generated. This is a part for calculating the value cos θ. The comparison unit 25 is a unit that compares the cosine value obtained by the inner product calculation unit 24 with the reference value (cos θ ₀ ). The storage unit 26 is a circuit that stores a reference value (cos θ ₀ ).

Based on the sampling point determination signal output from the comparison unit 25, the storage unit 27 outputs the sampling value (output sample) of the output signal of the unknown system 3 output from the input unit 21 to the corresponding unknown value. System 3
This is a part to be stored together with the input value (input sample) of The output unit 28 is a unit that outputs the input sample to the unknown system 3. Each part of the sampling device 1 described above may be configured as hardware or software. Further, the output unit 28 and other portions may be configured on the same computer or on different computers.

The operation will be described below. FIG. 4 is a diagram explaining the principle of the sampling method of the present invention. Hereinafter, description will be given with reference to FIG. With respect to the input signal Y to the unknown system 3 which is output from the output unit 28 at a constant time interval, a signal (time series information) as shown in FIG. 4A is output from the unknown system 3. .

The input unit 21 outputs the unknown system 3 at each point (..., t _j-1 , t _j , t _{j + 1} , ...) Shown in FIG. 4 at a constant time interval. Sample the signal. The sampling point at this point is t _j , and the sampling values at the above points are x _j−1 , x _j , and x _{j + 1} , respectively. That is, the sampling points at the input unit 21 are arranged at equal time intervals, like the sampling points shown in FIGS. 7 and 8.

The sampled output signal x _j is input to the differentiator 22, and the differential vector x _j 'of the output signal x _j is calculated. Here, various methods of calculating the differential vector x _j 'can be considered. For example, sampling points are further provided near the above sampling points (..., t _j-1 , t _j , t _{j + 1} , ...) And the output signal of the unknown system 3 is sampled. The differential vector x from the sampling points and the sample values at the sampling points in the vicinity
approximate _j '. Alternatively, the differential vector x can be calculated from the sampling values at the sampling point and the sampling point immediately before it.
_A method of approximating _j 'can be considered.

The differential vector x _j calculated by the differentiator 22
Is input to the storage unit 23 and the inner product calculation unit 24. The storage unit 23 stores the differential vector x _j 'up to the next sampling point. In addition to the differential vector x _j ′, the inner product computing unit 24 receives the differential vector x _j−1 ′ at the previous sampling point from the storage unit 23. Inner product calculation unit 2
In 4, the inner product of the differential vector x _j 'and the differential vector x _j-1 ' is calculated, and the cosine value of both is calculated. That is,

[Equation 8] Is calculated.

Cosine value cos calculated by the inner product calculating unit 24
θ _j is input to the comparison unit 25, and the angle θ _j is calculated.
This angle θ _j is compared with the reference value θ ₀ stored in the storage unit 26. When the angle θ _j is equal to or larger than the reference value θ ₀ , the sample point determination signal is activated, the storage circuit 27 stores the sample value x _j of the output signal of the unknown system 3 as the output sample of the teacher signal, and The input signal Y _j to the unknown system 3 at this time is stored as an input sample of the teacher signal and is output to the multilayer perceptron 10.

When the angle θ _j is smaller than the reference value θ ₀ , the sampling value x _j at the sampling point t _j is not stored and is not adopted as a teacher signal. With that, the processing for one sampling value is completed. Such processing is performed for each sampling value. Note that the handling of the sampling value x _j when the angle θ _j and the reference value θ ₀ are the same is not limited to the above description, and the sampling value x _j may not be stored and adopted as a teacher signal. .

In FIG. 4, the sampling value x _{j at} the sampling point t _j is adopted as a teacher signal because the angle θ _j formed by the differential vector x _j-1 'and the differential vector x _j ' is smaller than the reference value θ _0. Not done. On the other hand, the sampling value x _{j + 1 at} the sampling point t _{j + 1} is the angle θ _{j + 1} formed by the differential vector x _j ′ and the differential vector x _{j + 1} ′.
Is larger than the reference value θ ₀ , it is used as a teacher signal.

The input samples and output samples of the teacher signal input to the multilayer perceptron 10 are used for learning of the multilayer perceptron 10. Multilayer Perceptron 1
The operation at the time of learning with 0 is as described above.

