JP2021135896A - Learning data generator and learning data generation method - Google Patents

Abstract

An object of the present invention is to efficiently provide learning data with appropriate content for a required period. Kind Code: A1 A learning data generation device connects a plurality of duplicated data, which are data obtained by duplicating generation source data, which is time-series data for a predetermined period, and gives noise to each of the duplicated data to obtain a desired period of time. Artificial data, which is time-series data for a period corresponding to , is generated, and learning data used for learning a machine learning model is generated using the artificial data. The learning data generation device generates, for example, artificial data in a period past the generation source data, and links the artificial data to the generation source data in the past direction to generate learning data. [Selection drawing] Fig. 6

Description

The present invention relates to a learning data generator and a learning data generation method.

When realizing a machine learning system, it is necessary to prepare effective learning data in order to ensure the accuracy of the machine learning model.

As a technique for generating learning data, for example, Patent Document 1 describes a learning data generation device for learning a determination device using a neural network. The training data generator changes the data value of the collected time-series data, changes the time interval of each data of the time-series data, adds distortion to the time-series data, and adds noise to the time-series data. ..

Further, Patent Document 2 describes a technique for generating data that contributes to improvement of learning by processing the learning data when the learning data is small. Specifically, the neural network learning device extracts features from the training data using the learning neural network, generates hostile features from the extracted features using the learning neural network, and is hostile to the training data. The recognition result of the neural network is calculated using the characteristics, and the neural network is trained so that the recognition result approaches the desired output.

Further, Patent Document 3 describes an abnormality detection system configured for the purpose of promptly detecting a state abnormality of a monitoring target. The anomaly detection system collects observation data for the monitoring target, saves it as time-series observation data, classifies the observation data into either training data or verification data, and based on the training data, the linear state space of the monitoring target. When the model parameters of the model are identified, the estimated value of the probability distribution of the state variable to be monitored is calculated by inputting the model parameters and the verification data, the degree of abnormality of the monitored target is calculated based on the estimated value, and the observation data is collected. The newly collected observation data is added to the time-series observation data, and if the number of time-series observation data is larger than the threshold value, the earliest collected observation data is discarded.

JP-A-2019-87106 International Publication No. 2018/167900 Japanese Unexamined Patent Publication No. 2019-191836

RB Cleveland, 3 outsiders, “STL: a seasonal-trend decomposition procedure based on loess”, [online], 1990, Journal of official statistics, [Search January 31, 2020], Internet <URL: https: / /www.wessa.net/download/stl.pdf>

In order to realize a machine learning system that performs inference processing such as predictive diagnosis and abnormality detection based on time series data, it is necessary to ensure the accuracy of the machine learning model that performs the above inference processing, and for that purpose, effective learning data is efficient. You need to prepare well. In addition, in order to ensure the accuracy of the machine learning model, it is necessary to prepare time-series data for the period required for that purpose as learning data.

However, neither Patent Document 1 nor Patent Document 2 discloses any technique for generating time-series data for a required period. Further, in the technique described in Patent Document 3, it is necessary to collect observation data for a monitored object, and it is not possible to deal with a case where observation data is not obtained, for example, when a machine learning system is introduced.

An object of the present invention is to provide a learning data generation device and a learning data generation method capable of efficiently providing learning data having appropriate contents for a required period.

One of the present inventions for achieving the above object is a learning data generator which is configured by using an information processing device and generates learning data used for learning a machine learning model, and is a learning data generator for a predetermined cycle. By concatenating a plurality of duplicated data which are duplicated data of the generation source data which is serial data and giving noise to each of the duplicated data, artificial data which is time-series data of a period corresponding to a required period is generated. An artificial data generation unit for generating data and a learning data generation unit for generating training data using the artificial data are provided.

In addition, the problems disclosed in the present application and the solutions thereof will be clarified by the column of the form for carrying out the invention and the drawings.

According to the present invention, it is possible to efficiently provide learning data having appropriate contents for a required period.

It is a figure which shows the schematic structure of the machine learning system. This is an example of the configuration of an information processing device. It is a figure which shows the main function which a learning data generation apparatus has. It is a flowchart explaining the learning data generation process. It is a flowchart explaining the detail of the artificial data generation processing. It is a figure which showed typically the data generated in the execution process of an artificial data generation process. This is an example of source data. This is an example of artificial data. It is a flowchart explaining the detail of the learning data period setting process. This is an example of observation data. It is a flowchart explaining the detail of the learning data generation process. This is an example of training data. It is a flowchart explaining the artificial data generation processing in 2nd Embodiment. It is a figure which shows typically the data generated in the execution process of the artificial data generation process in 2nd Embodiment. This is an example of source data. This is an example of intermediate data. This is an example of replication source data. This is an example of artificial data.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In the following description, the same reference numerals may be given to configurations having the same or similar functions, and duplicate description may be omitted. Further, in the following description, the letter "S" prefixed with the reference numeral means a processing step. In addition, the term "learning data" is used in the following explanation, but it is synonymous with "training data". Further, the learning data used for so-called supervised machine learning includes so-called label information, but for the sake of simplicity, explanations and examples of labels are omitted in this embodiment. Further, in the following description, the period may be specified by the date and time, or may be specified only by the day or only the time.

[First Embodiment]
FIG. 1 shows a schematic configuration of an information processing system (hereinafter, referred to as “machine learning system 1”) to which the learning data generation device 100 shown as the first embodiment is applied. As shown in the figure, the machine learning system 1 includes an inference device 2 and a learning data generation device 100.

The inference device 2 has the functions of a learning processing unit 21 that learns the machine learning model 23 using the learning data 114, which is time-series data, and an inference processing unit 22 that performs inference processing using the machine learning model 23. .. The inference processing unit 22 performs inference processing by inputting the observation data 113, which is time series data, into the machine learning model 23, and outputs the result as the inference result 7. The machine learning model 23 performs inference processing for predictive diagnosis, abnormality detection, and the like based on time-series data, for example.

The learning data generation device 100 generates learning data 114 based on generation source data 111 and observation data 113, which are time-series data. The generated learning data 114 is input to the inference device 2 via communication or a recording medium.

FIG. 2 shows an example of the information processing device 10 used in the configuration of the inference device 2 and the learning data generation device 100. As shown in the figure, the illustrated information processing device 10 includes a processor 11, a main storage device 12, an auxiliary storage device 13, an input device 14, an output device 15, and a communication device 16. These are connected so as to be able to communicate via a communication means such as a bus.

The information processing device 10 is realized by using virtual information processing resources provided by using virtualization technology, process space separation technology, or the like, for example, a virtual server provided by a cloud system. May be good. Further, all or a part of the functions of the information processing device 10 may be realized by, for example, a service provided by a cloud system via an API (Application Programming Interface) or the like. Further, for example, the learning data generation device 100 may be configured by using a plurality of information processing devices 10 connected so as to be communicable. Software such as an operating system, a file system, and a DBMS (DataBase Management System) (relational database, NoSQL, etc.) may be installed in the information processing device 10.

The processor 11 includes, for example, a CPU (Central Processing Unit), an MPU (Micro Processing Unit), a GPU (Graphics Processing Unit), an AI (Artificial Intelligence) chip, an FPGA (Field Programmable Gate Array), an ASIC (Application Specific Integrated Circuit), or the like. Is configured using.

The main storage device 12 is a device that stores programs and data, and is, for example, a ROM (Read).
Only Memory), RAM (Random Access Memory), non-volatile memory (NVRAM (Non Volatile RAM)) and the like.

The auxiliary storage device 13 includes, for example, an SSD (Solid State Drive), a hard disk drive, an optical storage device (CD (Compact Disc), DVD (Digital Versatile Disc), etc.), a storage system, an IC card, an SD card, or an optical recording device. A reading / writing device for a recording medium such as a medium, a storage area for a virtual server, and the like. The auxiliary storage device 13 can read programs and data via the reading device of the recording medium and the communication device 16. Programs and data stored (stored) in the auxiliary storage device 13 are read out to the main storage device 12 at any time.

The input device 14 is an interface that accepts input from the outside, and is, for example, a keyboard, a mouse, a touch panel, a card reader, a voice input device, and the like. The output device 15 is an interface that outputs various information such as processing progress and processing results.

The output device 15 is, for example, a display device (liquid crystal monitor, LCD (Liquid Crystal Display), projector, etc.) that visualizes the above-mentioned various information, a device (audio output device (speaker, etc.)) that visualizes the above-mentioned various information. It is a device (printing device, etc.) that converts the above various information into characters.

The input device 14 and the output device 15 form a user interface. For example, the information processing device 10 may be configured to input / output information to / from another device (smartphone, tablet, notebook computer, various personal digital assistants, etc.) via the communication device 16.

The communication device 16 realizes communication with other devices. The communication device 16 is a wireless or wired communication interface that realizes communication with another device via a communication network, and is, for example, a NIC (Network Interface Card), a wireless communication module, or a USB (Universal Serial Bus). ) Modules, serial communication modules, etc. Subsequently, the functions provided by each device will be described.

FIG. 3 shows the main functions of the learning data generation device 100. As shown in the figure, the learning data generation device 100 includes a storage unit 110, an observation data acquisition unit 120, a generation source data acquisition unit 130, an artificial data generation unit 140, a learning data period setting unit 150, and a learning data generation unit 160. And each function of the learning data output unit 170 is provided. These functions are performed by the processor 11 of the information processing device 10 constituting the learning data generation device 100 reading and executing the program stored in the main storage device 12 of the information processing device 10 or by executing the information processing device. It is realized by the hardware (FPGA, ASIC, AI chip, etc.) included in 10.

Among the above functions, the storage unit 110 stores and manages the generation source data 111, the artificial data 112, the observation data 113, and the learning data 114. The storage unit 110 stores and manages each data as, for example, a database table provided by the DBMS or a file provided by the file system.

The generation source data 111 is data used for generating the artificial data 112. The artificial data 112 is data used for generating the learning data 114. The observation data 113 is data input to the machine learning model 23 after the machine learning system 1 starts the production operation, for example. The learning data 114 is data generated based on artificial data 112 and observation data 113, and is data used for learning (training) of the machine learning model 23.

Among the functions shown in FIG. 3, the observation data acquisition unit 120 acquires the observation data 113 from the inference device 2 via communication or a recording medium. The storage unit 110 stores the observation data 113 acquired by the observation data acquisition unit 120.

The generation source data acquisition unit 130 acquires or generates the generation source data 111. The generation source data acquisition unit 130 may, for example, generate source data 1 from the user via the user interface.
Get 11. The generation source data acquisition unit 130 generates the generation source data 111 based on the observation data 113, for example. The generation source data 111 may be generated by the user editing the observation data 113 via the user interface. The storage unit 110 stores the generation source data 111 acquired or generated by the generation source data acquisition unit 130.

The artificial data generation unit 140 generates artificial data 112 based on the generation source data 111. The storage unit 110 stores the artificial data 112 generated by the artificial data generation unit 140.

The learning data period setting unit 150 performs processing related to setting the period of the learning data 114 (from the start time to the end time of the learning data. Hereinafter, referred to as a “learning data period”). The learning data period setting unit 150 receives, for example, information regarding the setting of the learning data period from the user via the user interface.

The learning data generation unit 160 generates learning data 114 based on the artificial data 112 and the observation data 113.

The learning data output unit 170 outputs the learning data 114 generated by the learning data generation unit 160. The output learning data 114 is input to the inference device 2 via communication or a recording medium.

FIG. 4 is a sequence diagram illustrating a process (hereinafter, referred to as “learning data generation process S400”) performed by the learning data generation device 100 when generating the learning data 114. Hereinafter, the learning data generation process S400 will be described with reference to the figure. At the start of the process shown in the figure, the storage unit 110 has already stored the observation data 113 acquired by the observation data acquisition unit 120 and the generation source data 111 acquired or generated by the generation source data acquisition unit 130. It is assumed that there is.

As shown in the figure, first, the artificial data generation unit 140 reads out the generation source data 111 stored in the storage unit 110 (S411).

Subsequently, the artificial data generation unit 140 specifies the length of the period (hereinafter, referred to as “request period”) of the learning data 114 to be generated by the user via the user interface, and is included in the generation source data 111. It accepts the input of the number of cycles (hereinafter referred to as "the number of cycles") and the dispersion σ ^ 2 used for generating noise given to the artificial data 112 (S412).

Subsequently, the artificial data generation unit 140 generates artificial data 112 based on the read generation source data 111, the received request period, the number of cycles, and the variance σ ^ 2 (hereinafter, “artificial data generation process S413”). (S413).

Subsequently, the learning data period setting unit 150 reads out the observation data stored in the storage unit 110 (S421).

Subsequently, the learning data period setting unit 150 uses the period of the generation source data 111 read by the artificial data generation unit 140 in S412, the period of the artificial data 112 generated by the artificial data generation unit 140 in S413, and the observation read in S421. The period of data 113 is acquired (S422).

Subsequently, the learning data period setting unit 150 acquires the request period received by the artificial data generation unit 140 in S412 (S423).

Subsequently, the learning data period setting unit 150 sets the learning data period based on each period acquired in S422 and the request period acquired in S423 (hereinafter, referred to as "learning data period setting process S424"). (S424).

Subsequently, the learning data generation unit 160 reads out the generation source data 111 read by the artificial data generation unit 140 in S411, the artificial data 112 generated by the artificial data generation unit 140 in S413, and the learning data period setting unit 150 in S421. Based on the observed data 113, the training data 114 is generated for the training data period set by the training data period setting process S424 (S431).

After that, the learning data output unit 170 outputs the generated learning data 114. The output learning data 114 is transmitted (provided) to the learning processing unit 21 of the inference device 2 via communication or a recording medium.

FIG. 5 is a flowchart illustrating the details of the artificial data generation process S413 shown in FIG. Further, FIG. 6 is a diagram schematically showing data generated in the execution process of the artificial data generation process S413. The artificial data generation unit 140 duplicates the generation source data 111 for the number of cycles corresponding to the request period received in S412, replaces the date and time of the generation source data 111 with an appropriate date and time, and further adds noise to the observed value. As a result, artificial data 112 is generated. Hereinafter, the artificial data generation process S413 will be described with reference to FIGS. 5 and 6.

FIG. 7 shows an example of the generating source data 111. Hereinafter, the artificial data generation process S413 will be described using the generation source data 111 shown in the figure as an example. As shown in the figure, the illustrated source data 111 includes a plurality of entries (records) having each item of the date and time 701 and the observed value 702.

Among the above items, the date and time 701 is set to the date and time when the observed value 702 is acquired. The value of the date and time 701 is also used as an identifier for uniquely identifying each entry. The observed value is set in the observed value 702. The time series data may include categorical variable information, but unless otherwise specified, the observed values are assumed to be quantitative variable information. The observed value 702 is, for example, a value obtained by processing the value itself (raw data) acquired from a sensor device or the like or a value obtained from a plurality of observed objects (addition / subtraction / multiplication / division, aggregation processing, statistical processing, etc.). be. The above values are, for example, the communication volume and the operating rate when the observation target is an information communication system. Further, the above value is, for example, a value "total communication amount" which is the sum of the values of two observation targets, "uplink communication amount" and "downlink communication amount". Further, for example, the above value is obtained by calculation based on the observed value at a certain time point and the observed value at another time point. Further, for example, the above value is the difference between the previous communication amount and the current communication amount (time change amount of the communication amount).

The example generation source data 111 comprises the content of recording the information from 0:00:00 on November 15, 2019 to 0:00:00 on November 22, 2019 at 10-minute intervals, and is shown in FIG. It is the data for one cycle shown in A). In the following description, it is assumed that the cycle of the generating source data 111 received in S412 is one week, and 1008 entries are included in each cycle. Further, it is assumed that 28 weeks is accepted as the request period in S412.

As shown in FIG. 5, first, the artificial data generation unit 140 has a minimum value (hereinafter referred to as "minimum number of cycles") that is a multiple of the period of one cycle of the generation source data 111, which is a period equal to or longer than the request period received in S412. To be referred to) (S501).

Subsequently, the artificial data generation unit 140 obtains a value obtained by subtracting the number of cycles included in the generation source data 111 from the obtained minimum number of cycles, and divides the obtained value by the number of cycles included in the generation source data 111. The value obtained by rounding up the value after the decimal point is taken as the number of duplications (S502). The number of cycles included in the generating source data 111 is subtracted from the minimum number of cycles obtained in S501 in order to exclude the amount of the generating source data 111 of the duplication source from the number of duplications. Since the number of cycles of the generating data 111 illustrated is 1 and the required period is 28 weeks, 27 is obtained as the number of duplications in this example. The number of copies may be obtained by a method other than the above method. For example, the number of copies may be specified by the user via the user interface.

Subsequently, the artificial data generation unit 140 generates data (hereinafter, referred to as “replication data”) in which the generation source data 111 is duplicated as many times as the number of reproductions acquired in S502 (S503).

Subsequently, the artificial data generation unit 140 assigns natural numbers starting from 1 to each duplicate data in order. The storage unit 110 stores a number assigned to each duplicated data (hereinafter, referred to as a “duplicate number”) in association with each of the duplicated data (S504).

Subsequently, the artificial data generation unit 140 generates data (hereinafter, referred to as "primary artificial data") in which the duplicated data is concatenated in chronological order in the reverse order of the assigned duplicate numbers (S505).

Subsequently, the artificial data generation unit 140 duplicates the value of the date and time 701 of each entry of the generation source data 111 for each entry of the primary artificial data generated in S507 (hereinafter, referred to as “reference source date and time”). (S506).

Subsequently, the artificial data generation unit 140 updates the reference source date and time of each entry of the given primary artificial data to a value retroactive from the reference date and time (for example, date and time t in the figure) (generation source data from date and time t). The date and time of each entry is generated by multiplying the period of 111 by the duplication number of each entry to go back by the date and time obtained (S507). For example, the date and time after the change of 00:00 on November 15, 2019 in the duplicated data of the duplicate number 27 will be 00:00 on May 10, 2019, which is 27 weeks back. The primary artificial data generated by executing S507 is as shown in FIG. 6 (B).

Subsequently, the artificial data generation unit 140 generates white noise during the period of the artificial data using the variance σ ^ 2 received in S412 of FIG. 4 (S508). The white noise generated by executing S508 is as shown in FIG. 6 (C).

Subsequently, the artificial data generation unit 140 generates artificial data 112 by adding the white noise generated in S508 as a variable value to the primary artificial data (S509). The artificial data 112 generated by executing S509 is as shown in FIG. 6D.

FIG. 8 shows an example of the artificial data 112 generated by the artificial data generation process S413. As shown in the figure, the illustrated artificial data 112 includes a plurality of entries having the respective items of date and time 811, observed value 812, reference source observed value 813, variation value 814, reference source date and time 815, and replication number 816. ..

Among the above items, the date and time generated in S507 is set in the date and time 811. The value of the date and time 811 is also used as an identifier that uniquely identifies each entry. The observation value 812 is set to the observation value of the artificial data 112 generated in S509 at the relevant date and time. The observation value 702 of the generation source data 111 corresponding to the date and time is set in the reference source observation value 813. The value of the white noise generated in S508 and corresponding to the date and time is set in the fluctuation value 814. The date and time 701 of the generation source data 111 corresponding to the date and time is set in the reference source date and time 815. The value of the reference source date and time 815 indicates that the entry is based on the entry of the generator data 111 of the date and time of the value of the reference source date and time 815. The duplication number assigned in S504 is set in the duplication number 816.

FIG. 9 is a flowchart illustrating the learning data period setting process S424 shown in FIG. Hereinafter, the learning data period setting process S424 will be described with reference to the figure. When the observation data 113 for a period overlapping the request period acquired in S423 has already been acquired, the artificial data generation unit 140 sets the learning data period so that the observation data 113 is preferentially adopted as the learning data 114. Set.

FIG. 10 is an example of the observation data 113 used in the following description. As shown in the figure, the illustrated observation data 113 includes a plurality of entries having each item of the date and time 1011 and the observed value 1012. Of the above items, the date and time 1011 is set to the date and time when the observed value of the entry was acquired. The observed value actually acquired from the observed object is set in the observed value 1012.

As shown in FIG. 9, first, the learning data period setting unit 150 confirms whether or not the storage unit 110 stores the observation data 113 (whether or not the learning data generation device 100 has acquired the observation data 113). (S901). When the storage unit 110 stores the observation data 113 (S901: YES), the learning data period setting unit 150 sets the end point of the period of the observation data 113 as the end point t_end of the learning data period (S902). .. After that, the process proceeds to S904. On the other hand, when the storage unit 110 does not store the observation data 113 (S901: NO), the learning data period setting unit 150 sets the end time of the period of the generation data 111 as the end time t_end of the learning data period. (S903). After that, the process is S90.
Proceed to 4.

In S904, the learning data period setting unit 150 goes back to the past by the request period (the request period acquired by the artificial data generation unit 140 in S412) from the end point t_end of the learning data period set in S902 or S903 (hereinafter). , "Temporary start point tmp_Tstart") is acquired.

Subsequently, the learning data period setting unit 150 compares the period of the artificial data 112, the period of the generation source data 111, and the period of the observation data 113 with the provisional start time point tmp_Tstart (S905). When the temporary start time tmp_Tstart is in the period of the artificial data 112 (S905: during the period of the artificial data), the temporary start time tmp_Tstart is set in the start time t_start of the learning data period (S906). On the other hand, when the provisional start time tmp_Tstart is in the period of the generation source data 111 or the observation data 113 (S905: during the period of the generation source data or the observation data), the generation source data is set to the start time t_start of the learning data period. Set the start time point of 111 (S907)
..

By the above processing, the start time t_start and the end time t_end of the training data period are set, and the setting of the training data period is completed. By the processing of S902, S903, and S907, in the learning data generation processing S431, the observation data 113 or the generation source data 111 is adopted as the learning data 114 in preference to the artificial data 112.

FIG. 11 is a flowchart illustrating the learning data generation process S431 shown in FIG. Hereinafter, the learning data generation process S431 will be described with reference to the figure.

First, the learning data generation unit 160 acquires observation data 113, generation source data 111, and artificial data 112 for a period overlapping the learning data period set by the learning data period setting process S424 from the storage unit 110 (S1101- S1103).

Subsequently, the learning data generation unit 160 generates intermediate connection data in which the acquired generation source data 111 and the acquired artificial data 112 are connected in the time series direction (artificial data 112 and generation source data 111 are connected in chronological order). (S1104).

Subsequently, the learning data generation unit 160 generates the learning data 114 by connecting the intermediate connection data and the acquired observation data 113 in the time series direction (S1105). If the observation data 113 corresponding to the entire learning data period exists, the learning data 114 is all based on the observation data 113. When the observation data 113 overlaps with a part of the training data period, the observation data is based on the intermediate connection data from the start time of the training data period to the start time of the observation data 113 in the entire period of the training data 114. From the start time of 113 to the end of the training data period, the observation data 113 is used. When the observation data 113 exists in this way, the observation data 113 is preferentially adopted as the learning data 114. Therefore, the observation data which is the data actually acquired after the production operation of the machine learning system 1 is started. It is possible to quickly shift to an operational state in which only 113 is used as learning data 114 for learning.

FIG. 12 shows an example of the learning data 114 generated by the learning data generation process S431. The illustrated learning data 114 includes a plurality of entries having each item of the date and time 1201 and the observed value 1202. Among the above items, the date and time 1201 is set to a date and time based on any one of the date and time 811 of the artificial data 112, the date and time 701 of the generation source data 111, and the date and time 1011 of the observation data 113. The observation value 1202 is set to an observation value based on any of the observation value 812 of the artificial data 112, the observation value 702 of the generation source data 111, and the observation value 1012 of the observation data 113.

As described above, according to the learning data generation device 100 of the first embodiment, even when it is difficult to prepare the learning data for the period necessary for ensuring the accuracy of the machine learning model 23, the above period Effective learning data can be efficiently generated and provided without bothering the user.

Further, since the learning data generation device 100 generates the learning data 114 by using the artificial data 112 in which noise is individually added to each duplicated data, it has a variety that is expected to have an effect of suppressing overfitting of the machine learning model 23. The training data 114 can be generated, and the inference accuracy of the machine learning model 23 can be improved. Further, by adding white noise to the artificial data 112, it is possible to reproduce fluctuations close to actual fluctuations. For example, it is possible to improve the inference accuracy of the machine learning model 23 that functions on the premise that the observed data follows a normal distribution. Can be done.

Further, as shown in FIG. 9, in the training data period setting process S424, the training data period is set so that the observation data 113 is preferentially adopted over the training data 114, so that the actual operation of the machine learning system 1 can be performed. After the start, it is possible to quickly shift to the operational state of learning by using only the observation data 113, which is the actually acquired data, as the learning data 114. Therefore, the inference accuracy of the inference device 2 can be improved at an early stage after the start of the actual operation.

Further, the learning data generation device 100 can generate the artificial data 112 in a period earlier than the generation source data 111, and generates the artificial data 112 so as not to overlap with the period of the newly acquired observation data 113. For example, it is possible to prevent complicated processing such as replacing the artificial data 112 with the observation data 113.

In the above, the case where the artificial data 112 in the period past the generation source data 111 is generated has been illustrated, but the artificial data 112 in the future period than the generation source data 111 may be generated. Thereby, for example, the learning data 114 of the desired future time can be generated by using the past time series data at the time considered to best reflect the actual behavior as the generation source data 111. In this case, for example, in S504 of FIG. 5, a negative integer starting from -1 is assigned to each duplicate data to be artificial data 112 in the future period, and a positive integer assigned to each duplicate data in the past period is assigned. Make sure that the sum of the duplication number and the absolute value of the negative duplication number matches the number of duplications acquired in S502 of FIG. By doing so, the date and time (period) information can be easily calculated only by adding the value obtained by multiplying the period of the duplicated data by the duplicate number in S507 to the date and time as the reference.

[Second Embodiment]
Subsequently, the second embodiment will be described. The learning data generation device 100 of the second embodiment generates artificial data 112 based on the components (trends, periodic fluctuations, and residuals described later) obtained by decomposing the generation source data 111. The basic configuration of the machine learning system 1 of the second embodiment and the flow of processing executed in the machine learning system 1 are basically the same as those of the machine learning system 1 of the first embodiment described with reference to FIGS. 1 to 4. However, some of the functions of the artificial data generation unit 140 are different. Hereinafter, the parts different from those of the first embodiment will be mainly described.

FIG. 13 is a flowchart illustrating the artificial data generation process S413 shown as the second embodiment. Further, FIG. 14 is a diagram schematically showing data generated in the execution process of the artificial data generation process S413. Further, FIG. 15 is an example of the generator data 111 used in the following description. Hereinafter, the artificial data generation process S413 of the second embodiment will be described in detail with reference to these figures.

As shown in FIG. 14 (A), the illustrated source data 111 has a small cycle Tp (= 1 day) and a large cycle T (= 7 days) on November 15, 2019 at 0:00:00. It consists of data for 8 days at 10-minute intervals from 23:50:00 on November 22, 2019 (8 small cycle Tp (large cycle 7 days x 1 + small cycle 1 day)). In the following description, the start time point of the generating source data 111 is t. Further, in the following description, the request period received in S412 of FIG. 4 is 28 weeks. Further, in S412 of FIG. 4, it is assumed that one cycle (one large cycle) is accepted as the number of cycles of the generating source data 111.

As shown in FIG. 13, first, the artificial data generation unit 140 solves the user interface and accepts the inputs of the small cycle Tp (1 day) and the large cycle T (7 days) (S1301).

Subsequently, the artificial data generation unit 140 obtains a minimum value (minimum number of cycles) that is a multiple of the period of one cycle of the generation source data 111, which is a period equal to or longer than the request period (28 weeks) received in S412 (S1302). ..

Subsequently, the artificial data generation unit 140 obtains a value obtained by subtracting the number of cycles included in the generation source data 111 from the minimum number of cycles obtained in S1301, and the calculated value is included in the generation source data 111. The value divided by the number is rounded up to the nearest whole number, and the value obtained by further adding 1 is taken as the number of duplications (S1303). It should be noted that 1 is added because the period of the generated artificial data 112 satisfies the required period due to the time difference (Tp / 2 described later) caused by obtaining the trend described later by the moving average for the generating source data 111. This is because there is a possibility that it will disappear. In the case of this example, since the number of cycles of the generation source data 111 is 1 and the request period received in S412 is 28 weeks, 28 can be obtained as the number of duplications.

Subsequently, the artificial data generation unit 140 decomposes the generation source data 111 into components (trend, periodic fluctuation, residual) with the small cycle Tp as the fluctuation cycle (S1304). Here, the trend means an element (Trend component) representing a long-term fluctuation in time series data. In addition, periodic fluctuation refers to an element (Seasonal component) that appears periodically at regular intervals in time series data. The residual is a fine fluctuation element (Redidual component) that remains by excluding the trend and the periodic fluctuation in the time series data. In this embodiment, the above decomposition is described in Non-Patent Document 1 as STL (Seasonal-Trend Decomposition).
Procedure Based on Loess) shall be used, but the above decomposition method is not necessarily limited.

FIG. 14B is a component obtained by decomposing the generating data 111 of FIG. 14A. In the figure, (B-1) is a trend, (B-2) is a periodic fluctuation, and (B-3) is a residual.

FIG. 16 shows the data obtained in S1304 (hereinafter, referred to as “intermediate data 1600”). As shown in the figure, the intermediate data 1600 includes a plurality of entries having each item of date and time 1601, observed value 1602, trend 1603, periodic variation 1604, and residual 1605. In the figure, "-" indicates that data is missing. The date and time 1601 and the observed value 1602 correspond to the date and time 1201 and the observed value 1202 in the generating data 111. The trend 1603, the periodic variation 1604, and the residual 1605 are set with values indicating the trend, the periodic variation, and the residual, which are the components of the observed value 1602, obtained in S1304, respectively. In both the trend 1603 and the residual 1605, the values of half the period (= Tp / 2) of the small cycle specified when executing STL are missing at both ends of the period. In this example, the period from 0:00:00 on November 15, 2019 to 11:50:00 on November 15, 2019, and from 12:00:00 on November 22, 2019, 2019 The values of trend 1603 and residual 1605 are missing in the period until 23:50:00 on November 22nd.

Returning to FIG. 13, the artificial data generation unit 140 subsequently determines the total value of the trend 1603 and the periodic variation 1604 of the same date and time 1601 with respect to the date and time when the value of the trend 1603 of the intermediate data 1600 exists (is not missing). , "Replica source observation value") is obtained (S1305). The process of S1305 corresponds to the process of synthesizing the trend shown in (B-1) and the periodic fluctuation shown in (B-2) in FIG. By executing this process, the data shown in FIG. 14 (C) (hereinafter, referred to as “copy source data 1700”) can be obtained.

FIG. 17 shows an example of the copy source data 1700. As shown in the figure, the replication source data 1700 includes a plurality of entries having each item of date and time 1701, observation value 1702, trend 1703, periodic variation 1704, residual 1705, and replication source observation value 1706. Among the above items, the date and time 1701 is set to the value of the date and time 1601 of the entry having the value of the trend 1703 among the entries of the intermediate data 1600. The observation value 1702 is set to the value of the observation value 1602 corresponding to the value of the date and time 1701 among the entries of the intermediate data 1600. The value of the trend 1603 corresponding to the value of the date and time 1701 is set in the trend 1703 among the entries of the intermediate data 1600. The value of the periodic variation 1604 corresponding to the value of the date and time 1701 is set in the periodic variation 1704 among the entries of the intermediate data 1600. The residual 1705 is set to the value of the residual 1605 corresponding to the value of the date and time 1701 among the entries of the intermediate data 1600. The duplication source observation value 1706 is set to the sum of the value of the trend 1603 corresponding to the value of the date and time 1701 and the value of the periodic variation 1604 corresponding to the value of the date and time 1701 among the entries of the intermediate data 1600.

Returning to FIG. 13, the artificial data generation unit 140 subsequently duplicates the replication source data 1700 as many times as the number of replications determined in S1303 (S1306). Hereinafter, each duplicated data will be referred to as "replicated data".

Subsequently, the artificial data generation unit 140 assigns a natural number starting from 1 to each of the generated duplicated data in order, and the storage unit 110 assigns a number assigned to each duplicated data (hereinafter, referred to as a “duplicate number”). It is stored in association with each of the duplicated data (S1307). As described in the first embodiment, in the process, a negative integer starting from -1 is assigned as the duplication number in addition to the natural number starting from 1, so that the artificial data in the future period from the generating data 111 is artificial. Data 112 may be generated. In this case, as in the case of the first embodiment, the total value of the absolute value of the positive duplication number and the negative duplication number is made to match the number of duplications acquired in S1303.

Subsequently, the artificial data generation unit 140 generates primary artificial data by concatenating the duplicated data in the time series direction in the reverse order of the assigned duplicate numbers (S1308).

Subsequently, the artificial data generation unit 140 deletes the entry corresponding to the period of the duplicate number 1 and the date and time from t + T to t + T + Tp / 2 among the primary artificial data (S1309).

Subsequently, the artificial data generation unit 140 assigns data (hereinafter, referred to as “reference source date and time”) that duplicates the value of the date and time 701 of each entry of the generation source data 111 to each entry of the primary artificial data. (S1310).

Subsequently, the artificial data generation unit 140 generates the date and time of each entry by updating the reference source date and time of each entry of the primary artificial data to a value retroactive from the reference date and time (S1311). By this processing, for example, the date and time after the change of 12:00:00 on November 15, 2019 in the duplicated data of the duplicate number 2 is set to 12:00:00 on November 1, 2019, which is two weeks back. Become.

Subsequently, the artificial data generation unit 140 obtains the difference d of the periodic variation of the duplicated data before and after the boundary at the time point when the duplicated data is concatenated. Specifically, the artificial data generation unit 140 obtains the difference d between the value of the entry at the time point of t + Tp / 2 and the value of the entry at the time immediately before the time point with respect to the periodic fluctuation (S1312). For example, in the example of the intermediate data 1600 of FIG. 16, since t + Tp / 2 is 12:00:00 on November 15, 2019, 152 is obtained as the periodic variation 1604 of the same date and time. Further, 151 is obtained as the periodic variation 1604 at 11:50:00 on November 15, 2019, which is one time before the same date and time. Therefore, in this example, 1 is obtained as the difference d.

Subsequently, the artificial data generation unit 140 reflects the value obtained as the product of the difference d and the duplication number of each entry for each entry of the primary artificial data in the observed value of each entry of the primary artificial data (for example). , Addition or subtraction) (S1313). That is, the primary artificial data is generated on the assumption that the trend (difference d) acquired from the short-term data continues in the required period. After executing the process, the primary artificial data is as shown in FIG. 14 (D).

Subsequently, the artificial data generation unit 140 obtains the variance s ^ 2 of the residual obtained in S1304 (S1314).

Subsequently, the artificial data generation unit 140 generates white noise for a period corresponding to the period of the primary artificial data having the variance s ^ 2 (S1315). The white noise generated by executing the process is as shown in FIG. 14 (E).

Subsequently, the artificial data generation unit 140 generates a duplicated value of the observed value of each entry (hereinafter referred to as “reference source observed value”) with respect to the primary artificial data (S1316).

Subsequently, the artificial data generation unit 140 adds the white noise generated in S1315 as a fluctuation value to the primary artificial data (S1317).

Subsequently, the artificial data generation unit 140 generates artificial data by setting a value obtained by adding each fluctuation value to each reference source observation value to the observation value of each entry of the artificial data 112 (S1318). The artificial data 112 generated by executing the process is as shown in FIG. 14 (F).

FIG. 18 shows an example of artificial data 112. As shown in the figure, the artificial data 112 includes a plurality of entries having each item of date and time 1801, observed value 1802, reference source observed value 1803, variation value 1804, reference source date and time 1805, and replication number 1806. Among the above items, the date and time 1801 is set to the date and time generated in S1311. The value of the date and time 1801 also functions as an identifier for uniquely identifying each entry. The observed value obtained in S1318 is set in the observed value 1802. The reference source observation value generated in S1316 is set in the reference source observation value 1803. The value of the white noise given in S1317 is set in the fluctuation value 1804. The date and time assigned in S1310 is set in the reference source date and time 1805. The reference source date and time 1805 corresponds to the date and time 701 of the generating source data 111, and indicates that the entry is based on the entry of the date and time 701 of the generating source data 111. As the duplication number 1806, the duplication number assigned in S1307 is set.

As described above, the learning data generation device 100 of the second embodiment decomposes the generation source data 111, which is time series data including two cycles, into a trend, a periodic fluctuation, and a residual, and the trend and the periodic fluctuation. Based on the above, noise-free replication source data is generated, white noise is generated based on the dispersion s ^ 2 obtained from the residual, and artificial data is generated from the replication source data and the white noise. Therefore, it is possible to generate learning data that reproduces a fluctuation process that is close to the fluctuation process that actually occurs, and by using this to train the machine learning model 23, the inference accuracy of the inference device 2 can be improved. ..

Further, the learning data generation device 100 acquires the difference d (short-term trend) of the periodic fluctuation of the duplicated data before and after the time point at which the duplicated data is connected, and the product of the duplicate number and the difference d at the boundary. Artificial data 112 is generated by concatenating a plurality of duplicated data while changing the observed value by the value. Therefore, the learning data 114 in consideration of the long-term trend can be generated, and the machine learning model 23 can be learned accurately.

Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and various modifications are included. Further, for example, the above-described embodiment describes the configuration in detail in order to explain the present invention in an easy-to-understand manner, and is not necessarily limited to the one including all the described configurations. Further, a part of the configuration of each embodiment can be added, deleted, or replaced with another configuration.

Further, each of the above configurations, functions, processing units, processing means and the like may be realized by hardware by designing a part or all of them by, for example, an integrated circuit. It can also be realized by the program code of the software that realizes each function shown in the embodiment. In this case, a storage medium in which the program code is recorded is provided to the information processing device (computer), and the processor included in the information processing device reads out the program code stored in the storage medium. In this case, the program code itself read from the storage medium realizes the functions of the above embodiments, and the program code itself and the storage medium storing the program code itself constitute the present invention. Examples of the storage medium for supplying such a program code include a hard disk, an SSD (Solid State Drive), an optical disk, a magneto-optical disk, and a CD-R.
Flexible disks, CD-ROMs, DVD-ROMs, magnetic tapes, non-volatile memory cards, ROMs and the like are used.

In the above embodiments, the control lines and information lines indicate those considered necessary for explanation, and do not necessarily indicate all the control lines and information lines in the product. All configurations may be interconnected. In the above, various types of information are illustrated in a tabular format, but these information may be managed in a format other than the table.

1 Machine learning system, 2 Inference device, 21 Learning processing unit, 22 Inference processing unit, 23 Machine learning model, 100 Learning data generator, 110 Storage unit, 112 Artificial data, 111 Source data, 113 Observation data, 114 Learning data , 120 Observation data acquisition unit, 130 Generation source data acquisition unit, 140 Artificial data generation unit, 150 Learning data period setting unit, 160 Learning data generation unit, 170 Learning data output unit, S400 Learning data generation processing, S413 Artificial data generation processing , S424 training data period setting processing, S431 training data generation processing

Claims (18)

It is a learning data generator that is configured using an information processing device and generates learning data used for learning a machine learning model.
By concatenating a plurality of duplicated data which are duplicated data of the generation source data which is the time-series data for a predetermined cycle and giving noise to each of the duplicated data, the time-series data of the period corresponding to the required period is used. An artificial data generator that generates certain artificial data,
A learning data generation unit that generates learning data using the artificial data,
A learning data generator equipped with. The learning data generator according to claim 1.
The requested period is a period received from the user via the user interface.
Training data generator. The learning data generator according to claim 1.
The artificial data generation unit generates the artificial data in a period earlier than the generation source data, and generates the artificial data.
The learning data generation unit generates the learning data by connecting the artificial data to the generation source data in a time series.
Training data generator. The learning data generator according to claim 1.
The artificial data generation unit generates the artificial data in a future period from the generation source data, and generates the artificial data.
The learning data generation unit generates the learning data by connecting the artificial data to the generation source data in a time series.
Training data generator. The learning data generator according to claim 1 or 2.
The learning data generation unit connects the artificial data with observation data which is time-series data actually input to the machine learning model during the actual operation of the machine learning system that performs inference processing using the machine learning model. By generating the training data,
Training data generator. The learning data generator according to claim 5.
The learning data generation unit generates the learning data by adopting the observation data in preference to the artificial data.
Training data generator. The learning data generator according to claim 6.
A learning data period setting unit for setting a learning data period, which is a period from the start time to the end time of the learning data, is provided.
The learning data period setting unit sets the latest time point of the observation data as the end time point of the learning data period, and sets it.
A time point that goes back by the required period from the end time point is set as the start time point.
Training data generator. The learning data generator according to claim 1.
The artificial data generation unit individually gives noise to each of the duplicated data.
Training data generator. The learning data generator according to claim 1.
The noise is white noise,
Training data generator. The learning data generator according to claim 1.
The artificial data generation unit decomposes the generation source data into trend, periodic fluctuation, and residual components, and generates the artificial data based on the trend and the periodic fluctuation among the components. ,
Training data generator. The learning data generator according to claim 10.
The artificial data generation unit generates the noise based on the dispersion of the residual, and imparts the generated noise to the generated artificial data.
Training data generator. The learning data generator according to claim 10.
The artificial data generation unit obtains a difference in the periodic variation of the duplicated data before and after the boundary at a time when the duplicated data is connected, and generates a plurality of the duplicated data while reflecting the difference. Generate the artificial data by concatenating,
Training data generator. Information processing device
By concatenating a plurality of duplicated data which are duplicated data of the generation source data which is the time-series data for a predetermined cycle and giving noise to each of the duplicated data, the time-series data of the period corresponding to the required period is used. The steps to generate some artificial data,
A step of generating training data used for learning a machine learning model using the artificial data, and
How to generate training data. The learning data generation method according to claim 13.
The information processing device connects the artificial data with observation data which is time-series data actually input to the machine learning model during the actual operation of the machine learning system that performs inference processing using the machine learning model. To generate the training data by
A training data generation method that further executes. The learning data generation method according to claim 14.
A step in which the information processing device generates the learning data by adopting the observation data in preference to the artificial data.
A training data generation method that further executes. The learning data generation method according to claim 13.
A step in which the information processing device decomposes the generation source data into trend, periodic fluctuation, and residual components, and generates the artificial data based on the trend and the periodic fluctuation among the components. ,
A training data generation method that further executes. The learning data generation method according to claim 16.
A step in which the information processing device generates the noise based on the dispersion of the residuals and imparts the generated noise to the generated artificial data.
A training data generation method that further executes. The learning data generation method according to claim 16.
The information processing device obtains the difference in the periodic variation of the duplicated data before and after the boundary at the time when the duplicated data becomes a boundary when connecting the duplicated data, and concatenates the duplicated data while sequentially adding the differences. Steps to generate the artificial data by going,
A training data generation method that further executes.

Images