Respiratory disease diagnosis system and ear tag sensor

Abstract

To provide a respiratory disease diagnosis system and ear tag sensor that can more accurately and early determine the presence of respiratory disease in a living organism, regardless of individual differences in living organisms. [Solution] The respiratory disease determination system 100 includes a determination unit 40 that determines whether or not each of several pigs P has a respiratory disease. The determination unit 40 includes autoencoders 50A, 50B, 50C, ... which compress learning time-series data representing the characteristics of internal conduction sound in a healthy pig P, learn to restore the learning time-series data, and then input and restore inspection time-series data representing the characteristics of internal conduction sound in the same pig P; and detection units 50A, 50B, 50C, ... which detect whether or not a pig has a respiratory disease based on the difference between the inspection time-series data input to the autoencoders 50A, 50B, 50C, ... and the inspection time-series data restored by the autoencoders 50A, 50B, 50C, ... for each pig P.

Description

[Technical Field]

[0001] The present invention relates to a respiratory disease determination system and an ear tag sensor. [Background Art]

[0002] In Japan's livestock industry, scale expansion has progressed: while the number of pig farmers has decreased, the number of pigs raised per household has increased. Such a situation may increase the risk of large-scale damage caused by the rapid spread of infectious diseases such as ASF (African swine fever). That is, appropriate health management and early detection of infectious diseases are becoming increasingly difficult. Surprisingly many pigs suffer from respiratory diseases such as pneumonia, and such respiratory diseases hinder the growth of affected pigs and surrounding pigs, delay shipping, and reduce the cost-effectiveness of farmers. It is considered that damage can be minimized if countermeasures such as treatment can be taken at an early stage through early detection. Therefore, a system that can detect an infection at an early stage is needed. Accordingly, an infection determination model using LSTM (Long Short-Term Memory) has been proposed, which receives as input the cepstrum and delta cepstrum of in-body transmitted sounds of pigs (see, for example, Non-Patent Document 1). [Prior Art Literature] [Non-Patent Literature]

[0003] [Non-Patent Document 1] Yuki Matoba et al., "Construction of a Pig Infection Determination Model Using In-Body Transmitted Sounds", Proceedings of the Acoustical Society of Japan, 2023 [Summary of the Invention] [Problem to be Solved by the Invention]

[0004] However, there are individual differences among living organisms such as pigs, and the patterns of in-body transmitted sounds vary even particularly when the individual is healthy. The individual differences in these in-body transmitted sounds are a factor that causes a decrease in the identification accuracy of respiratory diseases by the infection determination model.

[0005] This invention was made under the circumstances described above, and aims to provide a respiratory disease diagnosis system and ear tag sensor that can more accurately and early determine whether a living organism has a respiratory disease, regardless of individual differences in living organisms. [Means for solving the problem]

[0006] To achieve the above objective, the respiratory disease diagnosis system according to the first aspect of the present invention is: It is equipped with a determination unit that determines whether or not each of multiple living organisms has a respiratory disease. The determination unit, An autoencoder that first compresses time-series data for training that represents the characteristics of internal conduction sounds in a healthy pig P, then learns to reconstruct the said time-series data for training, and then inputs and reconstructs time-series data for testing that represents the characteristics of internal conduction sounds in the same organism. A detection unit detects whether or not a person has a respiratory disease based on the difference between the time-series data for testing input to the autoencoder and the time-series data for testing restored by the autoencoder. Each organism is equipped with its own unique features.

[0007] The aforementioned autoencoder is an LSTM (Long Short-Term Memory) autoencoder. It would be acceptable to do so.

[0008] The autoencoder is a PerceiverIO autoencoder. It would be acceptable to do so.

[0009] The system includes a feature calculation unit that calculates two-dimensional waveform data representing the time variation of the frequency spectrum of the internally conducted sound or the components obtained by Fourier transforming the frequency spectrum, as data representing the characteristics of internally conducted sound in a living organism. The autoencoder inputs the two-dimensional waveform data calculated by the feature calculation unit as the time-series data. It would be acceptable to do so.

[0010] An ear tag sensor according to a second aspect of the present invention is: An ear tag sensor that is attached to the ear of a living organism, A piezoelectric element that adheres closely to the ear of a living organism, A retaining plate is positioned to sandwich the ear portion between the piezoelectric element and the piezoelectric element, A circuit board on which a signal circuit is mounted that amplifies the signal detected by the piezoelectric element and wirelessly outputs the amplified signal, and a battery that supplies power to the signal circuit, and a housing that holds the piezoelectric element, A through rod extending through the piezoelectric element, the ear portion, and the retaining plate, with its tip engaging with the housing, A spring portion, in a compressed state, is sandwiched between the base end of the through rod and the retaining plate, and biases the retaining plate toward the lug portion, Equipped with, In the aforementioned housing, the circuit board and the battery are positioned at locations offset from a straight line along the longitudinal direction of the through-rod.

[0011] The aforementioned spring portion is The aforementioned through rod is composed of a hollow cylindrical elastic member that penetrates the interior. The through rod is curved spherically along the direction in which it extends, and has a shape that is expandable and contractible in the direction in which the through rod extends. The distance between the piezoelectric element and the retaining plate is formed such that the rate of change of elastic force when it is within a range that can be taken by the thickness of a living ear is smaller than the rate of change when it is in other ranges. It would be acceptable to do so.

[0012] The piezoelectric element is A plate-shaped first electrode layer made of a conductive material having a first through-hole penetrating in the thickness direction, A plate-shaped piezoelectric layer made of electrostrictive material is laminated on the first electrode layer, having a second through-hole that penetrates in the thickness direction, and the second through-hole is coaxial with the first through-hole. A plate-shaped second electrode layer made of a conductive material is laminated on the piezoelectric layer, having a third through-hole that penetrates in the thickness direction of the plate, and the third through-hole is coaxial with the first through-hole. a plate-shaped sealing layer made of an insulating material, the sealing layer having a fourth through hole penetrating in the plate thickness direction and being laminated on the second electrode layer such that the fourth through hole communicates with the first through hole; when viewed in the lamination direction of the first electrode layer, the piezoelectric layer, the second electrode layer and the sealing layer, the sealing layer and the first electrode layer enclose the piezoelectric layer and the second electrode layer, the piezoelectric layer and the second electrode layer are sealed by the sealing layer and the first electrode layer, and this configuration may be adopted. Effects of the Invention

[0013] According to the present invention, it is possible to more accurately and early determine whether a living organism is affected by a respiratory disease regardless of individual differences among living organisms. Brief Description of the Drawings