FIG. 5 is a diagram showing sampling points corresponding to sampling values adopted as teacher signals when the sampling apparatus 1 is applied to the output signal waveform of the unknown system 3 shown in FIGS. 7 and 8. . In FIG.
Sampling points corresponding to the sampling values adopted as the teacher signal are indicated by vertical dotted lines in the figure. Figure 5
In the portion where there are many changes in the output signal of the unknown system 3 shown in (A), the frequency of appearance of sampling points corresponding to the sampling value adopted as the teacher signal is high. On the other hand, in the portion shown in (B) of FIG. 5 where the change is relatively small, the frequency of appearance of sampling points corresponding to the sampling value adopted as the teacher signal is low.

FIG. 6 is a flow chart showing the processing of the sampling apparatus 1 in this embodiment. In FIG. 6, in step 00 (S00), the input unit 21
The output signal x _j of the unknown system 3 is sampled at regular time intervals. In step 01 (S01), the sampled output signal x _j is input to the differentiator 22,
The differential vector x _j 'of the output signal x _j is calculated. In step 02 (S 02), the differential vector x _j calculated by the differentiator 22 is input to the storage 23 and the inner product calculator 24. The storage unit 23 stores the differential vector x _j 'up to the next sampling point. Inner product calculation unit 24, 'than the other in the storage unit 23 of the differential vector in the previous sampling point x _j-1' the differential vector x _j is input,
The inner product of the differential vector x _j 'and the differential vector x _j-1 ' is calculated, and the cosine value of both is calculated. That is,

In step 03 (S03), the cosine value cos θ _j calculated by the inner product calculating unit 24 is input to the comparing unit 25, and the angle θ _j is calculated. The angle θ _j is stored in the storage unit 2
6 is compared with the reference value θ ₀ stored in 6. When the angle θ _j is equal to or larger than the reference value θ ₀ , the process proceeds to S04,
If it is smaller, the process ends. Step 04 (S04)
, The sample point determination signal is activated, the memory circuit 27 stores the sample value x _j of the output signal of the unknown system 3 as the output sample of the teacher signal, and the input to the unknown system 3 at this time. The signal Y _j is stored as an input sample of the teacher signal and output to the multilayer perceptron 10.

In this embodiment, the unknown system 3
Described the case of 1 input, 1 output, but multiple inputs,
The sampling device 1 can be applied to a system having a plurality of outputs with a similar configuration. In this case, the sampling point determination signal is activated when a significant sampling value is generated in any one of the plurality of output signals. Alternatively, various configurations are conceivable, such as activating the sampling point determination signal when a significant sampling value occurs in an important signal among the plurality of output signals.

In the above embodiments, the multi-layer perceptron having one intermediate layer has been described as an example, but a multi-layer perceptron having two or more intermediate layers also has a teaching signal sampling method and apparatus in the multivariable system of the present invention. Is applicable. Further, although the multi-layer perceptron is realized by software on a computer, other configurations, for example, one or a plurality of unit units can be realized on one computer and a large number of these computers can be connected in a so-called multiprocessor system. A multi-layer perceptron may be implemented. The teaching signal sampling method and apparatus in the multivariable system of the present invention can have various configurations other than those described above. The embodiments described above are merely examples.

[0045]

As described above, according to the present invention, among the output signal waveforms of the unknown system, the portion where the signal changes largely (the portion where the amount of information is large) is mainly sampled, and the multi-layer perceptron is learned. It is possible to make the number of teacher signals to be made appropriate, and as a result, the time required for the calculation processing is reasonable, and the teacher signal sampling method in the multivariable system in which the input / output relation of the multilayer perceptron is well converged. And its device can be provided.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration of a multi-layer perceptron to which a teaching signal sampling method and an apparatus thereof in a multivariable system of the present invention are applied.

FIG. 2 is a diagram showing a configuration of a sampling device for implementing the present invention.

FIG. 3 is a diagram showing a configuration of a sampling unit of the present invention.

FIG. 4 is a diagram illustrating the principle of the sampling method of the present invention.

FIG. 5 is a diagram showing a case where the sampling device of the present invention is applied to an output signal waveform of an unknown system shown in FIGS. 7 and 8,
It is a figure which shows the sampling point corresponding to the sampling value employ | adopted as a teacher signal.

FIG. 6 is a flowchart showing the processing of the sampling apparatus of the present invention.

FIG. 7 is a diagram showing a case of sampling the time-series information A with a short sampling interval in the conventional method and apparatus for sampling a teacher signal in a multivariable system.

FIG. 8 is a diagram showing a case of sampling the time-series information A at a long sampling interval in a conventional teacher signal sampling method and apparatus in a multivariable system.

