JPH05282270A - Method and device for constructing neural net - Google Patents

Abstract

PURPOSE:To execute learning by using only the data of good quality, and to attain the high speed of the learning by providing a data selecting function to adopt, reject, and select the data on the basis of a standard set beforehand. CONSTITUTION:The learning is executed in a neural net 1000, and the data from a data base 2000 and on-line data 3000 are screened by the data selecting function 4000 by excluding the data lacking in information necessary to an event to be intended to achieve in the neural net 1000, and only the necessary data is sent to an automatic data generating function 5000. In the automatic data generating function 5000, only the information necessary to the function intended to achieve in the neural net is inputted to the neural net 1000 as digitized information in a range usually from '0' to '1'. Besides, the data required as the learning data of the output of the function intended to achieve at the network is outputted to a teacher data base 6000 by the automatic data generating function 5000.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for constructing a neural network and a neural network system.

[0002]

2. Description of the Related Art Although there are few papers that describe learning methods for neural networks, the Institute of Electrical Engineers of Japan, Volume B, 110
Vol. 12, No. 12, p995-998 (December 1990), it is discussed in the neural network for the electric power system of Sasaki (hereinafter referred to as prior art 1).
In addition, Trans. Of IEEE, PES, Summer Meeting (1990,
6) "Electric Load Forecasting Using AnArti
ficial Neural Network "(hereinafter referred to as Prior Art 2). The outline will be described below. Prior art 1 describes that the following procedure is required as a general procedure when applying a neural network to one problem.

(I) Appropriate selection and clustering of input signals (ii) Determination of neural network structure-input layer,
Determine the number of neurons in the middle layer and output layer (iii) Prepare training data and verification data (iv) Learning by training data (v) Verification of learning results by verification data For this reason, the neural network manually selects and prepares data. , Learning and verification were done, and the composition was decided by the case-by-case correspondence.

In the prior art 2, the maximum, minimum and average temperatures of the day are input to the neural net to predict the peak demand value of the day or the total demand value of the day. Further, the prediction target time and the estimated temperature at that time, the demand actual value and the temperature one hour and two hours before, are input, and the demand value at the target time is predicted as an output value.

[0005]

The data stored in the database includes not only the data that can be used for input, but also the data due to an input error. The first issue is how to select the data to be used from these.

The second problem is how to create training data and verification data that satisfy the conditions set from the selected data.

Further, the third problem is how to determine the structure of the neural network which is currently handled on a case-by-case basis from the prepared data. It is an object of the present invention to provide a method of constructing a neural network or a neural network device capable of solving these problems.

[0008]

In order to solve the first problem, the present invention is characterized by providing a data selection function for selecting / sorting data based on a preset standard.

In order to solve the second problem, the data selected by the data selection function for dealing with the first problem from the prepared database is analyzed, clustered, and training and training are performed from the data. It is characterized by having a function of selecting data for verification.

Further, in order to solve the third problem, an event in which the neural network achieves the data selected by the data selection function for responding to the first problem based on the preset standard. Is characterized by having a function of performing analysis using a statistical method such as regression analysis or multivariate analysis of the correlation between the data stored in the database and the data stored in the database.

[0011]

The data selection function for solving the first problem is achieved by, for example, the following two types of selection methods.
One is achieved by detecting the default value when a default value, which is a preset value, is inserted in the list of information stored in the data stored in the database. To be done. Another example is
This is achieved by defining the accuracy of the event to be achieved by the neural network in advance. That is, by analyzing the data stored in the database using a statistical method such as regression analysis or multivariate analysis, only data that matches the required accuracy is selected, and data that deviates from it is not adopted. It is achieved by adopting means.

The data creating function for solving the second problem is to quantify the data selected by the data selecting function for solving the first problem, for example, by measuring the distance between the data by means of mean square error. Clustering according to the defined data similarity, classify the data so that each class and the data of the event to be achieved by the neural network correspond in advance, select representative data for each class, and perform learning This is achieved by storing the data for verification and verification in association with each other.

A means for solving the third problem is to solve the first problem by correlating the correlation between the event to be achieved by the neural network and the data stored in the database, based on a preset standard. From the data selected by the data selection function to correspond to, using statistical methods such as regression analysis and multivariate analysis beforehand, the data with a large correlation are used as factors in the input layer of the neural network. It is achieved by adopting.

