WO2021220715A1 - Determination device and determination program - Google Patents

Abstract

The objective of the present invention is to determine accurately the specific behavior class of a ruminant. This determination device comprises: a generation unit performing frequency analysis on time-series one-dimensional data delivered from an acceleration sensor attached to a ruminant and representing acceleration, and generating two-dimensional array data representing the intensity of each frequency at each time point; and a determination unit taking as input the two-dimensional array data for each of preset time ranges and determining the behavior class of the ruminant for each of the preset time ranges.

Description

Judgment device and judgment program

The present invention relates to a determination device and a determination program.

In general, for ruminant animals such as cows, the length of rumination time is an indicator of their health condition. Therefore, if the rumination time can be appropriately managed, it is possible to accurately detect that the health condition has changed.

On the other hand, as a method of determining whether or not a ruminant state is present, for example, a method of attaching an acceleration sensor to a ruminant animal and analyzing acceleration data can be mentioned.

Japanese Unexamined Patent Publication No. 2018-1770969 JP-A-2017-158509 Japanese Unexamined Patent Publication No. 2017-051146

However, the acceleration data is only one-dimensional time-series data indicating the magnitude of acceleration at each time, and the amount of information is small. On the other hand, the acceleration data contains variable elements based on various behaviors other than rumination behavior, and based on the one-dimensional time-series data, the characteristics of rumination behavior, which is one of the behavior classifications, are grasped and rumination. It is not easy to accurately determine whether or not it is in a state.

One aspect is aimed at accurately determining a specific behavioral classification of ruminants.

According to one aspect, the determination device
A generator that frequency-analyzes one-dimensional time-series data indicating acceleration output from an acceleration sensor attached to an anticorrosive animal and generates two-dimensional array data indicating the intensity of each frequency at each time.
It has a determination unit for determining the behavior classification of the ruminant for each predetermined time range by inputting the two-dimensional array data for each predetermined time range.

It is possible to accurately determine the specific behavioral classification of ruminants.

FIG. 1 is a diagram showing an example of a system configuration of a determination system and a functional configuration of a determination device. FIG. 2 is a diagram showing an example of the hardware configuration of the determination device. FIG. 3 is a first diagram showing a specific example of the functional configuration and processing of the learning data generation unit. FIG. 4 is a diagram showing an example of the functional configuration of the learning unit. FIG. 5 is a diagram showing a specific example of the functional configuration and processing of the inference data generation unit. FIG. 6 is a first diagram showing a specific example of the functional configuration and processing of the inference unit. FIG. 7 is a flowchart showing the flow of the determination process. FIG. 8 is a second diagram showing a specific example of the functional configuration and processing of the learning data generation unit. FIG. 9 is a third diagram showing a specific example of the functional configuration and processing of the learning data generation unit. FIG. 10 is a fourth diagram showing a specific example of the functional configuration and processing of the learning data generation unit. FIG. 11 is a second diagram showing a specific example of the functional configuration and processing of the inference unit.

Hereinafter, each embodiment will be described with reference to the attached drawings. In the present specification and the drawings, components having substantially the same functional configuration are designated by the same reference numerals, so that duplicate description will be omitted.

[First Embodiment]
<System configuration of judgment system and functional configuration of judgment device>
First, the system configuration of the determination system for determining the behavior classification of cattle (whether the cattle are in the ruminant state or the non-ruminant state), which is an example of ruminant animals, and the functional configuration of the determination device will be described. FIG. 1 is a diagram showing an example of a system configuration of a determination system and a functional configuration of a determination device.

As shown in FIG. 1, the determination system 100 includes a measuring device 110, a gateway device 120, and a determination device 130. In the determination system 100, the measuring device 110 and the gateway device 120 are connected via wireless communication, and the gateway device 120 and the determination device 130 are communicably connected via a network (not shown).

The measuring device 110 is an acceleration sensor in the three-axis directions (x-axis direction, y-axis direction, z-axis direction) attached to a predetermined portion (neck in the example of FIG. 1) of the cow 10. The x-axis direction is, for example, a direction along the body surface of the neck of the cow 10, and refers to a direction along the circumference of the neck, and the y-axis direction is, for example, the body surface of the neck of the cow 10. It shall be a direction along the direction from the head to the body. Further, the z-axis direction is defined as, for example, a direction perpendicular to the body surface of the neck of the cow 10.

The measuring device 110 measures one-dimensional time-series data indicating acceleration in each of the three axial directions at a predetermined sampling frequency (for example, 5 [Hz] to 20 [Hz]) and transmits it to the gateway device 120.

The gateway device 120 transmits the one-dimensional time series data indicating the accelerations in each of the three axial directions transmitted from the measuring device 110 to the determination device 130.

The determination device 130 determines whether the cow 10 is in a ruminant state or a non-ruminant state based on one-dimensional time-series data indicating acceleration in each of the three axial directions. A determination program is installed in the determination device 130, and when the program is executed, the determination device 130 serves as a learning data generation unit 131, a learning unit 132, an inference data generation unit 133, and an inference unit 134. Function.

The learning data generation unit 131 extracts one-dimensional time-series data indicating acceleration in the z-axis direction from one-dimensional time-series data indicating acceleration in each of the three axis directions. Further, the learning data generation unit 131 performs frequency analysis of the extracted one-dimensional time series data for each predetermined analysis time range (for example, 15 seconds) to show the intensity of each frequency at each time. Generate sequence data (spectrogram image).

The predetermined analysis time range is shifted at a predetermined shift interval (for example, 1 second interval), and the result of frequency analysis in each analysis time range is a two-dimensional array at each start time of each analysis time range. Generated as data. That is, each time referred to here refers to the start time of each time range (hereinafter, the same applies).

