JP3184388B2 - Time series data generator - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for generating time-series data, and more particularly to a technique for generating time-series data indicating characteristics as close as possible to desired characteristics.

[0002]

2. Description of the Related Art A natural wind that a human feels comfortable like a wind blowing on a plateau includes "fluctuation". Heretofore, attempts have been made to realize this "fluctuation" by setting the wind speed and temperature of a fan, an air conditioner, and the like. In general, it is known that time-series frequency components such as wind speed and temperature in the natural world exhibit so-called “1 / f fluctuation” characteristics. It is more comfortable for humans to control the wind speed or temperature with 1 / f fluctuations than to control the wind speed or temperature at a fan or air conditioner, etc., at a constant wind speed or constant temperature. I know.

[0003] As a conventional example for engineeringly realizing such 1 / f fluctuation, there is, for example, a time-series data generation device as described later. This device generates time-series data with fluctuations by applying chaos theory. The device uses equation (1) as the conversion function.

[0004]

X _{n + 1} = 2 (x _n ) ² + x _n (0 <x _n ≦ 0.5) 2x _n −1 (0.5 <x _n <1) (1) Input the previous output, such as inputting an initial value x ₀ (0 <x ₀ <1) to obtain an output value x ₁ , and using this x ₁ as a next input value to obtain an output value x _2. By repeating the process for obtaining the current output, time series data _xn showing 1 / f fluctuation characteristics is generated. FIG. 4 shows an example of the time-series data generated by this device. In the drawing, the horizontal axis represents the transition of time, and the vertical axis represents the magnitude of the output value. FIG. 5 shows natural wind data as a comparative example. The vertical axis in FIG. 5 represents the magnitude of the wind speed value.

[0005]

The frequency component of data generated using the above-described time-series data generation device is 1 /
f shows fluctuation characteristics. However, as can be seen by comparing FIGS. 4 and 5, the phase of the output time-series data is significantly different from the phase of the wind speed value of the natural wind. Even when the frequency component of the time-series data shows the 1 / f fluctuation characteristic, if the phase characteristic is different from that of the natural wind, a fan or an air conditioner controlled using the data is used. Artificial wind generated by the human being cannot provide sufficient comfort to humans. Therefore, a technology for generating time-series data having a 1 / f fluctuation characteristic, which is a characteristic of wind speed and temperature in the natural world, and having a phase similar to that of the natural world is now desired.

Further, there is a problem in accuracy in data processing that hinders the engineering realization of the wind speed and temperature in the natural world. Generally, the conversion function of equation (1) is programmed and executed by a computer system included in the time-series data generation device. When the calculation accuracy of the computer system is high, a section in which the same data is not repeatedly output can be lengthened in the sequentially generated data. However, when the calculation accuracy is low, “digit loss” occurs, and during the generation of the time-series data, the input data having the same value as the previously input value is generated after a relatively short period. May occur. In that case, the time-series data output after that point is a repetition of the previously output time-series data. As described above, when data repetition occurs, the data conversion function provided in the time-series data generation device is not sufficiently performed, and the characteristics of the output data may be different from those theoretically required.

SUMMARY OF THE INVENTION The present invention has been made to solve such a problem, and an object of the present invention is to provide time-series data capable of generating time-series data indicating characteristics as close as possible to desired characteristics. It is to provide a generating device.

[0008]

In order to achieve the above-mentioned object, the invention according to claim 1 is for learning the fluctuation characteristics of natural wind.
Using a neural network習済, a first generating means to convert the input data to generate output data, the second generating means for generating time-series data, the output from the first generation means It is provided with data and third generating means for generating data to be supplied to the first generating means based on the output data from the second generation means, when data is output from the first generation means Forming the series data, the second
Generating means includes storage means for storing data, the data read from the storage means, functions to generate the chaotic value
And a converting means for converting using a series data generating device when and storing in conjunction with the storage means outputs the output data from the conversion means as the second time series data.

The invention described in claim 2 is the same as the invention described in claim 1
A sequence data generation device, wherein the first generation means is a natural data generation device.
Includes a neural network that has learned the wind fluctuation characteristics.
No.