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The basic construction of the present invention is shown in FIG. 1, and detailed examples of the basic construction are shown in FIGS.

The present invention comprises a neural network 1000,
Database 2000, online data 3000, data selection function 4000, data automatic creation function 500
0 is the basic function.

In the neural network 1000, the conventional technique 2 is used.
Learning is performed by the learning method represented by. Data from database 2000 and online data 30
00 is a data selection function 4000, which is a neural network 1
At 000, the data lacking the information necessary for the event to be achieved is excluded and only the necessary data is sent to the data automatic creation function 5000.

In the automatic data generation function 5000, only the information necessary for the function to be achieved in the network is normally input to the neural network 1000 as the digitized information 10 converted into the numerical information in the range of 0 to 1. .. Further, the data required as learning data for the output of the function to be achieved in the network by the automatic data generation function 5000 is output to the teacher database 6000.

At the time of learning, the output from the neural network 1000 and the output 20 from the teacher database 6000 are compared, and the weighting coefficient in the neural network 1000 is modified so that the difference between them is reduced. Neural net 1000 for all learning data
When the error is within the error to be achieved in, the learning ends, and the entire weighting coefficient at that time is the neural network 100.
It is stored in the memory of the computer that is performing the operation of the neural network 1000 as a weighting coefficient matrix that achieves the function of 0.

The verification of the event to be achieved by the neural network 1000 is compared with the output output from the teacher database 6000 through the similar process, and when the error is within the initial set value, it is verified. It At this time, the weight coefficient is not modified. Details of the data selection function 4000 are shown in FIG. In the data selection function 4000, it is desirable that the data appropriateness detection function 420 is a basic function and the data analysis function 440 is provided. Data adequacy detection function 4
In 20, the default value which is a value preset in the system and the user who uses it are listed in the list of information stored in the data stored in the database 2000 or in the information of the data in the online data 3000. When the specified default value is inserted, the default value is detected, and the data in which the default value is detected is excluded from the data sent to the neural network 1000.

On the other hand, in the data analysis function 440, the appropriate data 42 output from the data appropriateness detection function 420 is statistically processed by regression analysis or multivariate analysis with reference to the accuracy specified by the event to be achieved by the neural network in advance. By performing analysis using, only data that matches within the required accuracy is selected and output as analysis data 44. This allows you to sort out data based on preset standards.
It becomes possible to select, and it is possible to learn by using only qualitatively good data, so that learning can be speeded up.

Details of the automatic data generation function 5000 are shown in FIG.
Shown in. The data automatic creation function 5000 has a data similarity comparison function 520, a data correspondence function 540, and a data interface function 560 as basic functions. In the data similarity comparison function 520, the appropriate data 42 or the analysis data 44 output from the data selection function 4000 is quantified by the mean square error of the distance between the data and clustered according to the defined data similarity, which classification Is output as the similarity data 52 to which is added information indicating how much the data is deviated from the flag.

The data correspondence function 540 uses the similarity data 5
In response to 2, the data is organized according to flags indicating which classification, and the data is classified in advance in association with the data of the event to be achieved by the neural network for each flag.
Representative data is selected for each classification and is output as the correspondence data 54 in association with the learning and verification data.

In the data interface function 560,
In order to convert the corresponding data 54 into numerical information in the range of 0 to 1, the maximum value for each corresponding data 54 is a standardized value, or the data that is not numerical information is the standardized value. The converted information 56 is output to the input layer of the neural network 1000 or each unit of the teacher database 6000.

The standardized information can verify the output from the neural network 1000 and is used in conversion for comparison with the data of the event to be achieved by the neural network when it is put to practical use. By this, it is possible to analyze using only the data selected from the prepared database, clustering is performed, and the convergence is good as the data for training and verification,
Moreover, it is possible to select the same number of data from each class, and the learning convergence is improved.

Further, the automatic data generation function 5000 includes
It is desirable to have an input signal generation function 580.
The input signal generation function 580 includes a data selection function 4000.
Proper data 42 or analysis data 44 output from
Is input, and a flag 59 indicating that the number of input layer units 58 of the neural network 1000 and the number of input layer units 58 have been output from the input signal generation function 580.
Is output to the data creation function 540.