Further, the learning data generation unit 131 associates the two-dimensional array data with the information indicating that it is in the rebellious state or the information indicating that it is in the non-reflexed state (correct answer data) at each time for learning. It is stored as data in the learning data storage unit 135.

The learning unit 132 reads the learning data from the learning data storage unit 135, and obtains the two-dimensional array data at each time included in the read learning data for each predetermined determination time range (for example, 15 seconds). Input to the judgment model in sequence. The predetermined determination time range is shifted at a predetermined shift interval (for example, 1 second interval), and the two-dimensional array data of each determination time range is input to the determination model.

Further, the learning unit 132 sets the model of the determination model so that the information indicating that the ruminant state or the information indicating the non-ruminant state (classification probability) output from the determination model approaches the corresponding correct answer data. Update the parameters. As described above, the learning unit 132 relates to the determination model that specifies the correspondence between the two-dimensional array data in the predetermined determination time range and the information indicating that the ruminant state is present or the information indicating that the ruminant state is present. Perform machine learning.

The model parameters of the learned determination model generated by the learning unit 132 performing machine learning are applied to the inference unit 134.

The inference data generation unit 133 acquires one-dimensional time-series data indicating acceleration in each of the three axial directions transmitted from the measuring device attached to the cow to be determined via the gateway device 120. Further, the inference data generation unit 133 extracts one-dimensional time-series data indicating acceleration in the z-axis direction from one-dimensional time-series data indicating acceleration in each of the three axis directions. Further, the inference data generation unit 133 generates two-dimensional array data indicating the intensity of each frequency at each time by frequency-analyzing the extracted one-dimensional time series data for each predetermined analysis time range.

Further, the inference data generation unit 133 associates the two-dimensional array data at each time with each time information and stores it in the inference data storage unit 136 as inference data.

The inference unit 134 is an example of the determination unit. The inference unit 134 reads the inference data from the inference data storage unit 136, and sequentially obtains the two-dimensional array data at each time included in the read inference data in a predetermined inference time range (with a predetermined determination time range). Input to the trained judgment model for each time range of the same length). Further, the inference unit 134 outputs the inference result (information indicating that it is in the ruminant state or information indicating that it is in the non-ruminant state) output from the learned determination model together with the corresponding time information.

As described above, in the determination device 130 according to the present embodiment, the amount of information used for determination is increased by frequency-analyzing the acceleration data which is one-dimensional time series data and generating the two-dimensional array data. Further, in the determination device 130 according to the present embodiment, by inputting the two-dimensional array data into the determination model and performing machine learning, it is possible to capture the characteristics of the ruminant behavior of the cow such as periodicity and continuity. Generate a judgment model for.

Thereby, according to the determination device 130 according to the present embodiment, it is possible to accurately determine the behavior classification of the cow (whether the cow is in the ruminant state or the non-ruminant state).

<Hardware configuration of judgment device>
Next, the hardware configuration of the determination device 130 will be described. FIG. 2 is a diagram showing an example of the hardware configuration of the determination device. As shown in FIG. 2, the determination device 130 includes a processor 201, a memory 202, an auxiliary storage device 203, an I / F (Interface) device 204, a communication device 205, and a drive device 206. The hardware of the determination device 130 is connected to each other via the bus 207.

The processor 201 has various arithmetic devices such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). The processor 201 reads various programs (for example, a determination program, etc.) onto the memory 202 and executes them.

The memory 202 has a main storage device such as a ROM (Read Only Memory) and a RAM (Random Access Memory). The processor 201 and the memory 202 form a so-called computer, and the processor 201 executes various programs read on the memory 202, so that the computer performs the above functions (learning data generation unit 131 to inference unit 134). Realize.

The auxiliary storage device 203 stores various programs and various data used when various programs are executed by the processor 201. For example, the learning data storage unit 135 and the inference data storage unit 136 are realized in the auxiliary storage device 203.

The I / F device 204 is a connection device that connects the operation device 210 and the display device 211, which are examples of external devices, and the determination device 130. The I / F device 204 receives an operation on the determination device 130 via the operation device 210. Further, the I / F device 204 outputs the result of processing by the determination device 130 and displays it on the display device 211.

The communication device 205 is a communication device for communicating with another device. In the case of the determination device 130, it communicates with the gateway device 120, which is another device, via the communication device 205.

The drive device 206 is a device for setting the recording medium 212. The recording medium 212 referred to here includes a medium such as a CD-ROM, a flexible disk, a magneto-optical disk, or the like that optically, electrically, or magnetically records information. Further, the recording medium 212 may include a semiconductor memory or the like for electrically recording information such as a ROM or a flash memory.

The various programs installed in the auxiliary storage device 203 are installed, for example, by setting the distributed recording medium 212 in the drive device 206 and reading the various programs recorded in the recording medium 212 by the drive device 206. Will be done. Alternatively, the various programs installed in the auxiliary storage device 203 may be installed by being downloaded from the network via the communication device 205.

<Specific example of functional configuration and processing of the learning data generation unit>
Next, among the functions realized by the determination device 130, the functional configuration of the learning data generation unit 131 and a specific example of the processing by the learning data generation unit 131 will be described. FIG. 3 is a first diagram showing a specific example of the functional configuration and processing of the learning data generation unit. As shown in FIG. 3, the learning data generation unit 131 includes an acceleration data acquisition unit 301, a z-axis direction extraction unit 302, a frequency analysis unit 303, and a data generation unit 304.

The acceleration data acquisition unit 301 acquires one-dimensional time-series data indicating acceleration in each of the three axial directions transmitted from the measuring device 110 via the gateway device 120. In FIG. 3, graph 311 is an example of one-dimensional time-series data indicating acceleration in each of the three axial directions acquired by the acceleration data acquisition unit 301, with the horizontal axis indicating time and the vertical axis indicating acceleration. ..