[0010]

[0011]

[0012]

According to the configuration of the first aspect, when the neural network has learned the fluctuation characteristics of the natural wind, and the initial input data is input to the first generating means, the first generating means starts the initial processing. The input data is converted by a neural network to generate output data. The second generation means generates time series data. The third generating means includes the first generating means.
Then, data to be given to the first generation unit is generated based on the data output from the generation unit and the data output from the second generation unit. A data group output from the first generation unit through a series of these processes forms time-series data. Further, the second generating means converts the data read from the storage means by the converting means into a chaotic value using a predetermined function including a function for generating a chaotic value. The chaos value output from the conversion means is output as time-series data and stored in the storage means. The time series data stored in the storage means is read out by the conversion means. By repeating these series of processes,
The second generation means sequentially outputs time-series data.

According to the second aspect of the present invention, the first generating means is provided.
The stage learns the fluctuation characteristics of natural wind from the input data.
The output data is converted by the neural network
Generate.

[0014]

[0015]

[0016]

An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a time-series data generation device according to the present embodiment. The time-series data generation device (hereinafter simply referred to as “device”) 1 is a device that generates data for controlling the wind speed of an air conditioner (not shown). The device 1 includes a first conversion unit 11 including a neural network configured to learn a 1 / f fluctuation characteristic of a natural wind and predict a wind speed value at a next time based on data up to the current time. Further, the device 1 has a first
A first storage unit 13 for storing input data of the conversion unit 11, a first initial value storage unit 15 for storing an initial value of data to be stored in the first storage unit, An output unit 17 for outputting the data output from the conversion unit 11 to the outside of the device 1.

The apparatus 1 further includes a second converter 21 for converting the input data using a predetermined conversion function.
A second storage unit 22 for storing input data of the second conversion unit 21; and a second initial value storage unit for storing an initial value of data stored in the second storage unit 22. 23. The second conversion unit 21 generates output data based on the input data read from the second storage unit 22. The output from the second conversion unit 21 is replaced with the data stored in the second storage unit 22 at that time. Thereby, the second conversion unit 21 repeats the process of generating the next output data using the previous output data as input data.

The apparatus 1 further includes a third converter 31 for converting the output from the second converter 21 using a predetermined conversion function, and a third converter for converting the output from the first converter 11 into a third. And an addition unit 32 for adding the output from the conversion unit 31 and replacing the output with data stored in the first storage unit 13.

The first storage unit 13 stores four time-series data x _{1 to} x ₄ . Also, the first initial value storage unit 15
Stores the initial value x _I1 ~x _I4 time series data x ₁ ~x _4. When a drive command signal is input from outside the device 1, the initial values x _{I1 to} x _I4 stored in the first initial value storage unit 15 are sent to the first storage unit 13 in response to the input signal. Can be First converting section 11 reads the initial data x _I1 ~x _I4 stored in the first storage unit 13, as input data it, by treatment with a neural network, which will be described later, outputs the output data z I do.

The data output from the first conversion unit 11 is output to the outside via the output unit 17 and sent to the addition unit 32. The adding unit 32 includes the first converting unit 11
The output data w (described later) from the third conversion unit 31 is added to the output data from, to generate new data, and the data is replaced with the data stored in the first storage unit 13. As a result, the first conversion unit 11 repeats the process of generating the next output data using the data obtained by adding the output data from the third conversion unit 31 to the previously output data as input data. The data group generated by this repetitive processing forms time-series data.

FIG. 2 is a schematic diagram showing the configuration of the neural network included in the first conversion unit 13. The neural network 41 includes an input layer 51 and an intermediate layer 53.
And an output layer 55. The input layer 51 is composed of four input units 61a to 61d, the intermediate layer 53 is composed of three intermediate units 63a to 63c, and the output layer 55
Is composed of one output unit 65.

A method of learning the neural network 41 will be described. As an example, a case where the wind speed value of the natural wind shown in FIG. 5 is learned will be described. The time series data shown in FIG.
(T). The input units 61a to 61d of the neural network 41 receive v (t−3) and v (t
-2), v (t-1) and v (t) are input. further,
The neural network 41 sends v (t
+1), and the neural network 41 learns the wind speed value by the generally known “error back propagation method”.

In this way, the neural network 41 that has learned the data at all points in time becomes a system capable of predicting the wind speed value at the next point in time based on the data up to this point. Then, by using the neural network 41, the output value with respect to the current input is fed back to the input side, and the learned time-series data can be approximately obtained.

For example, in FIG. 1, the first storage unit 1
The x ₁ ~x ₄ stored in the 3 inputs to the input unit 61a~61d shown in FIG. 2, through the intermediate layer 53, the output data z from the output unit 65 of the output layer 55
Is output.