[0014] [Figure 1] It is a block diagram showing the configuration of the respiratory disease determination system according to Embodiment 1 of the present invention. [Figure 2] It is a schematic diagram showing the configuration of an ear tag sensor. [Figure 3] (A) is a top view of the piezoelectric element. (B) is a cross-sectional view taken along line III-III of (A). [Figure 4] It is an exploded perspective view of a housing constituting the ear tag sensor. [Figure 5] It is a graph showing the characteristics of a spring portion constituting the ear tag sensor. [Figure 6] It is a perspective view of a part of a tag pliers for attaching the ear tag sensor to an ear part. [Figure 7] It is a timing chart showing the timing of detecting bioconducted sound in the ear tag sensor. [Figure 8] It is a block diagram showing the configuration of an information processing unit. [Figure 9] It is a diagram showing changes in the spectrum of internally conducted sound in three pigs when they are healthy and after contracting the disease (7th day after inoculation). [Figure 10](A) is a diagram showing the operation of the autoencoder when normal data is input. (B) is a diagram showing the operation of the autoencoder when abnormal data is input. [Figure 11] This is a schematic diagram showing the configuration of an LSTM autoencoder. [Figure 12] This figure shows an example of the probability density of reconstruction errors. [Figure 13] Figure 8 is a block diagram showing the hardware configuration of the information processing unit. [Figure 14] This is a flowchart of the learning process. [Figure 15] This is a flowchart of the file management unit's processing during the judgment process. [Figure 16] This is a flowchart of the processing steps taken by the information processing unit and the user terminal during the decision-making process. [Figure 17] (A) to (C) are graphs showing examples of the discrimination rate between healthy data and diseased data. [Figure 18] (A) and (B) are diagrams illustrating the Kullback-Leibler divergence (KL divergence). [Figure 19] Figures (A) to (D) compare the disease identification rate when a threshold is set based on reconstruction error and when a threshold is set based on KL divergence. [Figure 20] This is a block diagram showing the configuration of PerceiverIO, which constitutes the respiratory disease diagnosis system according to Embodiment 2 of the present invention. [Figure 21] This figure compares the disease identification rate when a threshold is set based on reconstruction error and when a threshold is set based on KL divergence. [Modes for carrying out the invention]

[0015] Embodiments of the present invention will be described in detail below with reference to the drawings. In each drawing, the same or equivalent parts are denoted by the same reference numerals. In the embodiments described below, expressions such as “having,” “including,” or “containing” also include the meaning of “consisting of” or “composed of.”

[0016] Embodiment 1 First, Embodiment 1 of the present invention will be described. As shown in Figure 1, the respiratory disease determination system 100 according to this embodiment determines whether or not each of several pigs P, which are livestock raised in groups in a barn, are suffering from a respiratory disease. In this embodiment, pigs P correspond to living organisms. The respiratory disease determination system 100 comprises an ear tag sensor 1, a router 2, a file management unit 3, an information processing unit 4, and a user terminal 5. The user terminal 5 is, for example, a terminal owned by a farmer who raises pigs P.

[0017] [Ear tag sensor] Pigs P are fitted with ear tags for individual identification. In this respiratory disease diagnosis system 100, an ear tag sensor 1 is used, which integrates a sensor with the ear tag to facilitate easy introduction by farmers. The ear tag sensor 1 is attached to the ear Y of the pig P. As shown in Figure 2, the ear tag sensor 1 comprises a piezoelectric element 10, a retaining plate 11, a through rod 12, a housing 13, and a spring part 14. The ear tag sensor 1 is attached to the ear Y by sandwiching the ear Y between the piezoelectric element 10 and the retaining plate 11. A thin film sheet (not shown) is inserted between the piezoelectric element 10 and the ear Y.

[0018] [Piezoelectric element] The piezoelectric element 10 is in close contact with the ear portion Y of the pig P. The piezoelectric element 10 detects the signal of internally conducted sound in the pig P. As shown in Figures 3(A) and 3(B), the piezoelectric element 10 is a disc-shaped element with a through hole formed in the center. The piezoelectric element 10 comprises a first electrode layer 10a, a piezoelectric layer 10b, a second electrode layer 10c, and a sealing layer 10d.

[0019] [First electrode layer] The first electrode layer 10a is a plate-shaped member made of a conductive material, such as metal, having a first through-hole 20a that penetrates in the thickness direction of the plate. In this embodiment, the first electrode layer 10a is disc-shaped, and the first through-hole 20a is a circular hole centered on the center O of the first electrode layer 10a. A soldering portion 10e for connecting to a cable is provided on the outer circumferential surface of the first electrode layer 10a. This cable is connected to a signal circuit on the circuit board 21 shown in Figure 2.

[0020] [Piezoelectric layer] The piezoelectric layer 10b is a plate-shaped member made of electrostrictive material, having a second through-hole 20b that penetrates in the thickness direction of the plate, and laminated on the first electrode layer 10a such that the second through-hole 20b is coaxial with the first through-hole 20a. In this embodiment, the piezoelectric layer 10b is disc-shaped, and its center O coincides with the first electrode layer 10a. The second through-hole 20b is a circular hole centered on the center O of the piezoelectric layer 10b.

[0021] [Second electrode layer] The second electrode layer 10c is a plate-shaped member made of a conductive material, such as silver, laminated on the piezoelectric layer 10b such that the third through-hole 20c penetrates in the thickness direction of the plate and the third through-hole 20c is coaxial with the first through-hole 20a. In this embodiment, the second electrode layer 10c is disc-shaped, and its center O coincides with the first electrode layer 10a and the piezoelectric layer 10b. The third through-hole 20c is a circular hole centered on the center O of the second electrode layer 10c. A soldering portion 10e for connecting to a cable is provided on the outer circumferential surface of the second electrode layer 10c. This cable is connected to a signal circuit on the circuit board 21 shown in Figure 2. The signal circuit detects a signal indicating the potential difference between the first electrode layer 10a and the second electrode layer 10c via the two cables described above.

[0022] [Seal layer] The sealing layer 10d is a plate-shaped protective sheet made of an insulating material, laminated on the second electrode layer 10c such that the fourth through-hole 20d penetrates in the thickness direction and communicates with the first through-hole 20a. In this embodiment, the sealing layer 10d is disc-shaped, and its center O coincides with the first electrode layer 10a, the piezoelectric layer 10b, and the second electrode layer 10c. The fourth through-hole 20d is a circular hole centered on the center O of the sealing layer 10d.

[0023] As shown in Figure 3(A), when viewed in the stacking direction of the first electrode layer 10a, piezoelectric layer 10b, second electrode layer 10c, and seal layer 10d, the seal layer 10d and the first electrode layer 10a enclose the piezoelectric layer 10b and the second electrode layer 10c. Furthermore, as shown in Figure 3(B), the piezoelectric layer 10b and the second electrode layer 10c are sealed by the seal layer 10d and the first electrode layer 10a. This prevents short circuits between the first electrode layer 10a and the second electrode layer 10c, for example, due to blood flowing from the ear Y of a pig P. To prevent a decrease in sealing performance, the seal layer 10d is provided inside the soldered portion 10e of the first electrode layer 10a.

[0024] [Retaining plate] As shown in Figure 2, the retaining plate 11 is a plate-shaped member that is positioned coaxially with the piezoelectric element 10 on the opposite side of the piezoelectric element 10 from the lug Y. In other words, the retaining plate 11 is positioned to sandwich the lug Y between itself and the piezoelectric element 10. The retaining plate 11 also has a through hole in its center that penetrates through in the direction of its thickness.