As shown in Graph 311, the one-dimensional time-series data indicating the acceleration in each of the three axial directions acquired by the acceleration data acquisition unit 301 of the learning data generation unit 131 shows that the cow is in a rebellious state (or non-reflexive state). It is assumed that the information indicating that the state) is associated in advance.

The z-axis direction extraction unit 302 extracts one-dimensional time-series data indicating acceleration in the z-axis direction from the one-dimensional time-series data indicating acceleration in each of the three axes directions acquired by the acceleration data acquisition unit 301. In FIG. 3, graph 312 is an example of one-dimensional time-series data showing acceleration in the z-axis direction extracted from one-dimensional time-series data showing acceleration in each of the three axes by the z-axis direction extraction unit 302. Is.

The frequency analysis unit 303 frequency-analyzes one-dimensional time-series data indicating acceleration in the z-axis direction extracted by the z-axis direction extraction unit 302 for each predetermined analysis time range, and determines the intensity of each frequency at each time. Generate the two-dimensional array data shown. In FIG. 3, graph 313 is an example of two-dimensional array data generated by the frequency analysis unit 303, with the horizontal axis representing time and the vertical axis representing frequency. Further, the difference in color in the graph 313 represents the difference in the intensity of each frequency at each time. For example, in Graph 313, red has the highest intensity, and hereafter, the intensity decreases in the order of orange → yellow → green → blue → purple. However, in the present embodiment, the difference in intensity of each frequency at each time is treated as a gray scale (for example, 0 to 255) (that is, it is not a value of three types of R, G, and B (0 to 255)). It shall be treated as one type of shade value (0 to 255)).

The rectangle shown by the broken line in the graph 312 indicates a predetermined analysis time range, and the one-dimensional time series data included in the analysis time range is frequency-analyzed, so that the corresponding time is set in the graph 313. Two-dimensional array data (rectangular indicated by a broken line) is generated.

The data generation unit 304 extracts the two-dimensional array data generated by the frequency analysis unit 303 at each time for each predetermined determination time range. In FIG. 3, graph 314 shows how the data generation unit 304 extracts the two-dimensional array data for each predetermined determination time range (see reference numeral 315).

The data generation unit 304 sequentially extracts two-dimensional array data while shifting a predetermined determination time range (for example, 15 seconds) at a predetermined shift interval (for example, 1 second interval).

Further, the data generation unit 304 associates the extracted two-dimensional array data of each determination time range with the information indicating that it is in the rebellious state or the information indicating that it is in the non-rebellious state, and learns it as learning data. It is stored in the data storage unit 135.

In FIG. 3, the learning data 320 is an example of the learning data stored by the data generation unit 304, and includes "image data" and "correct answer data" as information items.

The "image data" stores the two-dimensional array data of each determination time range extracted by the data generation unit 304. In the "correct answer data", either information indicating that the ruminant state is present or information indicating that the ruminant state is present is stored as the correct answer data in the corresponding determination time range.

<Functional configuration of the learning department>
Next, among the functions realized by the determination device 130, the functional configuration of the learning unit 132 will be described. FIG. 4 is a diagram showing an example of the functional configuration of the learning unit 132. As shown in FIG. 4, the learning unit 132 has a CNN unit 401 and a comparison / change unit 402.

The learning unit 132 reads the learning data 320 from the learning data storage unit 135, and inputs each two-dimensional array data included in the "image data" of the read learning data 320 to the CNN unit 401. Further, the learning unit 132 reads the learning data 320 from the learning data storage unit 135, and is included in the "correct answer data" of the read learning data 320, and is information indicating that the learning data is in a rut state or a non-warping state. Information indicating that is input to the comparison / change unit 402.

The CNN unit 401 is, for example, a determination model formed by a convolutional neural network, and when two-dimensional array data is input, it outputs information indicating that it is in a reflexive state or information indicating that it is in a non-reflexive state. do.

Further, the CNN unit 401 updates the model parameters by the comparison / change unit 402 according to the output of the information indicating the ruminant state or the information indicating the non-ruminant state.

When the comparison / change unit 402 outputs the information indicating that it is in the ruminant state or the information indicating that it is in the non-ruminant state by the CNN unit 401, it compares it with the correct answer data input by the learning unit 132 and makes an error. (Error of classification probability) is calculated. Further, the comparison / change unit 402 back-propagates the calculated error and updates the model parameters of the CNN unit 401.

<Specific example of functional configuration and processing of data generation unit for inference>
Next, among the functions realized by the determination device 130, the functional configuration of the inference data generation unit 133 and a specific example of the processing by the inference data generation unit 133 will be described. FIG. 5 is a diagram showing a specific example of the functional configuration and processing of the inference data generation unit. As shown in FIG. 5, the inference data generation unit 133 includes an acceleration data acquisition unit 501, a z-axis direction extraction unit 502, a frequency analysis unit 503, and a data generation unit 504.

Since the acceleration data acquisition unit 501 to the data generation unit 504 of the inference data generation unit 133 have the same functions as the acceleration data acquisition unit 301 to the data generation unit 304 of the learning data generation unit 131. , The description is omitted here.

However, the one-dimensional time-series data indicating the acceleration in each of the three axial directions acquired by the acceleration data acquisition unit 501 is the time-series data transmitted from the measuring device attached to the cow to be determined, and is in a rebellious state (or). Information indicating that it is in a non-reflexive state) is not associated.