Next, the first conversion unit 1
The added value w added to the output data z from 1 will be described. The addition value w is calculated by the second conversion unit 21 as described above.
Is the data converted and output by the third conversion unit 31 from the output from. Equation (2) below is a conversion function used in the second conversion unit 21.

[0027]

In Equation 2] _{y n + 1 = 4 * y} n (1-y n) ... (2) the second storage unit 22, data y _n in the formula (2) is stored. The second initial value storage unit 23 stores y
_n Initial value y _I is stored in. When the drive command signal is inputted from the outside, the initial and y _I stored in response to the input signal to the second initial value storage 23 is input to the second storage unit 22 as y _n . Second converter 21
Reads data y _n of the second storage unit 22, Equation (2)
To obtain output data y _{n + 1} . The range of the initial values y _I stored in the second initial value storage 23 is a 0 <y _I <1.

Output data y _{n + 1} from the second converter 21
, Together they are sent to the third converter 31 is stored in the second storage unit 22 as input data y _n of the next conversion. Thereby, the second conversion unit 21 repeats the conversion processing of the equation (2) and sequentially outputs data yn _{+ 1} .

It is known that the time series data formed by the data yn _{+ 1} group is "chaotic" data.
Here, chaos means that a set of data has such a variety that it cannot be predicted by a deterministic mechanism. The expression of the conversion function used in the second conversion unit 21 is not limited to the expression (2), and may be another conversion expression that generates chaotic data.

[0030]

W = c * y _{n + 1} (3) Expression (3) is a conversion function used in the third conversion unit 31. c is a constant, and in this embodiment, c = 0.2
5 is assumed. The third conversion unit 31 converts the data yn _{+ 1} output from the second conversion unit 21 by the conversion formula of Expression (3) and outputs output data w. As described above, the output data w is a value obtained by multiplying the data y _{n + 1} by a constant value, and forms chaotic time-series data similarly to the data y _{n + 1} .
Note that the value of the constant c is not limited to 0.25,
An appropriate value may be set based on the relationship between z and yn _{+ 1} .

The adder 32 adds w output from the third converter to z output from the first converter 11, and inputs this z + w to the first storage 13. The first storage unit 13 replaces the value of x ₃ to x ₄ of that time, replace the value of x ₂ in x _3, it replaces the value of x ₁ to x _2. Then, replace the z + w, which is input to x _1. Thereby, the time-series data x stored in the first storage unit 13
₁ ~x ₄ is updated to the next input data.

As described above, the z output from the first conversion unit 11 is not directly input to the first storage unit 13 but the addition unit 32
The reason why the data to which w is added is used for the following reason.

The first conversion unit 11 is configured to generate time-series data by repeating the processing using the output data as the next input data. However, the neural network 41 included in the first conversion unit 11
Depending on the accuracy of the data (see FIG. 2), the output data z
May be the same as the value of the input data used previously. If the input data is the same, all output data after that point will be a repetition of the previous data, and the characteristics learned by the neural network 41 may not be sufficiently reflected in the output data.

In order to prevent such a problem, in the present embodiment, the input data x _{1 to} x ₄ of the first storage unit 13 is updated using data obtained by adding the addition value w to the output data z. I made it. This addition value w is the second value as described above.
Is a value obtained by multiplying the chaotic time-series data y _{n + 1} generated by the conversion unit 21 using equation (2) by a certain value.

Equation (2) is generally called a logistic function, and it is known that if there is infinite calculation accuracy, time-series data without any repetition can be generated indefinitely. Therefore, the input data z + w is
The frequency with which the same data as previously used can be minimized.

However, in practice, since the calculation accuracy is limited, the same value as the value that appeared before may appear during the generation of the time-series data. When the same value appears, the time-series data output thereafter becomes a repetition of the previously output data.

Here, the relationship between the maximum number of time-series data without repetition and the practicality of the apparatus 10 will be described. For example, when equation (2) is calculated with 8-bit precision,
Up to 30 time series data without repetition can be continuously generated. Similarly, the maximum number is 262 for 16-bit precision, and 7,000 or more for 24-bit precision. On the other hand, the calculation accuracy of the neural network in the first conversion unit 11 does not affect the maximum number of data without repetition of time-series data.