In the input signal generation function 580, the event to be achieved by the neural network and the database 2
Correlation with the data stored in 000 is automatically obtained using statistical methods such as regression analysis and multivariate analysis, and the number of data with a large correlation coefficient is adopted as the unit of the input layer of the neural network. The number of input layer units is 5
Output as 8.

On the other hand, in the data correspondence function 540, when the flag 59 indicating that the input layer unit number 58 is output from the input signal generation function 580 is input, the data and the neural network adopted in the input layer unit number 58 are input. Data classification is performed by associating with the data of the event to be achieved in 1., representative data is selected for each classification, and data corresponding to learning and verification data is associated with the corresponding data 54.
Output as. As a result, it is possible to automatically generate a unit in the input layer, which was conventionally set by the system engineer based on intuition and experience.

An embodiment of the present invention is shown in FIG.

FIG. 4 is an example of a system for predicting the demand for district heating and cooling. The district heating / cooling database 2100 stores actual demand data for district heating / cooling and the season, day of the week, and weather information at that time. On the other hand, in the district heating / cooling online data 3100, information such as weather, days of the week, and events is input as measurement values along with demand data momentarily by operator's key input or voice input.

The outputs from the district cooling / heating database 2100 and the district cooling / heating online data 3100 are sent to the district cooling / heating data selection function 4100. In the district cooling / heating data selection function 4100, the district cooling / heating database 21
00 and data from the district heating / cooling online data 3100. For example, if the default value of the system is input due to the operator forgetting to input the data x and y at that time, the correlation between them is as shown in FIG. That is, the default value is represented as a point on the x and y axes, and this is represented by the data appropriateness detection function 42 in the district heating / cooling data selection function 4100.
At 1, the data excluded as invalid data is sent to the data analysis function 441 as the proper data 44.

Assuming that the required specifications of the system have a prediction accuracy of ± 10%, the data analysis function 441 of FIG.
It is determined that the data within the range of ± 10% from the test curve of the actual data is adopted as the data of (1), and the data excluding the data outside this range is output as analysis data 44 to the district cooling / heating data automatic creation function 5100.

If the demand data is time-series data, the data similarity comparison function 520 in the automatic district heating / cooling data generation function 5100 finds the correlation of the data at each time in the same manner as in FIG. As a degree, each data is classified into classes by comparing the characteristic equation and deviation for each time, and is output to the district cooling / heating data corresponding function 540 as similarity degree data 52.

The district heating / cooling data correspondence function 540 receives the similarity data 52, classifies each of the similar data, and associates with the output required by the system, the demand value in this embodiment. For example, if the data group has a spread as shown in FIG. 6 with respect to the data of pattern 1, not only the center of the set but also the peripheral data is selected as the data representing pattern 1, It outputs to the district cooling / heating interface function 560 as the corresponding data 54 for training and verification, respectively.

Next, the district heating / cooling interface function 5
In 60, the maximum value for each corresponding data 54 for converting the corresponding data 54 into numerical information in the range of 0 to 1 is a standardized value or data that is not numerical information is set as numerical data. The standardized information 56 is output to the input layer of the district heating / cooling neural net 1100, or the standardized teacher data 57 is output to each unit of the district cooling / heating teacher database 6100. The standardized information can verify the output from the district heating / cooling neural network 1000, and is used in the conversion for comparison with the data of the event to be achieved by the neural network when it is put into practical use.

Further, the district heating / cooling data automatic generation function 51
District heating and cooling input signal generation function 5 installed in 00
At 80, the proper data 42 or the analysis data 44 output from the district heating / cooling data selection function 4100 is input, and the number of units 58 in the input layer of the district heating / cooling neural network 1100 and the number of input layer units 58 from the district cooling / heating input signal generation function 580 are input. Is output to the district heating / cooling data creation function 540. The district heating / cooling input signal generation function 580 performs regression analysis of the correlation between the event to be achieved by the neural network and the data stored in the district heating / cooling database 2100. For example, suppose that the following equation is used.

Y = 3.25x1 + 0.5x2-2.5x3
-0.01x4 + ... + anxn Here, y is a forecast demand and xn is an influential factor.