[0025] [Penetrating rod] The through rod 12 is made of an insulating and slightly elastic material, such as resin. As shown in Figure 2, the through rod 12 comprises a base 12a, a rod-shaped portion 12b, and a return portion 12c. The base 12a is a disc-shaped base. The rod-shaped portion 12b is a cylindrical member extending from the center of the base 12a in the direction normal to the plate surface of the base 12a. The rod-shaped portion 12b is hollow. The return portion 12c is a portion at the tip of the rod-shaped portion 12b that is slightly thicker than the rod-shaped portion 12b.

[0026] [Cabinet] As shown in Figure 2, the housing 13 is a housing that houses signal circuits and the like. The housing 13 is made of an insulating and rigid material, such as resin. The outer surface of the housing 13 displays an ear tag number (not shown), which is identification information for individual identification of pig P. An ear tag may also be attached to the housing 13.

[0027] As shown in Figures 2 and 4, the housing 13 comprises a container 13a, a lid 13b, and a shaft hole 13c. The container 13a houses the shaft hole 13c in its center. The top of the container 13a is open, and with the shaft hole 13c housed in it, the top is closed with the lid 13b, and the housing 13 is formed by bonding the container 13a and the lid 13b together. The lid 13b has a through hole into which the rod-shaped portion 12b of the through rod 12 is inserted, and the shaft hole 13c has a hole into which the return portion 12c of the through rod 12 is inserted. The rod-shaped portion 12b of the through rod 12 extends through the piezoelectric element 10, the lugs Y, and the retaining plate 11, and the return portion 12c at its tip is locked to the housing 13.

[0028] As shown in Figure 2, a recess 13d for fitting the piezoelectric element 10 is formed in the center of the lid portion 13b. The piezoelectric element 10 is held in the housing 13 so as to be in close contact with the ear portion Y of the pig P and capable of vibrating in response to internally conducted sound transmitted from the ear portion Y.

[0029] As shown in Figure 2, in the container 13a, a circuit board 21 is housed in the internal space on one side of the axial hole 13c. The circuit board 21 is equipped with an amplifier that amplifies the signal detected by the piezoelectric element 10, a DA conversion circuit that converts the amplified analog signal into a digital signal with a predetermined sampling frequency, and a signal circuit that wirelessly transmits the converted digital signal to an external device. The signal circuit amplifies the signal detected by the piezoelectric element 10 and wirelessly outputs the amplified signal using the WiFi (IEEE 802.11) standard or the like.

[0030] Furthermore, in the container 13a, a battery 22 is housed in the internal space on the other side of the axial hole 13c. The battery 22 supplies power to the piezoelectric element 10 and the signal circuit on the circuit board 21. To miniaturize the ear tag sensor 1, only the WiFi module is mounted on the front side of the circuit board 21, while the other components are mounted on the back side. In addition, the circuit board 21, battery 22, piezoelectric element 10, and other components are connected by soldering rather than connectors, and the housing 13 is designed to be as small as possible. Moreover, since the battery 22 does not require recharging, there is no need to prevent over-discharge, and due to its capacity, there is no concern about overheating or ignition in the event of a short circuit. Therefore, the battery 22 can be miniaturized by omitting a protection circuit to stop high-current discharge.

[0031] The signal circuit detects the voltage output from the piezoelectric element 10 at a predetermined sampling rate, for example, 50 Hz. The signal circuit wirelessly outputs data in which the signal data of the internal conduction sound of the pig P detected by the piezoelectric element 10 is supplemented with identification information of the pig P, i.e., data representing the ear tag number.

[0032] [Spring section] As shown in Figure 2, in a compressed state, the spring portion 14 is sandwiched between the base 12a, which is the base end of the through rod 12, and the retaining plate 11, biasing the retaining plate 11 toward the lug portion Y. One end of the spring portion 14 in the direction of expansion and contraction and the base 12a of the through rod 12 are bonded together with double-sided tape or the like. Alternatively, a step or offset may be used to interlock them and prevent them from coming loose.

[0033] As shown in Figure 2, the spring portion 14 is composed of a hollow cylindrical elastic member through which the through rod 12 passes. The spring portion 14 is curved spherically along the direction in which the through rod 12 extends and has a shape that allows it to expand and contract in the direction in which the through rod 12 extends.

[0034] The spring portion 14 compresses more as the thickness L of the ear portion Y increases, and the clamping force on the ear portion Y increases accordingly. On the other hand, the thickness L of the ear portion Y of the pig P increases as the pig P grows. As the thickness L of the pig P's ear increases, the amount of compression of the spring portion 14 increases accordingly, so it is necessary to prevent the force of the spring portion 14, i.e., the spring pressure, from becoming too strong. Therefore, the spring portion 14 is formed such that the rate of change of the elastic force when the distance between the piezoelectric element 10 and the retaining plate 11 is within the range that the ear thickness L of the pig P's ear portion Y can take is smaller than the rate of change when it is in other ranges. For example, as shown in Figure 5, the elastic force of the spring portion 14 does not change much from 3.5 mm to just under 10 mm, and remains within the range of 1 N to 2 N. In this respect, the elastic force characteristics of the spring portion 14 differ from the characteristics of a coil spring (dotted line) which changes linearly (the elastic modulus is constant). Thus, the spring portion 14 is configured to suppress the increase in spring pressure even when the ear thickness L increases due to the growth of the pig P, thereby minimizing damage to the ear portion Y. Note that the spring portion 14 is not limited to having the characteristics shown in Figure 5, and its elastic force can be adjusted as appropriate. The spring portion 14 only needs to have the minimum elastic force necessary to hold the ear tag sensor 1 to the ear portion Y of the pig P.

[0035] In the housing 13, the circuit board 21 and the battery 22 are positioned at locations offset from a straight line along the longitudinal direction of the through rod 12. This prevents the force generated by the spring portion 14 from directly applying to the circuit board 21 and the battery 22, thereby preventing damage to them.

[0036] As shown in Figure 6, the ear tag sensor 1 having such a structure can be attached using commercially available tag pliers 23 for attaching ear tags. The ear tag sensor 1 can be attached to the ear Y of a pig P by attaching the housing 13 with the piezoelectric element 10 mounted on one end of the tag pliers 23, and the retaining plate 11, through rod 12, and spring part 14 to the other end of the tag pliers 23, and then gripping the ear Y of the pig P with the pliers. In other words, the ear tag sensor 1 can be attached to the ear Y using commercially available tag pliers 23 without using any special attachment tools.

[0037] [router] Returning to Figure 1, Router 2 connects to the ear tag sensor 1 wirelessly using Wi-Fi (IEEE 802.11) or similar standards, and also connects to the file management unit 3 via a wired or wireless communication network. Router 2 wirelessly outputs detection commands transmitted from the file management unit 3 to the ear tag sensor 1, and functions as a relay station that receives waveform data of internally conducted sound, which is wirelessly output from the ear tag sensor 1 and assigned the individual's ear tag number, and transmits it to the file management unit 3. By using a wired communication network between Router 2 and the file management unit 3, it is possible to reduce packet loss of transmitted data.