Therefore, in the data generation unit 504, the two-dimensional array data extracted while shifting the predetermined inference time range (see reference numeral 515) at a predetermined shift interval is used as the time information which is the start time of each inference time range. It is stored in the inference data storage unit 136 in association with each other.

The inference data 530 shows how the two-dimensional array data in the inference time range with time = t as the start time is stored. Further, the inference data 530 shows that the two-dimensional array data in the inference time range with time = t + α as the start time is stored. Further, it shows how the two-dimensional array data in the inference time range with time = t + β as the start time is stored.

Similar to the learning data generation unit 131, the predetermined inference time range is equal to the predetermined determination time range, for example, 15 seconds. The shift interval when the data generation unit 504 extracts the two-dimensional array data is equal to the shift interval when the data generation unit 304 extracts the two-dimensional array data, for example, 1 second.

<Specific examples of functional configuration and processing of the inference unit>
Next, among the functions realized by the determination device 130, the functional configuration of the inference unit 134 and specific examples of the processing of the inference unit 134 will be described. FIG. 6 is a diagram showing a specific example of the functional configuration and processing of the inference unit. As shown in FIG. 6, the inference unit 134 has a CNN unit 601 and a determination result output unit 602.

The inference unit 134 reads out the inference data 520 from the inference data storage unit 136, and inputs the two-dimensional array data included in the "image data" of the read inference data 520 into the CNN unit 601.

The model parameters of the learned determination model generated by performing machine learning by the learning unit 132 are applied to the CNN unit 601. The CNN unit 601 outputs an inference result (information indicating that it is in a ruminant state or information indicating that it is in a non-ruminant state) by inputting two-dimensional array data.

The determination result output unit 602 arranges and outputs the inference results output by the CNN unit 601 along the time axis. In FIG. 6, the determination result 610 shows how the inference results output from the CNN unit 601 are arranged along the time axis based on the time information.

<Flow of judgment processing>
Next, the flow of the entire determination process by the determination device 130 will be described. FIG. 7 is a flowchart showing the flow of the determination process.

As shown in FIG. 7, the determination device 130 first executes the learning phase. Specifically, in step S701, the acceleration data acquisition unit 301 of the learning data generation unit 131 is associated with information indicating that it is in a rebellious state or information indicating that it is in a non-reflexive state in the three-axis direction. Acquire one-dimensional time-series data showing each acceleration.

In step S702, the z-axis direction extraction unit 302 of the learning data generation unit 131 has the one-dimensional time-series data indicating the acceleration in the z-axis direction from the acquired one-dimensional time-series data indicating the accelerations in each of the three-axis directions. Is extracted. Further, the frequency analysis unit 303 of the learning data generation unit 131 performs frequency analysis while shifting the one-dimensional time series data indicating the acceleration in the z-axis direction at a predetermined shift interval for each predetermined analysis time range. As a result, the frequency analysis unit 303 of the learning data generation unit 131 generates two-dimensional array data indicating the intensity of each frequency at each time.

Further, the data generation unit 304 of the learning data generation unit 131 extracts the two-dimensional array data while shifting the predetermined determination time range at a predetermined shift interval. Further, the data generation unit 304 of the learning data generation unit 131 corresponds to the information indicating that the two-dimensional array data extracted for each predetermined determination time range is in the rebellious state or the information indicating that the non-rebellious state is present. It is attached and stored in the learning data storage unit 135 as learning data.

In step S703, the learning unit 132 compares the image data (two-dimensional array data) included in the learning data with the CNN unit 401 with the correct answer data (information indicating that it is in the rebellious state (or non-rebellious state)). It is input to the change unit 402 and machine learning is performed on the determination model.

When the machine learning is completed and the learned determination model is generated, the determination device 130 shifts to the inference phase. Specifically, in step S704, the acceleration data acquisition unit 501 of the inference data generation unit 133 is a one-dimensional time indicating acceleration in each of the three axial directions measured by a measuring device attached to the cow to be determined. Get series data.

In step S705, the z-axis direction extraction unit 502 of the inference data generation unit 133 increases the one-dimensional time-series data indicating the acceleration in the z-axis direction from the acquired one-dimensional time-series data indicating the accelerations in each of the three-axis directions. Is extracted. Further, the frequency analysis unit 503 of the inference data generation unit 133 performs frequency analysis while shifting the one-dimensional time series data indicating the acceleration in the z-axis direction at a predetermined shift interval for each predetermined analysis time range. As a result, the frequency analysis unit 503 of the inference data generation unit 133 generates two-dimensional array data indicating the intensity of each frequency at each time. Further, the data generation unit 504 of the inference data generation unit 133 extracts the two-dimensional array data while shifting the predetermined inference time range at a predetermined shift interval. Further, the data generation unit 504 of the inference data generation unit 133 stores the two-dimensional array data extracted for each predetermined inference time range in the inference data storage unit 136 as inference data in association with the time information. do.

In step S706, the inference unit 134 executes the trained determination model CNN unit 601 by inputting the image data (two-dimensional array data) included in the inference data. As a result, the CNN unit 601 outputs an inference result (information indicating that it is in a ruminant state or information indicating that it is in a non-ruminant state).

In step S707, the determination result output unit 602 of the inference unit 134 outputs the inference result output from the CNN unit 601 in association with the time information.

<Summary>
As is clear from the above description, the determination device 130 according to the first embodiment is
-Acquire one-dimensional time-series data indicating acceleration in each of the three axial directions, which is output from a measuring device attached to the neck of the cow.
-A two-dimensional array showing the intensity of each frequency at each time after extracting the one-dimensional time-series data showing the acceleration in the z-axis direction from the one-dimensional time-series data showing the acceleration in each of the three axes and analyzing the frequency. Generate data.
-Inputs two-dimensional array data for each predetermined judgment time range and outputs information indicating the action classification for each predetermined judgment time range (information indicating that the ruminant state is present or information indicating that the ruminant state is present). do.

As described above, in the determination device 130 according to the first embodiment, the amount of information used for determination is increased by frequency-analyzing the acceleration data which is one-dimensional time series data and generating the two-dimensional array data. Further, in the determination device 130 according to the present embodiment, by inputting the two-dimensional array data into the determination model and performing machine learning, it is possible to capture the characteristics of the ruminant behavior of the cow such as periodicity and continuity. Generate a judgment model for.

As a result, according to the determination device 130 according to the first embodiment, for example, the characteristic of the ruminant behavior of the cow, such as the continuation of a high intensity state near a predetermined frequency (having periodicity and continuity), is exhibited in the individual. It can be captured accurately regardless of the difference in frequency characteristics of each.

As a result, according to the determination device 130 according to the first embodiment, it is possible to accurately determine the behavior classification of the cow (whether the cow is in the ruminant state or the non-ruminant state).

[Second Embodiment]
In the first embodiment, when one-dimensional time-series data showing acceleration in the z-axis direction is extracted from one-dimensional time-series data showing acceleration in each of the three axes to generate two-dimensional array data. Was explained. On the other hand, in the second embodiment, time-series data in the x-axis direction, the y-axis direction, and the z-axis direction are extracted from the one-dimensional time-series data showing the accelerations in the three-axis directions, respectively. The case of generating the two-dimensional array data of the above will be described. Hereinafter, the second embodiment will be described focusing on the differences from the first embodiment.

<Specific example of functional configuration and processing of the learning data generation unit>
FIG. 8 is a second diagram showing a specific example of the functional configuration and processing of the learning data generation unit. The difference from FIG. 3 is that in the case of the learning data generation unit 800, the x-axis direction extraction unit 801 and the y-axis direction extraction unit 811 and the frequency analysis units 802 and 812 are provided, and the function of the data generation unit 820 is shown in FIG. This is different from the function of the data generation unit 304 of. Further, the difference from FIG. 3 is that the configuration of the learning data 830 stored in the learning data storage unit 135 is different from the configuration of the learning data 320 of FIG.

In FIG. 8, the graphs corresponding to the graphs 311 to 314 are omitted due to space limitations, but the graphs of the data output from each part of the learning data generation unit 800 are the same as those of the graphs 311 to 314. be.

The x-axis direction extraction unit 801 extracts time-series data indicating acceleration in the x-axis direction from the one-dimensional time-series data indicating acceleration in each of the three axes directions acquired by the acceleration data acquisition unit 301.

The y-axis direction extraction unit 811 extracts time-series data indicating acceleration in the y-axis direction from the one-dimensional time-series data indicating acceleration in each of the three axes directions acquired by the acceleration data acquisition unit 301.

The frequency analysis unit 802 frequency-analyzes the time-series data indicating the acceleration in the x-axis direction extracted by the x-axis direction extraction unit 801 for each predetermined analysis time range, and two-dimensionally indicates the intensity of each frequency at each time. Generate array data.

The frequency analysis unit 812 frequency-analyzes the time-series data indicating the acceleration in the y-axis direction extracted by the y-axis direction extraction unit 811 for each predetermined analysis time range, and two-dimensionally indicates the intensity of each frequency at each time. Generate array data.

The data generation unit 820 indicates that the two-dimensional array data for each predetermined determination time range, which is output from the frequency analysis units 802, 812, and 303, is information indicating that it is in the rebellious state or that it is in the non-rebellious state. It is stored in the learning data storage unit 135 in association with the information.

In FIG. 8, the learning data 830 is an example of the learning data stored by the data generation unit 820, and the information items are "image data (x)", "image data (y)", and "image data". (Z) "," Correct answer data "is included.

The "image data (x)" is two-dimensional array data of each determination time range extracted by the data generation unit 304, and is two-dimensional array data based on one-dimensional time-series data indicating acceleration in the x-axis direction. Is stored.

The "image data (y)" is two-dimensional array data of each determination time range extracted by the data generation unit 304, and is two-dimensional array data based on one-dimensional time-series data indicating acceleration in the y-axis direction. Is stored.

The "image data (z)" is two-dimensional array data of each determination time range extracted by the data generation unit 304, and is two-dimensional array data based on one-dimensional time-series data indicating acceleration in the z-axis direction. Is stored.

In the "correct answer data", either information indicating that the ruminant state is present or information indicating that the ruminant state is present is stored as the correct answer data in the corresponding determination time range.

<Functional configuration of learning unit, inference data generation unit, and inference unit>
Next, the functional configurations of the learning unit 132, the inference data generation unit 133, and the inference unit 134 in the second embodiment will be described.

Of these, the functional configuration of the learning unit 132 in the second embodiment is the same as the functional configuration of the learning unit 132 in the first embodiment shown in FIG. However, in the case of the learning unit 132 in the second embodiment, when inputting the learning data 830 read from the learning data storage unit 135,
-Two-dimensional array data included in "image data (x)" is 1ch (channel),
-Two-dimensional array data included in "image data (y)" is 2ch (channel),
-The two-dimensional array data included in the "image data (z)" is 3ch (channel),
By inputting the data to the CNN unit 401, machine learning is performed on the determination model (however, the two-dimensional array data of each channel is the two-dimensional array data to which the same time information is associated).

Further, the inference data generation unit 133 in the second embodiment has the same configuration as the learning data generation unit 800 in the second embodiment shown in FIG. However, the inference data generated by the inference data generation unit 133 includes "image data (x)", "image data (y)", "image data (z)", and "time information" as information items. "Is included.

Further, the inference unit 134 in the second embodiment has the same functional configuration as the inference unit 134 in the first embodiment shown in FIG. However, in the case of the inference unit 134 in the second embodiment, when inputting the inference data read from the inference data storage unit 136,
・ 1ch of 2D array data included in "image data (x)"
・ 2ch, 2ch, 2D array data included in "image data (y)"
-The two-dimensional array data included in the "image data (z)" is 3ch,
The inference processing is performed by inputting the data to the CNN unit 601 and the inference result is output (however, the two-dimensional array data of each channel is the two-dimensional array data to which the same time information is associated).

<Summary>
As is clear from the above description, the determination device 130 according to the second embodiment is
-Acquire one-dimensional time-series data indicating acceleration in each of the three axial directions, which is output from a measuring device attached to the neck of the cow.
・ After extracting the one-dimensional time-series data showing the acceleration in the x-axis direction, the y-axis direction, and the z-axis direction from the one-dimensional time-series data showing the acceleration in each of the three axis directions, frequency analysis is performed, and at each time. Two-dimensional array data showing the intensity of each frequency is generated.
-Information indicating the behavior classification for each predetermined judgment time range (information indicating that it is in a ruminant state or non-ruminant state) by inputting two-dimensional array data of multiple channels (3 channels) for each predetermined judgment time range. Information indicating) is output.

As described above, in the determination device 130 according to the second embodiment, the one-dimensional time series data indicating the accelerations in each of the three axial directions is frequency-analyzed, the two-dimensional array data is generated for each axis, and then the machine is used. Do learning.

Thereby, according to the determination device 130 according to the second embodiment, it is possible to accurately determine the behavior classification of the cow (whether the cow is in the ruminant state or the non-ruminant state).

[Third Embodiment]
In the second embodiment, the time series data in the x-axis direction, the y-axis direction, and the z-axis direction are extracted from the one-dimensional time-series data showing the accelerations in each of the three-axis directions, and the respective two-dimensional data are extracted. The case of generating array data has been described. Further, the two-dimensional array data at this time has been described as being handled in gray scale.

On the other hand, in the third embodiment, each two-dimensional array data is handled in color, and the two-dimensional array data is divided and processed for each color component (for each R value, G value, and B value). The case will be described. Hereinafter, the third embodiment will be described focusing on the differences from the second embodiment.

<Functional configuration of the learning data generator>
FIG. 9 is a third diagram showing a specific example of the functional configuration and processing of the learning data generation unit. The difference from FIG. 8 is that the learning data generation unit 900 has R value extraction units 901, 911, 921, G value extraction units 902, 912, 922, and B value extraction units 903, 913, 923. The function of the data generation unit 930 is different from the function of the data generation unit 820 of FIG. Further, the difference from FIG. 8 is that the configuration of the learning data 940 stored in the learning data storage unit 135 is different from the configuration of the learning data 830 of FIG.

The R value extraction units 901, 911, and 921 extract the two-dimensional array data of the R value from the two-dimensional array data output from the frequency analysis units 802, 812, and 303, respectively.

The G value extraction units 902, 912, and 922 extract the two-dimensional array data of the G value from the two-dimensional array data output from the frequency analysis units 802, 812, and 303, respectively.

The B value extraction units 903, 913, and 923 extract the B value two-dimensional array data from the two-dimensional array data output from the frequency analysis units 802, 812, and 303, respectively.

The data generation unit 930 is output from the R value extraction units 901, 911, 921, the G value extraction units 902, 912, 922, and the B value extraction units 903, 913, and 923, respectively, in two dimensions for each predetermined analysis time range. Get array data.

Further, the data generation unit 930 extracts two-dimensional array data while shifting a predetermined determination time range at a predetermined shift interval, and information indicating that the state is ruminant or information indicating that the state is non-ruminant. It is stored as learning data 940 in association with.

In FIG. 9, the learning data 940 is an example of the learning data stored by the data generation unit 930. The training data 940 contains "image data (x, R)", ~ "image data (x, B)", "image data (y, R)" ~ "image data (y, B)" as information items. "," Image data (z, R) "to" Image data (z, B) "," Correct answer data "is included.

The "image data (x, R)" to "image data (x, B)" are two-dimensional array data of each determination time range based on one-dimensional time-series data indicating acceleration in the x-axis direction. Two-dimensional array data of R value, G value, and B value are stored.

The "image data (y, R)" to "image data (y, B)" are two-dimensional array data of each determination time range based on one-dimensional time-series data indicating acceleration in the y-axis direction. Two-dimensional array data of R value, G value, and B value are stored.

The "image data (z, R)" to "image data (z, B)" are two-dimensional array data of each determination time range based on one-dimensional time-series data indicating acceleration in the z-axis direction. Two-dimensional array data of R value, G value, and B value are stored.

<Functional configuration of learning unit, inference data generation unit, and inference unit>
Next, the functional configurations of the learning unit 132, the inference data generation unit 133, and the inference unit 134 in the third embodiment will be described.

Of these, the functional configuration of the learning unit 132 in the third embodiment is the same as the functional configuration of the learning unit 132 in the first embodiment shown in FIG. However, in the case of the learning unit 132 in the third embodiment, when inputting the learning data 940 read from the learning data storage unit 135,
-Each two-dimensional array data included in "image data (x, R)" to "image data (x, B)" is 1ch to 3ch,
-The two-dimensional array data included in the "image data (y, R)" to "image data (y, B)" is 4ch to 6ch,
-The two-dimensional array data included in "image data (z, R)" to "image data (z, B)" is 7ch to 9ch,
Is input to the CNN unit 401 to perform machine learning on the determination model.

Further, the inference data generation unit 133 in the third embodiment has the same configuration as the learning data generation unit 900 in the third embodiment shown in FIG. However, the inference data generated by the inference data generation unit 133 has information as an item of information.
-"Image data (x, R)" to "Image data (x, B)",
-"Image data (y, R)" to "Image data (y, B)",
-"Image data (z, R)" to "Image data (z, B)",
"Time information",
Is included.

Further, the inference unit 134 in the third embodiment has the same functional configuration as the inference unit 134 in the first embodiment shown in FIG. However, in the case of the inference unit 134 in the third embodiment, when inputting the inference data read from the inference data storage unit 136,
-Each two-dimensional array data included in "image data (x, R)" to "image data (x, B)" is 1ch to 3ch,
-The two-dimensional array data included in the "image data (y, R)" to "image data (y, B)" is 4ch to 6ch,
-The two-dimensional array data included in "image data (z, R)" to "image data (z, B)" is 7ch to 9ch,
Is input to the CNN unit 601 to perform inference processing and output the inference result.

<Summary>
As is clear from the above description, the determination device 130 according to the third embodiment is
-Acquire one-dimensional time-series data indicating acceleration in each of the three axial directions, which is output from a measuring device attached to the neck of the cow.
・ After extracting the one-dimensional time-series data showing the acceleration in the x-axis direction, the y-axis direction, and the z-axis direction from the one-dimensional time-series data showing the acceleration in each of the three axis directions, frequency analysis is performed, and at each time. Two-dimensional array data showing the intensity of each frequency is generated.
-The generated two-dimensional array data is divided for each color component.
-Information indicating the behavior classification for each predetermined judgment time range (information indicating that it is in a ruminant state or non-ruminant state) by inputting two-dimensional array data of multiple channels (9 channels) for each predetermined judgment time range. Information indicating) is output.

In this way, the determination device 130 according to the third embodiment frequency-analyzes one-dimensional time-series data indicating acceleration in each of the three axial directions, generates two-dimensional array data for each axis, and for each color component. After dividing into, machine learning is performed.

Thereby, according to the determination device 130 according to the third embodiment, it is possible to accurately determine the behavior classification of the cow (whether the cow is in the ruminant state or the non-ruminant state).

[Fourth Embodiment]
In the first embodiment, it is assumed that the one-dimensional time series data indicating the acceleration in each of the three axial directions is associated with the information indicating the ruminant state or the information indicating the non-ruminant state. explained.

However, the information associated with the time series data is not limited to this. For example, instead of the information indicating that the ruminant state or the information indicating the non-ruminant state, the information indicating the behavior classification other than the information indicating the ruminant state or the information indicating the non-ruminant state is provided. It may be associated.

Alternatively, in addition to the information indicating the ruminant state or the information indicating the non-ruminant state, the behavior classification other than the information indicating the ruminant state or the information indicating the non-ruminant state is shown. Information may be associated.

In addition, the information indicating the behavior classification other than the information indicating the ruminant state or the information indicating the non-ruminant state mentioned here includes, for example, the information indicating the feeding state and the walking state. Information and the like indicating the above are included.

FIG. 10 is a fourth diagram showing a specific example of the functional configuration and processing of the learning data generation unit. As shown in graphs 1011 to 1013 of FIG. 10, when the information indicating the behavior classification 1 or the information indicating the behavior classification 2 is associated, the "correct answer data" of the learning data 1020 is associated with any of the behavior classifications. Is stored.

As a result, according to the determination device 130 according to the fourth embodiment, the behavior classification of the cow to be determined can be accurately determined based on the one-dimensional time series data indicating the acceleration in each of the three axial directions. ..

[Fifth Embodiment]
In the first to fourth embodiments, the determination result output unit 602 has been described as outputting the inference result output from the CNN unit 601 in association with the time information. Further, in the first to fourth embodiments, the determination result output unit 602 has been described as outputting the inference result for each shift interval.

On the other hand, the interval at which the cow transitions from the ruminant state to the non-ruminant state, or the interval at which the cow transitions from the non-ruminant state to the ruminant state, is sufficiently longer than the shift interval.

Therefore, in the fifth embodiment, when the inference result at a predetermined time is output, the inference results for a plurality of inference time ranges including the predetermined time are referred to, and the final inference result is obtained based on the referenced plurality of inference results. Determine the inference result.

FIG. 11 is a second diagram showing a specific example of the functional configuration and processing of the inference unit. The difference from the inference unit 134 in the first embodiment shown in FIG. 6 is that in the case of FIG. 11, it has a statistical processing unit 1100.

The statistical processing unit 1100 is an example of a determination unit. The statistical processing unit 1100 acquires the inference result for each predetermined inference time range output from the CNN unit 601. In FIG. 11, the inference result 1110 shows the inference results for a plurality of inference time ranges 1121 to 1125 including a predetermined time 1111. The statistical processing unit 1100 determines the inference result of the predetermined time 1111 included in the inference result 1110 based on the inference results for the plurality of inference time ranges 1121 to 1125.

Specifically, the statistical processing unit 1100 determines an inference result having a high inference frequency as an inference result at a predetermined time 1111 (that is, the final inference result is determined by a majority vote). In the example of FIG. 11, the case where the information indicating the non-ruminant state is output as the inference result is once, and the case where the information indicating the ruminant state is output as the inference result is four times. Therefore, the statistical processing unit 1100 determines the inference result at the predetermined time 1111 as the information indicating the ruminant state.

In this way, when a plurality of inference results are output for the same time, in the determination process according to the fifth embodiment, the final inference result at each time is obtained by performing statistical processing on the plurality of inference results. To determine.

Thereby, according to the determination device 130 according to the fifth embodiment, it is possible to accurately determine the behavior classification of the cow (whether the cow is in the ruminant state or the non-ruminant state).

In the above description, in determining the final inference result at the predetermined time 1111, the inference results for a plurality of inference time ranges including the predetermined time 1111 are targeted, but the target used for the determination is limited to this. Not done. For example, a predetermined time 1111 based on inference results for a predetermined number of inference time ranges before or after the predetermined time 1111 regardless of whether or not the predetermined time 1111 is included. You may determine the final inference result in.

In this way, by determining the final inference result at each time based on the inference results for a predetermined number of inference time ranges, the noise contained in the two-dimensional array data generated when the inference time range is shortened. The influence of can be eliminated.

That is, when the inference time range is shortened, the time resolution is improved as compared with the case where the inference time range is lengthened, but it is easily affected by noise. However, according to the statistical processing unit 1100, the influence of noise is eliminated. can do.

[Other Embodiments]
In the first embodiment, the case of extracting the time-series data showing the acceleration in the z-axis direction from the one-dimensional time-series data showing the accelerations in the three-axis directions has been described. However, instead of the one-dimensional time-series data showing the acceleration in the z-axis direction, the one-dimensional time-series data showing the acceleration in the x-axis direction or the y-axis direction may be extracted.

That is, one or more of the one-dimensional time-series data showing the accelerations in the three-axis directions to the one-dimensional time-series data showing the accelerations in the x-axis direction, the y-axis direction, and the z-axis direction. Time series data may be extracted.

Further, in the third embodiment, the one-dimensional time series data indicating the acceleration is divided into three axes for frequency analysis, two-dimensional array data is generated for each axis, and each two-dimensional array data is further obtained. The case of dividing each color component has been described. However, the two-dimensional array data generated for any of the one-dimensional time-series data indicating the acceleration in the x-axis direction, the y-axis direction, and the z-axis direction extracted from the one-dimensional time-series data indicating the acceleration is used as a color component. It may be divided for each.

Further, in each of the above embodiments, when generating two-dimensional array data indicating the intensity of each frequency at each time, the value obtained by the frequency analysis is imaged (converted to gray scale or R, G, B). It was explained as being treated as "image data".

However, the method of generating the two-dimensional array data is not limited to this, for example, the value obtained by the frequency analysis is not imaged (not converted to grayscale, R, G, B), and each frequency at each time. It may be configured to generate two-dimensional array data indicating the strength of.

Further, in each of the above embodiments, the acceleration sensor is attached to the neck portion, but the attachment portion of the acceleration sensor is not limited to the neck portion and may be attached to another portion.

It should be noted that the present invention is not limited to the configurations shown here, such as combinations with other elements in the configurations and the like described in the above embodiments. These points can be changed without departing from the spirit of the present invention, and can be appropriately determined according to the application form thereof.

This application claims its priority based on Japanese Patent Application No. 2020-081460 filed on May 1, 2020, and by referring to the entire contents of the Japanese patent application, the present application is made. Invite.

100: Judgment system 110: Measuring device 130: Judgment device 131: Learning data generation unit 132: Learning unit 133: Inference data generation unit 134: Inference unit 301: Acceleration data acquisition unit 302: z-axis direction extraction unit 303: Frequency Analysis unit 304: Data generation unit 320: Learning data 401: CNN unit 402: Comparison / change unit 501: Acceleration data acquisition unit 502: z-axis direction extraction unit 503: Frequency analysis unit 504: Data generation unit 520: Inference data 601: CNN unit 602: Judgment result output unit 830: Learning data 940: Learning data 1020: Learning data 1100: Statistical processing unit

Claims (8)

1. A generator that frequency-analyzes one-dimensional time-series data indicating acceleration output from an acceleration sensor attached to an anticorrosive animal and generates two-dimensional array data indicating the intensity of each frequency at each time.
   A determination device having a determination unit for determining the behavior classification of the ruminant for each predetermined time range by inputting the two-dimensional array data for each predetermined time range.
2. The generator
   Of the one-dimensional time-series data indicating the acceleration in the x-axis direction, the one-dimensional time-series data indicating the acceleration in the y-axis direction, and the one-dimensional time-series data indicating the acceleration in the z-axis direction output from the acceleration sensor. The determination device according to claim 1, wherein one or more of the two-dimensional array data is generated by frequency-analyzing one or more one-dimensional one-dimensional time-series data.
3. The generator
   The determination device according to claim 2, wherein a plurality of the two-dimensional array data is generated by dividing the generated two-dimensional array data for each color component.
4. The determination unit
   The third aspect of claim 3, wherein the behavior classification of the rut animal for each predetermined time range is determined by inputting the two-dimensional array data of one or a plurality of channels including the generated one or a plurality of the two-dimensional array data. Judgment device.
5. The determination device according to claim 1, wherein the behavior classification includes at least either information indicating that the ruminant is in a ruminant state or information indicating that the ruminant is in a non-ruminant state.
6. The determination unit
   The determination according to claim 1, further comprising a learned determination model generated by performing machine learning on a determination model that specifies the correspondence between the two-dimensional array data and the behavior classification for each predetermined time range of the rutant. Device.
7. When there are a plurality of the predetermined time ranges including the predetermined time, the determination unit further has a determination unit for determining the action classification at the predetermined time by statistically processing the result determined by the determination unit for each time range. The determination device according to claim 1.
8. A generation process that frequency-analyzes one-dimensional time-series data indicating acceleration output from an acceleration sensor attached to an anticorrosive animal and generates two-dimensional array data indicating the intensity of each frequency at each time.
   A determination program for causing a computer to perform a determination step of determining the behavior classification of the rut animal for each predetermined time range by inputting the two-dimensional array data for each predetermined time range.

Images