In this case, first, data x1 and x3 having a large correlation coefficient are adopted as the units of the input layer of the neural network, and 2 is output as the number of input layer units 58. On the other hand, in the district cooling / heating data correspondence function 540,
When the flag 59 indicating that the input layer unit number 58 is output from the district heating / cooling input signal generation function 580 is input, the data adopted in the input layer unit number 58 and the data of the event to be achieved by the neural network are input. The data is classified in correspondence with each other, representative data is selected for each classification, and the data for learning and the data for verification are output as the corresponding data 54. If the learning does not converge well with this, x2, x
4 is adopted as a new additional unit and learning is performed again. In this way, by sequentially adding possible influential factors, the number of units is automatically added to improve learning accuracy and convergence.

By the above procedure, data is automatically selected / generated from the database and the number of units in the input layer is generated, so that the learning speed can be increased or the convergence can be improved. FIG. 7 shows a modification of the present invention. This drawing differs from FIG. 4 in that a district heating / cooling neural network 1200 is separately provided. If the data selected by the district heating / cooling data selection function 4100 is data that will be in a different class when examined in detail, this is output to the district cooling / heating data automatic creation function 5100 as another analysis data 45. The analysis data 45 is learned as the signal of the district heating / cooling neural network 1200 in the same procedure as shown in FIG.
Verification is done.

By doing so, another neural net can be prepared even for data that is difficult to classify in data selection, and a plurality of neural net configurations can be automatically generated.

Even if one neural network cannot deal with it, a plurality of neural networks can deal with it.

[0041]

According to the present invention, data can be sorted or selected based on preset criteria, and only high quality data can be used for learning, and learning can be speeded up. ..

From the prepared database, good convergence as training and verification data, and
It is possible to select the same number of data from each class, which improves the convergence of learning.

Conventionally, a unit in the input layer, which has been set by the system engineer based on intuition and experience, can be automatically generated.

[Brief description of drawings]

FIG. 1 is a basic configuration diagram of the present invention.

FIG. 2 is a detailed configuration diagram of a data selection function.

FIG. 3 is a detailed configuration diagram of a data automatic generation function.

FIG. 4 is a neural network configuration diagram showing an embodiment of the present invention.

FIG. 5 is a correlation diagram of data.

FIG. 6 is a distribution chart of a data group.

FIG. 7 is a neural network configuration diagram showing a modified example of the present invention.

[Explanation of symbols]

1000 ... Neural network, 1100 ... Another neural network, 2000 ... Database, 3000 ... Online data, 4000 ... Data selection function, 5000 ... Data automatic generation function, 6000 ... Teacher database.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Masahiro Yoshioka 5-2-1 Omika-cho, Hitachi-shi, Ibaraki Hitachi Ltd. Omika factory

Claims (7)

[Claims] 1. A method for selecting / selecting data to be used for learning and verification depending on whether or not a preset criterion is satisfied in a method of constructing a multilayer neural network having at least an input layer, an intermediate layer and an output layer. A method for constructing a neural network, characterized by comprising: 2. A method of constructing a multi-layered neural network having at least an input layer, an intermediate layer and an output layer, a function for selecting / selecting data to be used for learning and verification depending on whether or not a preset standard is satisfied. And a function of accumulating the data, a method of constructing a neural network. 3. A multi-layer neural network device having at least an input layer, an intermediate layer and an output layer, and means for selecting / selecting data to be used for learning and verification depending on whether or not a preset standard is satisfied. An apparatus for a neural network, comprising: a database that stores the data. 4. A method for constructing a multi-layered neural network having at least an input layer, an intermediate layer, and an output layer, having a function of discarding / selecting data to be used for learning at least according to a preset convergence condition. Neural network configuration method. 5. A multi-layer neural network device having at least an input layer, an intermediate layer and an output layer, means for selecting / selecting data to be used for learning based on at least a preset convergence condition, and means for accumulating the data. And a neural network device. 6. A neural network device having at least an input layer, an intermediate layer, and an output layer, at least discarding data used for learning according to a preset convergence condition.
A neural network device comprising a selecting means and a database accumulating the data. 7. A neural network apparatus having at least an input layer, an intermediate layer and an output layer, means for selecting / selecting at least data to be used for learning according to preset learning conditions and convergence conditions, and a structure of the neural network. A neural network device having a changing unit capable of automatically changing a condition.