[0038] [File Management Department] As shown in Figure 1, the file management unit 3 collects waveform data of internal conduction sound from ear tag sensors 1 attached to each of multiple pigs P via the router 2. Specifically, the file management unit 3 transmits a detection command to the ear tag sensor 1 via the router 2 at regular time intervals T1 (see Figure 7). Upon receiving the detection command, the ear tag sensor 1 detects internal conduction sound at a predetermined time T2 (see Figure 7) and wirelessly outputs the detected internal conduction sound waveform data to the router 2 with the ear tag number added. The file management unit 3 acquires the internal conduction sound waveform data via the router 2, files the acquired waveform data for each pig P, and manages it. The files managed by the file management unit 3 have the ear tag number added along with the waveform data, and by referring to the file, it is possible to determine which pig P the waveform data belongs to. The file management unit 3 transmits the internal conduction sound waveform data for each pig P, which has been filed, to the information processing unit 4 at a predetermined timing (for example, once a day).

[0039] [Information Processing Department] As shown in Figure 1, the information processing unit 4 receives the filed waveform data. Based on the received waveform data, the information processing unit 4 determines whether or not each pig P has a respiratory disease. As shown in Figure 8, the information processing unit 4 includes a feature calculation unit 30 and a determination unit 40.

[0040] [Feature Analysis Unit] The feature calculation unit 30 calculates two-dimensional waveform data representing the time variation of the frequency spectrum of the internal conduction sound or the components obtained by Fourier transforming the frequency spectrum, as data representing the characteristics of the internal conduction sound of pig P. Such two-dimensional waveform data includes the spectrogram of the internal conduction sound, delta spectrogram, wavelet transform data, cepstrum, and delta cepstrum. Spectrograms are generated when performing time-frequency analysis to analyze the time variation of the heartbeat and respiration bands from the internal conduction sound and confirm their changes. In pig P, the heartbeat is generally in a higher energy band than the respiration, and although there are individual differences, it is known to appear around 1-3 Hz. The feature calculation unit 30 resamples the internal conduction sound to 50 Hz to make it easier to extract biological signals, and divides the waveform with a frame length of 2 seconds or more, for example 2.56 seconds, and an overlap of 94% to ensure a time of more than one heartbeat cycle, and calculates a spectrogram using a Blackman window. A spectrogram is, for example, a two-dimensional data set with 65 rows and 16 columns.

[0041] Five-week-old pigs P, fitted with ear tag sensors 1 on their ear Y, were infected with Pasteurella bacteria, one of the causes of pneumonia, and their spectrograms were compared before and after infection. When comparing the spectrograms before infection (healthy) and after infection (several days after vaccination), as shown in Figure 9, on day 7 after vaccination, significant distortion occurred in the internal conduction sound due to coughing and respiratory sound stimuli. As a result, the spectrogram on day 7 after vaccination showed a large increase in harmonic components and a widening of the fundamental frequency band of the heartbeat in the internal conduction sound, indicating a significant change in the waveform of the spectrogram. In this embodiment, the feature calculation unit 30 calculates the spectrogram. When time waveform, spectrogram, and wavelet transform data are used as inputs, it has been shown that the LSTM encoder described later has the highest discrimination rate for the spectrogram.

[0042] [Judgment section] As shown in Figure 8, the determination unit 40 determines whether or not a pig P has respiratory disease for each of the multiple pigs P. As shown in Figure 8, the determination unit 40 is equipped with individual determination units 40A, 40B, 40C, ... for each pig P to be determined. That is, if the pigs P are pig P(1A), pig P(1B), pig P(1C), ..., then the determination unit 40 is equipped with an individual determination unit 40A for pig P(1A), an individual determination unit 40B for pig P(1B), an individual determination unit 40C, ... for pig P(1C). The individual determination units 40A, 40B, 40C, ... receive the two-dimensional waveform data calculated by the feature calculation unit 30 as time-series data. In the determination unit 40, the individual determination units 40A, 40B, 40C, ... determine whether or not each pig P (1A, 1B, 1C, ...) has a respiratory disease based on the difference between the input time-series data and the restored time-series data.

[0043] Individual determination unit 40A comprises an autoencoder 50A and a detection unit 51A, individual determination unit 40B comprises an autoencoder 50B and a detection unit 51B, and individual determination unit 40C comprises an autoencoder 50C and a detection unit 51C. The other individual determination units are similar.

[0044] As shown in Figures 10(A) and 10(B), the autoencoder 50A is a neural network composed of an encoder that compresses the input data and a decoder that reconstructs the input data. The size of the hidden layer is smaller than that of the input layer, and the size of the input layer and the output layer are the same. During training, since the size of the hidden layer is smaller than that of the input layer, the data input by the encoder is compressed. The decoder attempts to reconstruct the original data from the compressed data. These operations are the same for autoencoders 50B, 50C, ... Autoencoders 50A, 50B, 50C, ... perform input compression and reconstruction. Autoencoders 50A, 50B, 50C, ... are trained to compress training time-series data representing the characteristics of internal conduction sound in a healthy pig P, and then reconstruct the training time-series data. After training, autoencoders 50A, 50B, 50C, ... perform input and reconstruction of test time-series data representing the characteristics of internal conduction sound in the same pig P.

[0045] As shown in Figure 8, the detection units 51A, 51B, 51C, ... detect whether the pig P to be judged is healthy, i.e., normal, or suffering from a respiratory disease, i.e., abnormal, i.e., whether or not it has a respiratory disease, based on the magnitude of the difference between the input and output, for example, the absolute value (reconstruction error x). Abnormality detection by the individual judgment units 40A, 40B, 40C, ... is effective when abnormalities rarely occur.

[0046] In autoencoders 50A, 50B, 50C, etc., the data used for training (training time-series data) is expected to be faithfully reconstructed, as shown in Figure 10(A). However, data not used for training cannot be properly feature extracted, as shown in Figure 10(B), and cannot be reconstructed to its original state. In anomaly detection, by training only with normal data (waveform data of internal conduction sounds from healthy pigs), if the reconstruction error is small, it can be determined to be normal, and if the reconstruction error is large, it can be determined to be abnormal (suffering from a respiratory disease).

[0047] In this embodiment, the autoencoders 50A, 50B, 50C, ... are LSTM (Long Short-Term Memory) autoencoders. As shown in Figure 11, the LSTM autoencoders use LSTM layers (LSTM Blocks) capable of holding long- and short-term memories in the encoder and decoder, respectively, to learn the characteristics of time-series data and compress and decompress the data based on that learning.

[0048] As shown in Figure 8, the detection units 51A, 51B, 51C, ... detect whether or not a pig has a respiratory disease based on the difference between the time-series data for testing input to the autoencoders 50A, 50B, 50C, ... and the time-series data for testing reconstructed by the autoencoders 50A, 50B, 50C, ... (hereinafter simply referred to as "time-series data"). The detection units 51A, 51B, 51C, ... determine whether the pig is normal (healthy) or abnormal (suffering from a respiratory disease) based on the magnitude of the absolute value of the difference between the time-series data before compression and the reconstructed time-series data (reconstruction error x), and detect abnormalities. A threshold is used in the detection units 51A, 51B, 51C, ... to determine whether the pig is normal or abnormal. In this embodiment, as shown in Figure 12, the threshold is determined using the properties of a normal distribution. The probability density of the reconstruction error x is the mean μ and the variance σ. 2 The normal distribution N(μ, σ) 2 If the following is followed, the probability that the reconstruction error x falls within the μ±σ range is approximately 68%, the probability that the reconstruction error x falls within the μ±2σ range is approximately 95%, and the probability that the reconstruction error x falls within the μ±3σ range is approximately 99.7%. The detection units 51A, 51B, 51C, ... detect abnormalities by treating values ​​outside this σ range as outliers. In other words, in this embodiment, autoencoders (LSTM autoencoders) 50A, 50B, 50C, ... are used as the model for reconstruction, and the threshold μ+σ is set using the mean μ and standard deviation σ of the reconstruction error x of the waveform data of the internal conduction sound of a healthy pig P (normal data). That is, if the reconstruction error x exceeds μ+σ, it is determined to be abnormal. Note that the thresholds may also be set to μ+2σ and μ+3σ.

[0049] Generally, machine learning-based diagnosis requires learning the disease state. However, in practice, it is necessary to introduce a model based on the premise that there are no diseased pigs, so waveform data of internal conduction sounds (abnormal data or diseased data) from diseased pigs is required in advance. This respiratory disease diagnosis system 100 introduces autoencoders 50A, 50B, 50C, ... and incorporates an LSTM autoencoder into its algorithm, making it possible to determine the presence or absence of respiratory disease without learning using data from diseased pigs. The respiratory disease diagnosis system 100 makes a determination using autoencoders 50A, 50B, 50C, ... and detection units 51A, 51B, 51C, ... which are learned for each pig P, so it can determine the presence or absence of respiratory disease without considering individual differences.

[0050] [Hardware configuration of the information processing unit] The information processing unit 4 shown in Figure 1 is realized, for example, by a computer having the hardware configuration shown in Figure 13 executing a software program. Specifically, the information processing unit 4 includes a CPU (Central Processing Unit) 61 that controls the entire device, a main memory 62 that operates as a workspace for the CPU 61, an external memory 63 that stores the CPU 61's operating program, a human-machine interface 64, a communication interface 65, and an internal bus 68 that connects these.

[0051] As described later, the CPU 61 implements various functions of the information processing unit 4 by executing the program 69 stored in the main memory 62. The CPU 61 has a timer for timing and executes processing according to the time measured by the timer.

[0052] The main memory 62 consists of RAM (Random Access Memory), etc. The program 69 executed by the CPU 61 is loaded into the main memory 62 from the external memory 63. The main memory 62 is also used as the CPU 61's work area (temporary data storage area). The functions of the feature calculation unit 30 and the determination unit 40 shown in Figure 8 are realized by the execution of the program by the CPU 61.

[0053] The external memory 63 consists of non-volatile memory such as flash memory or a hard disk. The external memory 63 is pre-stored with a program 69 to be executed by the CPU 61.

[0054] The human-machine interface 64 includes devices such as a keyboard and mouse, and an interface device that connects these devices to the internal bus 68. The human-machine interface 64 also includes a display device such as an organic EL or liquid crystal monitor. A touch panel can be used as the human-machine interface 64. The human-machine interface 64 displays information such as the results of a respiratory disease diagnosis.

[0055] The communication interface 65 is a data communication interface. Data communication with external devices is possible via the communication interface 65. Data communication with the file management unit 3, user terminal 5, or cloud computer is performed via this communication interface 65.

[0056] The functions of the information processing unit 4 can be implemented in a computer system consisting of one or more computers, each containing one or more processors and one or more storage devices, including non-temporary storage media. Multiple computers communicate with each other via a connected communication network to implement the functions of the information processing unit 4. For example, some of the functions of the information processing unit 4 may be implemented in one computer, while other functions may be implemented in other computers. The functions of the information processing unit 4 may also be implemented on a cloud computer. The file management unit 3 and the user terminal 5 also have the hardware configuration shown in Figure 13. The file management unit 3 and the information processing unit 4 may be implemented on the same hardware.

[0057] Next, the operation of the respiratory disease diagnosis system 100 will be described. First, the learning process in the determination unit 40 of the information processing unit 4 will be described, and then the process of determining whether or not a patient has a respiratory disease after learning will be described.

[0058] [Learning Process] As shown in Figure 14, first, the respiratory disease diagnosis system 100 acquires, for example, waveform data of internal conduction sound (healthy data) for three days from healthy pigs P (1A, 1B, 1C, ...) using the ear tag sensor 1 and the file management unit 3 (step S1). Next, the respiratory disease diagnosis system 100 is trained using training time-series data (step S2). Based on this healthy data, the feature calculation unit 30 generates a spectrogram of the healthy data and, referring to the ear tag number, inputs the spectrogram into the corresponding autoencoders 50A, 50B, 50C, ... for the pigs P (1A, 1B, 1C, ...). The autoencoders 50A, 50B, 50C, ... are trained to reconstruct and output the input spectrogram.

[0059] Next, the respiratory disease determination system 100 determines thresholds for determining whether or not respiratory disease is present in the detection units 51A, 51B, 51C, ... (step S3). Based on this healthy data, the feature calculation unit 30 generates a spectrogram of the healthy data and, referring to the ear tag number, inputs the spectrogram into the corresponding autoencoders 50A, 50B, 50C, ... of the pigs P (1A, 1B, 1C, ...). The autoencoders 50A, 50B, 50C, ... reconstruct and output the input spectrogram.

[0060] Next, the respiratory disease diagnosis system 100 performs a test using trained autoencoders 50A, 50B, 50C, ... and detection units 51A, 51B, 51C, ... with set thresholds (step S4). Abnormal data acquired in advance is used. In this test, the feature calculation unit 30 calculates a spectrogram based on this healthy data and abnormal data of infected pigs, the autoencoders 50A, 50B, 50C, ... reconstruct the spectrogram, and the detection units 51A, 51B, 51C, ... calculate the reconstruction error x and make a judgment by comparing the reconstruction error x with the set threshold. Here, it is confirmed whether the healthy data is judged as healthy and the abnormal data is judged as abnormal in this judgment result. If abnormal data, which is waveform data of internal conducted sound in infected pigs, is not available, step S4 may be omitted.

[0061] [Decision Processing] The determination process after the learning process is completed will now be explained. First, as shown in Figure 15, the file management unit 3 waits until a certain interval T1 (see Figure 7) has elapsed (step S11). Once the certain interval T1 has elapsed (step S11; Yes), the file management unit 3 sends a detection command to the ear tag sensor 1 via the router 2 (step S12). After that, the file management unit 3 waits until it receives waveform data from the ear tag sensor 1 (step S13; No).

[0062] Upon receiving a detection command, the ear tag sensor 1 detects the internal conduction sound of the pig P for a predetermined time T2 (see Figure 7) and transmits waveform data, with data representing the ear tag number added, to the file management unit 3 via the router 2. When the file management unit 3 receives the waveform data (step S13; Yes), it stores the received data (step S14). Subsequently, the file management unit 3 determines whether or not one day has passed (step S15). If one day has not passed (step S15; No), it returns to step S11. In this way, steps S11 to S15 are repeatedly executed, and at regular intervals T1, waveform data of the internal conduction sound of the pig P at time T2 is accumulated in the file management unit 3.

[0063] After one day has passed (Step S15; Yes), the file management unit 3 generates a file of waveform data for each pig P for that day (Step S16) and sends that file to the information processing unit 4 (Step S17). In this way, according to the respiratory disease diagnosis system 100, waveform data of internal conduction sounds for one day is compiled into a file for each pig P and sent from the file management unit 3 to the information processing unit 4.

[0064] As shown in Figure 16, when the information processing unit 4 receives a file from the file management unit 3 (step S21), the feature calculation unit 30 of the information processing unit 4 calculates a spectrogram, which is a feature of the waveform data contained in the file (step S22). Subsequently, the determination unit 40 of the information processing unit 4 refers to the ear tag number of the pig P indicated by the file, inputs the feature of the waveform data into the individual determination unit of the pig P corresponding to the ear tag number, and performs a determination using this as time-series data for inspection (step S23). In this step, the spectrogram is reconstructed in the autoencoder of the pig P corresponding to the ear tag number, and one of the detection units detects an anomaly using the reconstructed spectrogram.

[0065] Next, the information processing unit 4 sends multiple judgment result emails for the pigs P to the user terminal 5 (step S24). When the user terminal 5 receives the judgment result email (step S31), it displays the judgment results contained in the email (step S32).

[0066] In this way, the respiratory disease detection system 100 determines whether or not an abnormality has occurred in the waveform data for each pig P, and automatically sends the determination result to the user terminal 5 via email. The respiratory disease detection system 100 may also be configured to send an email to the user terminal 5 only when an abnormality has occurred in pig P. Alternatively, the respiratory disease detection system 100 may be configured to automatically notify whether or not an abnormality has occurred when a health status request is received from the user terminal 5.

[0067] Three pigs P(A, B, C) were infected with Pasteurella bacteria, and the learning process shown in Figure 14 was performed to train the individual judgment units 40A, 40B, and 40C corresponding to each pig P(A, B, C), determine thresholds, and perform tests. For example, in this training, approximately 50% of the acquired healthy data was used, approximately 20% of the healthy data was used for threshold determination, and the remaining 30% of the healthy data and abnormal data, which are waveform data of internal conduction sounds from infected pigs, were used for testing. In addition, the reconstruction error x between the input spectrogram and the reconstructed spectrogram was calculated, and the mean μ and variance σ of the reconstruction error x were calculated. 2 The value is calculated, and μ+σ is set as the threshold. The following symptoms are present in pigs P (A, B, and C), respectively. Pig A: Cough / breath sounds stressed, respiratory distress (temporary after vaccination) Pig B: Cough and respiratory distress (temporary after vaccination) Pig C: Increased respiratory rate (temporary after vaccination) Figure 17(A) shows the results for pig A in the test, Figure 17(B) shows the results for pig B in the test, and Figure 17(C) shows the results for pig C in the test. Here, pre and pid in Figures 17(A) to 17(C) represent healthy data and diseased data, respectively. Pre and Pid in Figures 17(A) to 17(C) represent pre-disease and post-disease, respectively, and the number after pid represents the number of days elapsed since bacterial inoculation. In other words, the graphs shown in Figures 17(A) to 17(C) represent the identification rate of healthy pigs and pigs with respiratory disease after bacterial inoculation in pigs A to C. As shown in Figures 17(A) to 17(C), healthy pigs could be identified with an overall accuracy of over 80% in the three pigs A to C. Furthermore, pigs A and B could identify disease from pid5 to pid7 with an accuracy of 80-90% or higher, and pig C could identify disease from pid3 to pid7 with nearly 100% accuracy. Note that the proportion of data used for training, threshold determination, and testing is not limited to this. In addition, other thresholds such as μ+σ or the maximum value may be used.

[0068] In this way, the respiratory disease diagnosis system 100 constructs a spectrogram reconstruction model for each pig P and detects abnormalities, thereby eliminating the decrease in accuracy in determining the presence or absence of respiratory disease due to individual differences. According to the respiratory disease diagnosis system 100, an LSTM autoencoder using the spectrogram as input can identify healthy individuals with an overall accuracy of over 80%, and identify diseased pigs from pid5 to pid7 with an accuracy of over 80-90%.

[0069] In classifying healthy and diseased pigs P using reconstruction error x, there may be cases where there is almost no difference between the distributions of healthy and diseased pigs, which can lead to a decrease in classification accuracy. Therefore, KL divergence can be used to determine the threshold. KL divergence is an index that measures the difference between probability distributions, and when P(x) and Q(x) are discrete probability distributions, it is defined by the following equation.

number

[0070] For example, if Q(x) and P(x) are the probability distributions of the reconstruction error x of the waveform data of internal conduction sound from a healthy pig P (healthy data), then, as shown in Figure 18(A), there is almost no difference between the probability distributions P(x) and Q(x), so the KL divergence value becomes small. However, if Q(x) is the probability distribution of the reconstruction error x of the internal conduction sound data from a healthy pig P (healthy data), and P(x) is the probability distribution of the reconstruction error x of the internal conduction sound data from a diseased pig P (disease data), then, as shown in Figure 18(B), there is a difference in the distributions, so the KL divergence value becomes large. Therefore, the accuracy of the threshold can be improved by replacing the reconstruction error x with the KL divergence.

[0071] The KL divergence is calculated in step S3 of the learning process shown in Figure 14 by creating probability distributions P(x) and Q(x) of the calculated reconstruction error x. The disease of pigs P can be detected by setting this KL divergence as the threshold for detection units 51A, 51B, 51C, etc.

[0072] For the four pigs A through D shown below, autoencoders 50A, 50B, 50C, 50D, ... were constructed for each pig, and healthy and diseased individuals were distinguished by reconstruction error x or KL divergence. Pig A: Pneumonia (severe), respiratory symptoms (cough) Pig B: Pneumonia (moderate), respiratory symptoms (cough) Pig C: Pneumonia (moderate to mild), respiratory symptoms (cough) Pig D: Pneumonia (mild), respiratory symptoms (cough)

[0073] As shown in Figures 19(A) to 19(D), when comparing the disease identification rate for four pigs A to D with a threshold based on reconstruction error x and a threshold based on KL divergence, the disease identification rate was generally improved when the threshold based on KL divergence was used. Thus, by applying KL divergence to an autoencoder that takes a spectrogram as input, the disease identification rate is generally improved, and diseases can be identified with an accuracy of approximately 60-90% or more.

[0074] In this embodiment, the characteristic quantities of the internally conducted sound are used as its spectrogram. However, the characteristic quantities may also be a delta spectrogram or wavelet transform data. Alternatively, the waveform data of the internally conducted sound may be input directly to the determination unit 40 without the feature quantity calculation unit 30.

[0075] Embodiment 2 Embodiment 2 of the present invention will now be described. The respiratory disease diagnosis system 100 according to this embodiment differs from the respiratory disease diagnosis system 100 according to Embodiment 1 in that the autoencoders 50A, 50B, 50C, ... are PerceiverIO autoencoders as shown in Figure 20, rather than LSTM autoencoders, and the time-series data input to the autoencoders 50A, 50B, 50C, ... is not a spectrogram, but the waveform data of internally conducted sound itself.

[0076] In the respiratory disease diagnosis system 100, a PerceiverIO autoencoder is used as time-series data. The PerceiverIO autoencoder is an encoder-decoder model that can handle time-series data, similar to the LSTM autoencoder in Embodiment 1 described above. First, the PerceiverIO autoencoder extracts features from a latent array in which the input data is embedded using an encoder composed of multiple attentions and an MLP. Then, the decoder reconstructs the input data based on the extracted features. The configuration of the PerceiverIO autoencoder is well known, so a detailed explanation is omitted. The learning process and the diagnosis process performed by the respiratory disease diagnosis system 100 are the same as in Embodiment 1 described above.

[0077] Figure 21 shows an example of a comparison between the recognition rates when using an LSTM autoencoder and when using a PerceiverIO autoencoder. Note that the input data for the LSTM autoencoder is a spectrogram for 5 seconds, so the input data is not strictly the same. In Figure 21, Pre and Pid represent before and after infection, respectively, and the number after Pid represents the number of days elapsed since bacterial inoculation. As shown in Figure 21, the PerceiverIO autoencoder is generally able to correctly identify pre-infection cases, although there are individual differences. In other examples, the PerceiverIO autoencoder does not always outperform the LSTM autoencoder, but it is suggested that using the PerceiverIO autoencoder (autoencoder 50A, 50B, 50C, ...) may improve recognition accuracy compared to using the LSTM autoencoder.

[0078] In this embodiment, the waveform data of the internally conducted sound itself is input to the determination unit 40. However, similar to Embodiment 1 above, the feature calculation unit 30 may calculate a spectrogram, delta spectrogram, or wavelet transform data as features and input them to the determination unit 40.

[0079] [summary] As described in detail above, the respiratory disease determination system 100 according to the above embodiment includes a determination unit 40 that determines whether or not a pig P has a respiratory disease. The determination unit 40 includes an autoencoder 50A, 50B, 50C, ... which compresses learning time-series data representing the characteristics of internal conduction sound in a healthy pig P, learns to restore the learning time-series data, and then inputs and restores inspection time-series data representing the characteristics of internal conduction sound in the same pig P, and a detection unit 50A, 50B, 50C, ... which detects whether or not a pig has a respiratory disease based on the difference between the inspection time-series data input to the autoencoder 50A, 50B, 50C, ... and the inspection time-series data restored by the autoencoder 50A, 50B, 50C, ... for each pig P. Therefore, by using a learning model that reflects individual differences in pigs P, it is possible to determine whether or not a pig has respiratory disease. This allows for more accurate and earlier detection of respiratory disease in living organisms, regardless of individual differences in pigs P.

[0080] The autoencoders 50A, 50B, 50C, ... may be LSTM (Long Short-Term Memory) autoencoders or PerceiverIO autoencoders. Any of these autoencoders can individually determine whether or not pig P has respiratory disease through unsupervised learning.

[0081] The respiratory disease determination system 100 according to the above embodiment includes a feature calculation unit 30 that calculates two-dimensional waveform data representing the time variation of the frequency spectrum of internally conducted sound as data representing the characteristics of internally conducted sound in a living organism. Autoencoders 50A, 50B, 50C, ... input the two-dimensional waveform data calculated by the feature calculation unit 30 as time-series data. In this way, it is possible to determine whether or not a pig P has a respiratory disease based on the characteristics of internally conducted sound, thereby improving the identification rate.

[0082] The ear tag sensor 1 according to the above embodiment is a sensor attached to the ear Y of a pig P. The ear tag sensor 1 comprises a piezoelectric element 10 that is in close contact with the ear Y of the pig P, a retaining plate 11 positioned to sandwich the ear Y between the piezoelectric element 10 and the housing 13 that houses a circuit board 21 on which a signal circuit is mounted that amplifies the signal detected by the piezoelectric element 10 and wirelessly outputs the amplified signal, and a battery 22 that supplies power to the signal circuit, and holds the piezoelectric element 10 so that it is in close contact with the ear Y of the pig P, a through rod 12 that extends through the piezoelectric element 10, the ear Y and the retaining plate 11 and whose tip is locked to the housing 13, and a spring portion 14 that, when compressed, has both ends sandwiched between the base 12a of the through rod 12 and the retaining plate 11 and biases the retaining plate 11 toward the ear Y. In the housing 13, the circuit board 21 and the battery 22 are positioned at locations offset from a straight line along the longitudinal direction of the through rod 12. This prevents the elastic force of the spring portion 14 from being directly transmitted to the circuit board 21 and the battery 22, thereby preventing damage to them. While the circuit board 21 and battery 22 are positioned on opposite sides of the hole, this is not the only arrangement. Both the circuit board 21 and the battery 22 may be positioned on the same side of the hole.

[0083] In the ear tag sensor 1 according to the above embodiment, the spring portion 14 is made of a hollow cylindrical elastic member through which the through rod 12 passes. The spring portion 14 is curved spherically along the direction in which the through rod 12 extends and has a shape that can expand and contract in the direction in which the through rod 12 extends. The spring portion 14 is formed such that the rate of change of the elastic force when the distance between the piezoelectric element 10 and the retaining plate 11 is within the range that the ear thickness L of the pig P can take is smaller than the rate of change when it is in other ranges. This makes it possible to minimize damage to the ear portion Y of the pig P.

[0084] According to the ear tag sensor 1 of the above embodiment, the piezoelectric element 10 comprises: a plate-shaped first electrode layer 10a made of a conductive material having a first through-hole 20a penetrating in the thickness direction; a plate-shaped piezoelectric layer 10b made of an electrostrictive material laminated on the first electrode layer 10a having a second through-hole 20b penetrating in the thickness direction, such that the second through-hole 20b is coaxial with the first through-hole 20a; a plate-shaped second electrode layer 10c made of a conductive material laminated on the piezoelectric layer 10b having a third through-hole 20c penetrating in the thickness direction, such that the third through-hole 20c is coaxial with the first through-hole 20a; and a plate-shaped sealing layer 10d made of an insulating material laminated on the second electrode layer 10c having a fourth through-hole 20d penetrating in the thickness direction, such that the fourth through-hole 20d communicates with the first through-hole 20a. When viewed in the stacking direction of the first electrode layer 10a, piezoelectric layer 10b, second electrode layer 10c, and seal layer 10d, the seal layer 10d and the first electrode layer 10a enclose the piezoelectric layer 10b and the second electrode layer 10c, and the piezoelectric layer 10b and the second electrode layer 10c are sealed by the seal layer 10d and the first electrode layer 10a, thereby preventing a short circuit between the first electrode layer 10a and the second electrode layer 10c.

[0085] In the above embodiment, the animal used to determine whether or not it has a respiratory disease is pig P. However, it is not limited to this. Other animals, such as cattle which are raised in groups, may also be used as the subject of determination.

[0086] Furthermore, the hardware and software configurations of the file management unit 3, information processing unit 4, and user terminal 5 are examples only and can be changed and modified as needed.

[0087] The core processing units of the file management unit 3, information processing unit 4, and user terminal 5, which consist of a CPU 61, main memory 62, external memory 63, human-machine interface 64, communication interface 65, and internal bus 68, can be implemented using a normal computer system, as described above, without the need for a dedicated system. For example, the computer program for performing the above operations may be stored on a computer-readable recording medium (flexible disk, CD-ROM, DVD-ROM, etc.) and distributed, and the file management unit 3, information processing unit 4, and user terminal 5 that perform the above processing may be configured by installing the computer program on a computer. Alternatively, the computer program may be stored on a storage device of a server device on a communication network such as the Internet, and the file management unit 3, information processing unit 4, and user terminal 5 may be configured by downloading it from a normal computer system.

[0088] When computer functions are realized through a division of labor between the OS (operating system) and application programs, or through collaboration between the OS and application programs, only the application program portion may be stored on a recording medium or storage device.

[0089] It is also possible to superimpose a computer program onto a carrier wave and distribute it via a communication network. For example, a computer program could be posted on a bulletin board system (BBS) on a communication network and distributed via the network. This computer program could then be launched and executed under the control of the OS, similar to other application programs, thereby enabling the aforementioned processing.

[0090] This invention allows for various embodiments and modifications without departing from the broad spirit and scope of the invention. Furthermore, the embodiments described above are for illustrative purposes only and do not limit the scope of the invention. In other words, the scope of this invention is indicated not by the embodiments, but by the claims. Various modifications made within the scope of the claims and the equivalent scope of the meaning of the invention are considered to be within the scope of this invention. [Industrial applicability]

[0091] This invention is used to determine whether a living organism has a respiratory disease. [Explanation of Symbols]

[0092] 1 Ear tag sensor, 2 Router, 3 File management unit, 4 Information processing unit, 5 User terminal, 10 Piezoelectric element, 10a First electrode layer, 10b Piezoelectric layer, 10c Second electrode layer, 10d Seal layer, 10e Soldering part, 11 Retaining plate, 12 Through rod, 12a Base, 12b Rod-shaped part, 12c Return part, 13 Housing, 13a Container, 13b Lid part, 13c Axle hole part, 13d Recess, 14 Spring part, 20a First through hole, 20b Second through hole, 20c Third through hole, 20d Fourth through hole, 21 Circuit board, 22 Battery, 23 Tag pliers, 30 Feature quantity calculation unit, 40 Judgment unit, 40A, 40B, 40C Individual judgment unit, 50A, 50B, 50C Autoencoder, 51A, 51B, 51C Detection unit, 61 CPU, 62 Main memory, 63 External memory, 64 Human-machine interface (I / F), 65 Communication interface (I / F), 68 Internal bus, 69 Program, 100 Respiratory disease diagnosis system, P Pig, Y Ear

Claims

1. It is equipped with a determination unit that determines whether or not each of multiple living organisms has a respiratory disease. The determination unit, An autoencoder that first compresses time-series data for training that represents the characteristics of internal conduction sounds in a healthy pig P, then learns to reconstruct the said time-series data for training, and then inputs and reconstructs time-series data for testing that represents the characteristics of internal conduction sounds in the same organism. A detection unit detects whether or not a person has a respiratory disease based on the difference between the time-series data for testing input to the autoencoder and the time-series data for testing restored by the autoencoder. Each organism is equipped with its own unique features. Respiratory disease diagnosis system.

2. The autoencoder is an LSTM (Long Short-Term Memory) autoencoder. The respiratory disease diagnosis system according to claim 1.

3. The autoencoder is a PerceiverIO autoencoder. The respiratory disease diagnosis system according to claim 1.

4. The system includes a feature calculation unit that calculates two-dimensional waveform data representing the time variation of the frequency spectrum of the internally conducted sound or the components obtained by Fourier transforming the frequency spectrum, as data representing the characteristics of internally conducted sound in a living organism. The autoencoder inputs the two-dimensional waveform data calculated by the feature calculation unit as the time-series data. A respiratory disease diagnosis system according to any one of claims 1 to 3.

5. An ear tag sensor that is attached to the ear of a living organism, A piezoelectric element that adheres closely to the ear of a living organism, A retaining plate is positioned to sandwich the ear portion between the piezoelectric element and the piezoelectric element, A circuit board on which a signal circuit is mounted that amplifies the signal detected by the piezoelectric element and wirelessly outputs the amplified signal, and a battery that supplies power to the signal circuit are housed in a housing that also holds the piezoelectric element. A through rod extending through the piezoelectric element, the ear portion, and the retaining plate, with its tip engaging with the housing, A spring portion, in a compressed state, is sandwiched between the base end of the through rod and the retaining plate, and biases the retaining plate toward the lug portion, Equipped with, In the aforementioned housing, the circuit board and the battery are positioned at locations offset from a straight line along the longitudinal direction of the through rod. Ear tag sensor.

6. The aforementioned spring portion is The aforementioned through rod is composed of a hollow cylindrical elastic member that penetrates the interior. The through rod is curved spherically along the direction in which it extends, and has a shape that is expandable and contractible in the direction in which the through rod extends. The distance between the piezoelectric element and the retaining plate is formed such that the rate of change of elastic force when it is within a range that can be taken by the thickness of a living ear is smaller than the rate of change when it is in other ranges. The ear tag sensor according to claim 5.

7. The piezoelectric element is A plate-shaped first electrode layer made of a conductive material having a first through-hole penetrating in the thickness direction, A plate-shaped piezoelectric layer made of electrostrictive material is laminated on the first electrode layer, having a second through-hole that penetrates in the thickness direction, and the second through-hole is coaxial with the first through-hole. A plate-shaped second electrode layer made of a conductive material is laminated on the piezoelectric layer, having a third through-hole that penetrates in the thickness direction, and the third through-hole is coaxial with the first through-hole. A plate-shaped sealing layer made of an insulating material is laminated on the second electrode layer, having a fourth through-hole that penetrates in the thickness direction of the plate, and the fourth through-hole communicates with the first through-hole. Viewed in the stacking direction of the first electrode layer, the piezoelectric layer, the second electrode layer, and the seal layer, the seal layer and the first electrode layer enclose the piezoelectric layer and the second electrode layer, The sealing layer and the first electrode layer seal the piezoelectric layer and the second electrode layer. The ear tag sensor according to claim 5.

Images