[Explanation of symbols]

 1 ... Sampling device 2 ... Computer 3 ... Unknown system 10 ... Multilayer perceptron 11 ... Input layer 110 ... Unit unit 12 ... Coupling of input layer and intermediate layer 13 ... -Intermediate layer 130 ... Unit unit 14 ... Coupling of intermediate layer and output layer 15 ... Output layer 150 ... Unit unit 20 ... Sampling unit 21 ... Input unit 22 ... Differentiating unit 23 ... Storage unit 24 ... Inner product calculation unit 25 ... Comparison unit 26 ... Storage unit 27 ... Storage unit 28 ... Output unit

Claims (6)

[Claims] 1. A first operation is performed on one independent signal input, a plurality of operation processing results are output, a second operation is performed on these operation results, and a plurality of operation processing results are obtained. Using a multi-layer perceptron that outputs, performs a third operation on the second operation result, and outputs a plurality of operation processing results, samples a time-series signal at sampling points having a certain time interval, The differential vector of the time-series information at the sampling point is sequentially calculated from the sampling value and temporarily stored, and the differential vector of the time-series information at the sampling point before the sampling point and the time at the sampling point. Based on the differential vector of the series information, to determine the significance of the sampling value of the time series information at the sampling point, in the determination of the significance, as significant Sampling method for the teacher signal in multivariable systems using constant sampling value as a teacher signal of the multi-layer perceptron. 2. The method of sampling a teacher signal in a multivariable system according to claim 1, wherein the significance of the sampling value of the time series information is determined by a differential vector of the time series information at the sampling point,
The angle formed by the differential vector of the time-series information at the sampling points before the sampling point is calculated based on the value of the inner product of these two differential vectors, and the angle is equal to or greater than a certain value. A sampling method for teacher signals in a characteristic multivariable system. 3. An input processing means having a plurality of input nodes for performing a first calculation for one independent signal input and outputting a plurality of calculation processing results, and a plurality of input nodes in the input processing means. From a plurality of intermediate nodes in the intermediate processing means having a plurality of at least one intermediate node for outputting a plurality of arithmetic processing results and outputting a plurality of arithmetic processing results A time series output from an unknown system having a multilayer perceptron having an output processing means having a plurality of output nodes for inputting a plurality of operation processing results and performing a third operation and outputting one output signal The signal is sampled at sampling points having a certain time interval, differential vectors at the sampling points of the time series information are sequentially calculated from the sampling values, and temporarily stored, The significance of the sampling value of the time series information at the sampling point based on the differential vector of the time series information at the sampling point before the sampling point and the differential vector of the time series information at the sampling point. And a sampling value determined to be significant in the significance determination is used as a teacher signal of a multilayer perceptron, and a teacher signal sampling method in a multivariable system for identifying an input-output relationship of an unknown system. 4. The method of sampling a teacher signal in a multivariable system according to claim 3, wherein the significance of the sampling value of the time series information is determined by a differential vector of the time series information at the sampling point,
The angle formed by the differential vector of the time-series information at the sampling points before the sampling point is calculated based on the value of the inner product of these two differential vectors, and the angle is equal to or greater than a certain value. A sampling method for teacher signals in a characteristic multivariable system. 5. An input processing means having a plurality of input nodes for performing a first calculation on one independent signal input and outputting a plurality of calculation processing results, and a plurality of input nodes in the input processing means. From a plurality of intermediate nodes in the intermediate processing means having a plurality of at least one intermediate node for outputting a plurality of arithmetic processing results and outputting a plurality of arithmetic processing results A multilayer perceptron having an output processing means having a plurality of output nodes for outputting a single output signal by inputting a plurality of arithmetic processing results and outputting a single output signal is provided. Means for sampling at a sampling point, means for sequentially calculating a differential vector of the time series information at the sampling point from the sampling value, and temporarily storing the differential vector; Based on the differential vector of the time series information at the sampling point before the sampling point and the differential vector of the time series information at the sampling point, the significance of the sampling value of the time series information at the sampling point is shown. A teacher signal sampling apparatus in a multivariable system, comprising: a determining unit; and a unit that uses a sampling value determined to be significant as a teacher signal of a multilayer perceptron in the significance determination. 6. The teacher signal sampling device in the multivariable system according to claim 5, wherein the means for determining the significance of the sampling value of the time series information is a differential vector of the time series information at the sampling point, Means for calculating the angle formed by the differential vector of the time-series information at the sampling points before the sampling point based on the value of the inner product of these two differential vectors, and detecting that the angle is a certain value or more. A sampling device for a teacher signal in a multivariable system characterized by: