JP7409420B2 - Model creation device, data generation device, model creation method, and data generation method - Google Patents

Description

The present disclosure relates to a model creation device, a data generation device, a model creation method, and a data generation method for generating data related to a vehicle.

BACKGROUND ART A system is known that acquires data such as speed and acceleration of a vehicle and manages the vehicle based on the acquired data (for example, see Patent Document 1).

In order to understand the state of a vehicle, it is desirable for the system to analyze time-series data consisting of a large amount of data that changes over time. However, since time series data has a large data size, if the vehicle continues to transmit time series data, the load during communication will be large. Therefore, there is a need for a method that makes it easier to understand the state of a vehicle while suppressing the amount of data transmitted by the vehicle.

The applicant of the present application has proposed a technique for generating time series data from frequency information in Patent Document 2. As an example, Patent Document 2 discloses that by performing machine learning using vehicle speed time series data and the corresponding vehicle speed frequency and acceleration/deceleration frequency as learning data, vehicle speed It describes how to create a machine learning model that outputs series data.

Japanese Patent Application Publication No. 2012-248087 JP 2021-51637 Publication

By the way, in the technology described in Patent Document 2, generated time series data is output from the machine learning model according to the input of frequency data, so the frequency data regarding the generated time series data is based on the input frequency data. It is desirable that the values match exactly.

As an improvement measure when there is a difference between the input frequency data and the frequency data related to the generated time series data, it is possible to make more complex representations of the machine learning model by changing the structure of the machine learning model, such as making it multi-layered. One example is a method of creating a structure.

However, with this method, the amount of calculation increases as the number of layers increases, the learning time becomes longer, and an expensive and high-performance computer may be required. Furthermore, the method requires a number of trials in order to find a model structure that can achieve the desired accuracy, which has the disadvantage of being time-consuming.

The present invention has been made in consideration of the above points, and when generating time series data from frequency information using a machine learning model, it is possible to generate time series data without changing the structure of the machine learning model. Provided are a model creation device, a data generation device, a model creation method, and a data generation method that can improve the reproducibility of.

One aspect of the model creation device of the present invention is
a time-series data acquisition unit that acquires time-series data while the vehicle is running, which is measured while the vehicle is running, as learning time-series data;
a learning frequency data acquisition unit that acquires learning frequency data indicating an occurrence frequency distribution regarding the learning time series data;
By performing machine learning using the learning time series data and the corresponding learning frequency data as teacher data using weighting processing based on the difference, time series data is generated in response to input of frequency data. a model creation unit that creates a machine learning model that outputs
a frequency data generation unit that obtains frequency data indicating an occurrence frequency distribution regarding the generated time series data output from the machine learning model as generated frequency data;
a difference calculation unit that calculates a difference between the frequency data input to the machine learning model and the generation frequency data;
a correction unit that corrects the generated time series data by dividing, deleting, and duplicating the generated time series data so that the difference is small;
Equipped with.

One aspect of the data generation device of the present invention is
a frequency data acquisition unit that acquires frequency data indicating an occurrence frequency distribution regarding the time series data from time series data while the vehicle is running, which is measured while the vehicle is running;
The frequency data is applied to a machine learning model obtained by performing machine learning using weighting processing based on differences using learning time series data of the vehicle while the vehicle is running and corresponding learning frequency data as teacher data. a data generation unit that generates generated time series data that is time series data corresponding to the frequency data by inputting the frequency data;
a frequency data generation unit that obtains frequency data indicating an occurrence frequency distribution regarding the generated time series data output from the machine learning model as generated frequency data;
a difference calculation unit that calculates a difference between the frequency data input to the machine learning model and the generation frequency data;
a correction unit that corrects the generated time series data by dividing, deleting, and duplicating the generated time series data so that the difference is small;
Equipped with.

One aspect of the model creation method of the present invention is
A computer-implemented model creation method, the method comprising:
a step of acquiring time-series data of the vehicle while it is running, which is measured while the vehicle is running, as time-series data for learning;
obtaining learning frequency data indicating an occurrence frequency distribution regarding the learning time series data;
By performing machine learning using the learning time series data and the corresponding learning frequency data as teacher data using weighting processing based on the difference, time series data is generated in response to input of frequency data. a step of creating a machine learning model that outputs
obtaining frequency data indicating an occurrence frequency distribution regarding the generated time series data output from the machine learning model as generated frequency data;
calculating a difference between the frequency data input to the machine learning model and the generated frequency data;
correcting the generated time series data by dividing, deleting, and duplicating the generated time series data so that the difference is small;
including.

One aspect of the data generation method of the present invention is
A data generation method performed by a computer, the method comprising:
acquiring frequency data indicating an occurrence frequency distribution regarding the time series data from time series data of the vehicle while the vehicle is running, which is measured while the vehicle is running;
The frequency data is applied to a machine learning model obtained by performing machine learning using weighting processing based on differences using learning time series data of the vehicle while the vehicle is running and corresponding learning frequency data as teacher data. a step of generating generated time series data that is time series data corresponding to the frequency data by inputting the frequency data;
obtaining frequency data indicating an occurrence frequency distribution regarding the generated time series data output from the machine learning model as generated frequency data;
calculating a difference between the frequency data input to the machine learning model and the generation frequency data;
correcting the generated time series data by dividing, deleting, and duplicating the generated time series data so that the difference is small;
including.

According to the present invention, frequency data regarding generated time series data is obtained as generation frequency data, the difference between the frequency data input to a machine learning model and the generation frequency data is calculated, and the generated time series data is Since the generated time series data is corrected by dividing, deleting, and duplicating the data, when generating time series data from frequency information using a machine learning model, the structure of the machine learning model must be changed. At the very least, it is possible to improve the reproducibility of generated time series data.

Diagram to explain the overview of the data generation system Diagram to explain the overview of the data generation system FIG. 3A is a diagram showing time series data, FIG. 3B is a diagram showing frequency data of vehicle speed, and FIG. 3C is a diagram showing frequency data of acceleration. Block diagram showing the configuration of the data generation device Diagram showing an overview of the process by which the model creation unit creates a machine learning model configured by conditional VAE Diagram showing the process in which the data generation unit generates time series data using a machine learning model Flowchart showing the flow of processing in the data generation device A block diagram showing the configuration of a data generation device for realizing correction according to an embodiment. Diagram showing an example of generated time series data output from a machine learning model FIG. 9A is a diagram illustrating the operation of the frequency data generation unit, the difference calculation unit, and the correction unit, and FIG. 9A is a diagram showing the difference between the frequency data input to the machine learning model by vehicle speed and the generated frequency data. 9C is a diagram showing generated vehicle speed time series data before correction, and FIG. 9C is a diagram showing generated vehicle speed time series data after correction. FIG. 10A is a diagram showing how the generated vehicle speed time series data is divided into a plurality of generated divided vehicle speed time series data, and FIG. 10B is the generated vehicle speed time series data after correction obtained by combining the generated divided vehicle speed time series data of FIG. 10A. Diagram showing

Embodiments of the present disclosure will be described below with reference to the drawings.

<1> Overview of data generation system S FIGS. 1 and 2 are diagrams for explaining an overview of the data generation system S. The data generation system S is a system for generating time series data of various parameters measured in the vehicle T based on frequency data of the parameters. The data generation system S includes a data collection device 1 and a data generation device 2. The data generation device 2 is a device that uses a machine learning model to generate time series data based on frequency data. The data generation device 2 also functions as a model creation device that creates a machine learning model. The machine learning model is configured to include, for example, a conditional VAE (Variational Auto Encoder) or a conditional GAN (Generative Adversarial Network).

FIG. 3 is a diagram showing an overview of time series data and frequency data. As shown in FIG. 3A, the time series data is data indicating values of parameters that change over time, and is composed of, for example, values of the speed of the vehicle T every second. As shown in FIG. 3B, the frequency data is data indicating the distribution of the frequency of occurrence of parameter values (velocity) within a predetermined period. The frequency data may be data indicating the distribution of the frequency of occurrence of a value (acceleration) obtained by first-order differentiation of a parameter, as shown in FIG. 3C.

When the parameter is the speed of the vehicle T, the frequency data may be Nkm/h (N is an integer greater than or equal to 0) is data indicating the time or rate at which the state occurs (see FIG. 3B). When the parameter is the speed of the vehicle T, the frequency data may be data indicating the time or rate at which a predetermined acceleration occurs within a unit time (see FIG. 3C). Note that while the vehicle T is accelerating, the acceleration has a positive value, and while the vehicle T is decelerating, the acceleration has a negative value.

For the parameters measured in the vehicle T, the data generation system S creates a machine learning model that performs machine learning (for example, deep learning) using time series data and frequency data of the parameters measured in the vehicle T as training data. Using a machine learning model, it is possible to generate time series data based on frequency data obtained from the vehicle T.

This makes it possible to generate time-series data with a large amount of data based on frequency data with a small amount of data.

By analyzing the time-series data generated by the data generation system S, the administrator of the vehicle T can obtain various information such as the fuel efficiency, degree of deterioration, and driving quality of the vehicle T.

Hereinafter, an overview of the data generation system S will be explained with reference to FIGS. 1 and 2. The data collection device 1 is a device that acquires data on parameters measured in a large number of vehicles T via a network N, and is, for example, a computer.

As shown in FIG. 1, the data generation device 2 generates a machine learning model that performs machine learning using time series data acquired from the vehicle T via the data collection device 1 and frequency data corresponding to the time series data as training data. It is a computer to create. Moreover, as shown in FIG. 2, the data generation device 2 generates time series data based on the frequency data obtained from the vehicle T using the created machine learning model.

FIG. 1 is a diagram showing the operation of the data generation system S when the data generation device 2 performs machine learning and creates a machine learning model. The data collection device 1 acquires measurement data of a predetermined parameter (for example, speed) from a vehicle T registered in advance ((1) in FIG. 1). The data collection device 1 transmits time-series data of the acquired measurement data to the data generation device 2 ((2) in FIG. 1).

The data generation device 2 is a machine that generates frequency data based on the time series data received from the data collection device 1, uses the time series data and frequency data as training data, and outputs time series data when the frequency data is input. Create a learning model ((3) in Figure 1). Instead of the data generation device 2 generating frequency data, the data collection device 1 may generate frequency data from time series data, and the data collection device 1 may transmit the time series data and frequency data to the data generation device 2. .

Next, with reference to FIG. 2, the operation after the data generation device 2 creates the machine learning model will be described. The vehicle T transmits frequency data of the measured parameters to the data collection device 1 ((4) in FIG. 2). The data collection device 1 transmits the frequency data received from the vehicle T to the data generation device 2 ((5) in FIG. 2). The data generation device 2 generates time series data by inputting the received frequency data into a machine learning model and acquiring time series data output from the machine learning model ((6) in FIG. 2). The data generation device 2 transmits the generated time series data to the data collection device 1 ((7) in FIG. 2).

Through the above-described flow, a user such as an administrator of the vehicle T using the data generation device 2 can obtain time series data of desired parameters based on the frequency data. The data generation device 2 may transmit the generated time series data to any computer other than the data collection device 1, display it on a display, or print it.

<2> Configuration and operation of data generation device 2 FIG. 4 is a diagram showing the configuration of data generation device 2. The data generation device 2 includes a communication section 21, a storage section 22, and a control section 23. The control unit 23 includes a time series data acquisition unit 231, a learning frequency data acquisition unit 232, a generation frequency data acquisition unit 233, a data output unit 234, a model creation unit 235, and a data generation unit 236. .

When the data generation device 2 does not have the generation frequency data acquisition unit 233, the data output unit 234, and the data generation unit 236, the data generation device 2 functions as a model creation device that creates the machine learning model M.

The communication unit 21 is a communication interface for transmitting and receiving data with the data collection device 1 or other external devices. The communication unit 21 sends received data to the control unit 23, and also sends data input from the control unit 23 to an external device.

The storage unit 22 includes storage media such as a ROM (Read Only Memory), a RAM (Random Access Memory), and a hard disk. The storage unit 22 stores programs executed by the control unit 23. Furthermore, the storage unit 22 temporarily stores the time series data and frequency data received from the data collection device 1.

The control unit 23 is, for example, a CPU (Central Processing Unit). The control unit 23 executes the program stored in the storage unit 22 to obtain a time series data acquisition unit 231, a learning frequency data acquisition unit 232, a generation frequency data acquisition unit 233, a data output unit 234, and a model creation unit. 235 and a data generation unit 236.

The time-series data acquisition unit 231 acquires time-series data of parameters measured while the vehicle T is running as learning time-series data. The time-series data acquisition unit 231 acquires, for example, time-series data of vehicle speed measured while the vehicle T is running as learning time-series data, and sends this to the model creation unit 235.

The learning frequency data acquisition unit 232 acquires learning frequency data corresponding to the learning time series data acquired by the time series data acquisition unit 231. The learning frequency data acquisition unit 232 acquires the learning frequency data via the communication unit 21, for example, but the learning frequency data acquisition unit 232 creates the learning frequency data based on the learning time series data. By doing so, the frequency data for learning may be obtained from the time series data for learning.

The learning frequency data acquisition unit 232 acquires, as the learning frequency data, for example, data indicating the occurrence frequency distribution of parameters related to the learning time series data (see FIG. 3B). Further, the learning frequency data acquisition unit 232 may acquire data (see FIG. 3C) indicating the occurrence frequency distribution of the first differential value of the learning time series data as the learning frequency data.

When the learning time series data is speed time series data, the learning frequency data acquisition unit 232 acquires the learning speed frequency data indicating the frequency occurrence distribution of speed in the learning time series data and/or the learning time series data. Obtain learning acceleration frequency data that indicates the frequency distribution of acceleration in series data. The learning frequency data acquisition unit 232 sends the acquired learning frequency data to the model creation unit 235.

The generation frequency data acquisition unit 233 acquires generation frequency data used to generate time series data using the machine learning model M. In this specification, time series data generated using the machine learning model M is referred to as generated time series data. The generation frequency data acquisition unit 233 acquires generation speed frequency data and/or generation acceleration frequency data as generation frequency data. The generation frequency data acquisition unit 233 sends the acquired generation frequency data to the data generation unit 236.

The data output unit 234 outputs generated time series data generated from the machine learning model M by the data generation unit 236 based on the generation frequency data. The data output unit 234 transmits the generated time series data output from the data generation unit 236 to an external device such as the communication unit 21.

The model creation unit 235 creates a machine learning model M, and stores the weights of the created machine learning model M in the storage unit 22. The model creation unit 235 may store the weights in a memory (not shown) that the model creation unit 235 has.

The model creation unit 235 performs machine learning using weighting processing using learning time series data and learning frequency data related thereto as teacher data, thereby responding to the frequency data in response to input of the frequency data. A machine learning model M that outputs generated time series data that is time series data is created.

The learning time series data is the speed time series data of the vehicle T, the learning frequency data is the speed frequency data and/or the acceleration frequency data of the vehicle T, and the generated time series data is the speed time series In the case of data, the machine learning model M outputs generated time series data that is time series data of the speed of the vehicle T in response to input of the speed frequency data and/or the acceleration frequency data.

The data generation unit 236 generates generation time series data by inputting the generation frequency data inputted from the generation frequency data acquisition unit 233 into the machine learning model M. When the generation frequency data input from the generation frequency data acquisition unit 233 is speed frequency data and/or acceleration frequency data of the vehicle T, the data generation unit 236 generates the speed frequency data and/or the acceleration frequency data of the vehicle T. Alternatively, by inputting the frequency data of the acceleration of the vehicle T, the time series data of the speed output from the machine learning model M is generated as the generated time series data. The generated time series data generated by the data generation section 236 is outputted to the outside via the data output section 234.

<3> Method for Creating Machine Learning Model M FIG. 5A is a diagram illustrating an overview of a process in which the model creation unit 235 creates a machine learning model M configured by conditional VAE as an example of the machine learning model M. FIG. 5B is a diagram showing a process in which the data generation unit 236 generates time series data using the machine learning model M.

As shown in FIG. 5, the machine learning model M is configured by a deep neural network (DNN) as an example. A DNN has a plurality of layers between an input layer and an output layer, and a variable weight is provided to each of a plurality of nodes included in each layer. The weights before the machine learning model M learns are initial values.

As shown in FIG. 5A, the model creation unit 235 creates a machine learning model M-1 in which a pair of input learning time series data and learning frequency data is input, and a machine learning model M-1 in which a latent variable vector z and a learning frequency data are input. The generated time series data output from the machine learning model M-2 is compared with the input learning time series data.

The model creation unit 235 generates generated time series data output from the machine learning model M-2 when the learning frequency data is input to the machine learning model M-1 and the machine learning model M-2, and the learning time series data. The weights of machine learning model M-1 and machine learning model M-2 are updated based on the difference between machine learning model M-1 and machine learning model M-2.

For example, when the difference between the generated time series data and the learning time series data is greater than or equal to a threshold, the model creation unit 235 back-propagates the difference and updates the weights of nodes on the back-propagated route. The model creation unit 235 generates data by inputting the learning frequency data to the machine learning model M-1 and the machine learning model M-2 until the difference between the generated time series data and the learning time series data becomes less than a threshold value. The comparison between the generated time series data and the training time series data and the updating of the weights are repeated. The model creation unit 235 executes the above processing using a large number of pairs of learning frequency data and learning time series data, thereby creating the machine learning model M-3 (machine learning model M-2 and machine learning model M-2) shown in FIG. 5B. (substantially identical models).

After the model creation unit 235 updates the machine learning model M-1 and the machine learning model M-2 and completes the machine learning model M-3, the data generation unit 236 updates the machine learning model M-1 and the machine learning model M-2 from the vehicle T as shown in FIG. By inputting the acquired frequency data (generation frequency data) to the machine learning model M-3, the machine learning model M-3 outputs generated time series data corresponding to the input frequency data.

<4> Flow of processing in data generation device 2 FIG. 6 is a flowchart showing the flow of processing in data generation device 2. The flowchart shown in FIG. 6 starts from the time when the data generation device 2 receives an instruction to start creating the machine learning model M.

Upon receiving the instruction to create the machine learning model M, the model creation unit 235 acquires learning time series data from the time series data acquisition unit 231 (step S11). The model creation unit 235 also acquires learning frequency data from the learning frequency data acquisition unit 232 (step S12). The order in which steps S11 and S12 are executed is arbitrary, and the model creation unit 235 may acquire the learning time series data and the learning frequency data at the same time. The model creation unit 235 updates the machine learning model M by performing machine learning using a set of learning time series data and learning frequency data as teacher data (step S13). Specifically, the model creation unit 235 updates the weights of the machine learning model M stored in the storage unit 22.

The model creation unit 235 evaluates the performance of the updated machine learning model M, and determines whether the evaluated result is equal to or higher than the reference level (step S14). For example, when the difference between the generated time series data output from the machine learning model M when frequency data is input to the machine learning model M and the learning time series data is less than a threshold, the model creation unit 235 It is determined that the evaluation result is equal to or higher than the reference level.

If the model creation unit 235 determines that the evaluated result has not reached the reference level (NO in step S14), the process returns to step S11 and repeats machine learning using further learning time series data and learning frequency data. . If the model creation unit 235 determines that the evaluated result has reached the reference level (YES in step S14), it ends the update of the machine learning model M (step S15).

After that, the data generation unit 236 acquires generation frequency data for generating time series data (step S16), and inputs the generation frequency data to the machine learning model M (step S17). The data generation unit 236 outputs the generated time series data output from the machine learning model M (step S18).

<5> Correction of Generated Time Series Data FIG. 7, in which parts corresponding to those in FIG. 4 are denoted by the same reference numerals, is a block diagram showing the configuration of a further improved data generation device 300. The data generation device 300 has the same configuration as the data generation device 2 in FIG. 4, except that the model creation section 235 includes a frequency data generation section 301, a difference calculation section 302, and a correction section 303.

Note that the frequency data generation unit 301, the difference calculation unit 302, and the correction unit 303 are for correcting the generated time series data output from the updated and completed machine learning model M, and in the example of FIG. Although these are provided within the model creation section 235, they may be provided between the model creation section 235 and the data generation section 236, or within the data generation section 236.

The frequency data generation unit 301 obtains frequency data indicating an occurrence frequency distribution regarding the generated time series data output from the machine learning model M as generation frequency data.

The difference calculation unit 302 calculates the difference between the generation frequency data obtained by the generation frequency data acquisition unit 233 and the generation frequency data obtained by the frequency data generation unit 301.

The correction unit 303 corrects the generated time series data by dividing, deleting, and duplicating the generated time series data from the machine learning model M so that the difference calculated by the difference calculation unit 302 becomes small.

The corrected generated time series data is output to the data generation unit 236. That is, the generation frequency data obtained by the generation frequency data acquisition unit 233 is input to the model creation unit 235 via the data generation unit 236, and generated time series data based on the generation frequency data is generated from the machine learning model M. Once output, this generated time series data is corrected by the correction unit 303 and output to the data generation unit 236.

As a result, when obtaining generated time series data from generation frequency data using machine learning model M, the reproducibility of generated time series data with respect to training time series data can be improved without changing the structure of machine learning model M. be able to improve.

Next, characteristic operations of the data generation device 300 will be described using FIGS. 8 and 9. Here, an example will be explained in which the data to be handled is vehicle speed data.

FIG. 8 shows an example of generated time series data (generated vehicle speed time series data) output from the machine learning model M. FIG. 9 is a diagram for explaining the operations of the frequency data generation unit 301, the difference calculation unit 302, and the correction unit 303, and FIG. 9A is a diagram showing the difference between the generation frequency data and the generation frequency data for each vehicle speed. FIG. 9B is a diagram showing generated vehicle speed time series data before correction, and FIG. 9C is a diagram showing generated vehicle speed time series data after correction.

In the example of FIG. 9A, regarding the vehicle speed in the vehicle speed range c, the difference between the generation frequency data obtained by converting the generated vehicle speed time series data by the frequency data generation unit 301 and the generation frequency data is greater than or equal to a predetermined threshold value. Specifically, the generation frequency data is smaller than the generation frequency data, and the difference therebetween is greater than or equal to a predetermined threshold. This means that there is a large shortage of vehicle speed data in vehicle speed range c in the generated vehicle speed time series data, so as shown in FIGS. 9B and 9C, the correction unit 303 generates data that includes vehicle speed in vehicle speed range c. Duplicate the divided vehicle speed time series data to create corrected generated vehicle speed time series data.

On the other hand, in the example of FIG. 9A, regarding the vehicle speed in vehicle speed range b, the difference between the generation frequency data obtained by converting the generated vehicle speed time series data by the frequency data generation unit 301 and the generation frequency data is greater than or equal to the predetermined threshold value. It is. Specifically, the generation frequency data is greater than the generation frequency data, and the difference is greater than or equal to a predetermined threshold. This means that there is too much vehicle speed data in vehicle speed range b in the generated vehicle speed time series data, so as shown in FIGS. 9B and 9C, the correction unit 303 uses Delete the series data to create corrected vehicle speed time series data.

In this embodiment, the correction unit 303 divides the generated time series data into a plurality of generated divided time series data, and the difference calculated by the difference calculation unit 302 for each arbitrary division unit is set to a predetermined threshold value. The generated time series data is corrected so that the difference is equal to or less than the predetermined threshold by determining whether the difference is greater than or equal to the predetermined threshold and deleting or duplicating the generated divided time series data that is the cause of the difference greater than or equal to the predetermined threshold.

Here, the unit of division of the generated time series data is the mountain-shaped section in the figure, in other words, the section where the value fluctuates from the reference value until it returns to the reference value. In the example of FIG. 9, the reference value is a vehicle speed of 0 [km/h].

As shown in FIGS. 9B and 9C, the generated time series data is deleted and duplicated in division units (mountain units), so deletion and duplication can be easily performed. The amount of processing for replication can be suppressed. Further, by deletion and processing, it is possible to easily obtain a corrected generated time-series signal having a desired length.

In the example of FIG. 9, corrected generation time series data was obtained by deleting and duplicating the generation divided vehicle speed time series data based on the difference between the frequency data for generation and the frequency data converted from the generation time series data. However, the corrected generation time series data may be obtained by processing as shown in FIG.

FIG. 10A is a diagram showing how the generated vehicle speed time series data is divided into a plurality of generated divided vehicle speed time series data, and FIG. 10B is a diagram showing the generated vehicle speed time series data after correction obtained by combining the generated divided vehicle speed time series data of FIG. 10A. FIG. 3 is a diagram showing series data.

The generated divided vehicle speed time series data shown in FIG. 10A are combined to create the corrected generated vehicle speed time series data shown in FIG. 10B so that the difference between the generation frequency data and the generation frequency data approaches 0. As in the process of FIG. 9, it becomes possible to match the generation frequency data with the frequency data converted from the generation time series data with high precision.

In addition, when using either method of FIG. 9 or FIG. 10, if there are various generation division time series data, it is possible to obtain corrected generation frequency data closer to the generation frequency data. Taking this into consideration, it is desirable to prepare a large amount of generated and divided time series data using various methods, such as generating time series data that is longer than the desired time series length. However, after obtaining generated time series data that is the same as or shorter than the desired time series length, as shown in Figures 9 and 10, the generated divided time series data contained therein is duplicated and deleted to generate corrected data. Time series data may also be created.

<6> Summary As explained above, the model creation device (control unit 23) of the present embodiment uses time-series data of the vehicle T while it is running, which is measured while the vehicle T is running, as learning time-series data. A learning time series data acquisition unit 231 acquires learning frequency data acquisition unit 231, a learning frequency data acquisition unit 232 acquires learning frequency data indicating an occurrence frequency distribution regarding the learning time series data, and the learning time series data and the learning corresponding thereto. a model creation unit that creates a machine learning model M that outputs generated time series data in response to input of the frequency data by performing machine learning using weighting processing based on differences using the usage frequency data as training data; 235, a frequency data generation unit 301 that obtains frequency data indicating an occurrence frequency distribution regarding the generated time series data outputted from the machine learning model M as generated frequency data, and a frequency data generation unit 301 that generates frequency data input to the machine learning model and generated frequency data. It includes a difference calculation unit 302 that calculates a difference, and a correction unit 303 that corrects the generated time series data by dividing, deleting, and duplicating the generated time series data so that the difference becomes small.

The data generation device (control unit 23) of the present embodiment acquires frequency data indicating an occurrence frequency distribution regarding time series data from time series data during running of the vehicle T measured while the vehicle T is running. A machine learning model obtained by performing machine learning using the acquisition unit 233, learning time series data while the vehicle T is traveling, and corresponding learning frequency data as teacher data using weighting processing based on differences. A data generation unit 236 that generates generated time series data that is time series data corresponding to the frequency data by inputting the frequency data to M, and an occurrence related to the generated time series data output from the machine learning model M. A frequency data generation unit 301 that obtains frequency data indicating a frequency distribution as generated frequency data, and a difference calculation unit 302 that calculates a difference between the frequency data and the generated frequency data, generate time series data so that the difference is small. and a correction unit 303 that corrects the generated time series data by dividing, deleting, and duplicating the data.

By doing this, when generating time series data from frequency information using a machine learning model, it is possible to improve the reproducibility of the generated time series data without changing the structure of the machine learning model. become.

Although the present invention has been described above using the embodiments, the technical scope of the present invention is not limited to the scope described in the above embodiments, and various modifications and changes can be made within the scope of the gist. be. For example, all or part of the device can be functionally or physically distributed and integrated into arbitrary units. In addition, new embodiments created by arbitrary combinations of multiple embodiments are also included in the embodiments of the present invention. The effects of the new embodiment resulting from the combination have the effects of the original embodiment.

In the above-described embodiment, the case where the time-series data is mainly the speed of the vehicle T has been described, but as described above, the time-series data is not limited to this, and for example, the acceleration of the vehicle T, and the case where the time-series data is used in the vehicle T are described. Time series of various parameters that change over time as the vehicle T travels, such as the temperature of the cooling water used in the vehicle T, the temperature of the oil used in the vehicle T, the accelerator opening of the vehicle T, and the amount of vibration of the vehicle T. It can be data.

Further, in the above-described embodiment, the case where the data generation device 2 acquires time series data and frequency data from the data collection device 1 is illustrated, but the data generation device 2 has the function of the data collection device 1. , the data generation device 2 may receive measurement data from a plurality of vehicles T.

In addition, in the above description, the computer functioning as the data generation device 2 has the function of a model creation device, and also generates a generation time series output from the machine learning model M in response to input of generation frequency data. Although the case where the data generation device 2 also has the function of outputting data has been illustrated, the configuration of the data generation device 2 is not limited to this. The data generation device 2 may include a first computer that functions as a model creation device, and a second computer that inputs frequency data for generation into the first computer and acquires generated time series data from the first computer. good.

INDUSTRIAL APPLICATION This invention can be widely used as a technique which reproduces the time series data regarding a vehicle while the vehicle is running from a small amount of data through learning.

1 Data collection device 2, 300 Data generation device 21 Communication unit 22 Storage unit 23 Control unit 231 Time series data acquisition unit 232 Learning frequency data acquisition unit 233 Generation frequency data acquisition unit 234 Data output unit 235 Model creation unit 236 Data generation Section 301 Frequency data generation section 302 Difference calculation section 303 Correction section M Machine learning model S Data generation system T Vehicle

Claims (5)

a learning time-series data acquisition unit that acquires time-series data during running of the vehicle, which is measured while the vehicle is running, as learning time-series data;
a learning frequency data acquisition unit that acquires learning frequency data indicating an occurrence frequency distribution regarding the learning time series data;
By performing machine learning using the learning time series data and the corresponding learning frequency data as teacher data using weighting processing based on the difference, time series data is generated in response to input of frequency data. a model creation unit that creates a machine learning model that outputs
a frequency data generation unit that obtains frequency data indicating an occurrence frequency distribution regarding the generated time series data output from the machine learning model as generated frequency data;
a difference calculation unit that calculates a difference between the frequency data input to the machine learning model and the generation frequency data;
a correction unit that corrects the generated time series data by dividing, deleting, and duplicating the generated time series data so that the difference is small;
A model creation device comprising: The correction unit is
The generated time series data is divided into a plurality of generated divided time series data,
Determining whether the difference or the difference obtained by recalculating the difference per arbitrary division unit is greater than or equal to a predetermined threshold;
Correcting the generated time series data so that the difference becomes less than or equal to the predetermined threshold by deleting or duplicating the generated time series data that is the cause of the difference that is greater than or equal to the predetermined threshold;
The model creation device according to claim 1. a frequency data acquisition unit that acquires frequency data indicating an occurrence frequency distribution regarding the time series data from time series data while the vehicle is running, which is measured while the vehicle is running;
The frequency data is applied to a machine learning model obtained by performing machine learning using weighting processing based on differences using learning time series data of the vehicle while the vehicle is running and corresponding learning frequency data as teacher data. a data generation unit that generates generated time series data that is time series data corresponding to the frequency data by inputting the frequency data;
a frequency data generation unit that obtains frequency data indicating an occurrence frequency distribution regarding the generated time series data output from the machine learning model as generated frequency data;
a difference calculation unit that calculates a difference between the frequency data input to the machine learning model and the generation frequency data;
a correction unit that corrects the generated time series data by dividing, deleting, and duplicating the generated time series data so that the difference is small;
A data generation device comprising: A computer-implemented model creation method, the method comprising:
a step of acquiring time-series data of the vehicle while it is running, which is measured while the vehicle is running, as time-series data for learning;
obtaining learning frequency data indicating an occurrence frequency distribution regarding the learning time series data;
By performing machine learning using the learning time series data and the corresponding learning frequency data as teacher data using weighting processing based on the difference, time series data is generated in response to input of frequency data. a step of creating a machine learning model that outputs
obtaining frequency data indicating an occurrence frequency distribution regarding the generated time series data output from the machine learning model as generated frequency data;
calculating a difference between the frequency data input to the machine learning model and the generated frequency data;
correcting the generated time series data by dividing, deleting, and duplicating the generated time series data so that the difference is small;
Model creation methods including. A data generation method performed by a computer, the method comprising:
acquiring frequency data indicating an occurrence frequency distribution regarding the time series data from time series data of the vehicle while the vehicle is running, which is measured while the vehicle is running;
The frequency data is applied to a machine learning model obtained by performing machine learning using weighting processing based on differences using learning time series data of the vehicle while the vehicle is running and corresponding learning frequency data as teacher data. a step of generating generated time series data that is time series data corresponding to the frequency data by inputting the frequency data;
obtaining frequency data indicating an occurrence frequency distribution regarding the generated time series data output from the machine learning model as generated frequency data;
calculating a difference between the frequency data input to the machine learning model and the generation frequency data;
correcting the generated time series data by dividing, deleting, and duplicating the generated time series data so that the difference is small;
data generation methods including;

Images