It is assumed that the time-series data generated in this way is output to the outside via the output unit 14, and one data value per 8 seconds is used as data for controlling the wind speed of the air conditioner. As described above, according to the 24-bit operation, 7,000 pieces of non-repeated data can be continuously generated. This is a large amount of data corresponding to about 15 hours of control data. Therefore, the periodicity of the wind speed control due to the repetition of the data output from the device 1 is as follows:
Almost negligible.

If the calculation accuracy of the second conversion unit 21 and the third conversion unit 31 with a small amount of calculation is sufficiently increased, the calculation accuracy of the first conversion unit 11 with a large calculation amount may be low. Thus, the calculation amount of the entire device 10 can be reduced.

FIG. 3 is a characteristic diagram showing an example of the time series data output from the device 1. The time-series data output from the device 1 is output so as to exhibit the 1 / f fluctuation characteristic of the natural world by learning of the neural network 41 as described above. Further, the phase is very similar to the phase of the wind speed value of the natural wind shown in FIG. 5, and by controlling the air conditioner using the time-series data output from the apparatus 1, the phase of the natural wind is minimized. It is possible to generate a near artificial wind.

The first conversion section 11 and the first storage section 13
These form the first generating means described in the claims. The second conversion unit, the second storage unit 22, and the second initial value storage unit constitute a second generation unit according to the present invention. The third converter 31 and the adder 3
2 constitutes a third generating means described in the claims.

The second storage section 22 constitutes the storage means described in the claims. The second conversion unit 21 constitutes a conversion unit described in claims.
The conversion function shown in the expression (2) corresponds to the predetermined function and the function of generating a chaos value in the claims.

The chaotic data corresponds to a chaotic value described in the claims. The 1 / f fluctuation characteristic corresponds to the fluctuation characteristic of a natural wind described in the claims.

[0044]

As described above, according to the first aspect of the present invention, the data input by the first generation means is converted by the neural network in which the fluctuation characteristics of the natural wind have been learned and output. . Based on the data output from the first generator and the time-series data output from the second generator, data to be subsequently provided to the first generator is generated. Therefore, while converting and outputting data reflecting the natural wind jaggedness learned by the neural network, the frequency at which the data input to the first generation means becomes the same as the previously input data is minimized. Becomes possible. In addition, the chaos value is converted and output using a predetermined function including a function for generating a chaos value, and the output chaos value is stored and read out in the next conversion processing. By repeating the chaos value output using the chaos value, the desired characteristics can be obtained.
Generate time-series data showing characteristics that are as close as possible to
Can be provided.

According to the second aspect of the present invention, the first generation
The means may generate data indicative of natural wind fluctuation characteristics.
Can be. As a result, characteristics that are as close to desired characteristics as possible
Time series data that can generate time series data
A data generation device can be provided.

[0046]

[0047]

[0048]

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a time-series data generation device according to an embodiment of the present invention.

FIG. 2 is a schematic diagram illustrating a configuration of a neural network used in the time-series data generation device.

FIG. 3 is a characteristic diagram illustrating an example of data output from a time-series data generation device.

FIG. 4 is a characteristic diagram showing an example of data output from a conventional time-series data generation device.

FIG. 5 is a characteristic diagram illustrating an example of time-series data of a wind speed value of a natural wind.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Time series data generation apparatus 11 1st conversion part 13 1st storage part 15 1st initial value storage part 17 Output part 21 2nd conversion part 22 2nd storage part 23 2nd initial value storage part 31 Third conversion unit 32 Addition unit 41 Neural network 51 Input layer 53 Intermediate layer 55 Output layer 61a, 61b, 61c, 61d Input unit 63a, 63b, 63c Intermediate unit 65 Output unit

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-3-233499 (JP, A) JP-A-57-71047 (JP, A) JP-A-4-335730 (JP, A) JP-A-5-71047 158662 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06N 3/00 G06N 3/02 G10H 5/00 G10H 7/00 G06F 7/58 F24F 11/02

Claims (1)

(57) [Claims] 1. A fluctuation characteristics of natural wind using neural <br/> Le network learned, a first generating means for generating output data converts the input data, time-series data A second generating unit for generating, and a third generating unit for generating data to be provided to the first generating unit based on data output from the first generating unit and data output from the second generating unit. Means for generating
The data output from the first generating means forms time-series data, The second generating means stores data for storing the data, and outputs the data read from the storing means as a chaotic value. and a converting means for converting with the functions that produce a storing output data from said converting means into said memory means and outputting the results as the second time series data
A time-series data generation device characterized by the following .