Learning system, estimation system, learning method, and estimation method

Abstract

To provide a leaning system that can improve estimation accuracy of a respiratory state of a person.SOLUTION: A learning system 1 comprises a signal acquisition unit 10, a state acquisition unit 20, and a model generation unit 30. The signal acquisition unit 10 acquires a plurality of sensor signals based on radio waves reflected from a person, from a plurality of first radio wave sensors for transmitting radio waves to the person from directions different from each other. The state acquisition unit 20 acquires a respiratory state of the person corresponding to the plurality of sensor signals. The model generation unit 30 generates a leaned model 31 on the basis of the plurality of sensor signals and the respiratory state of the person. The leaned model 31 estimates the respiratory state of the person on the basis of a sensor signal inputted from one or more second radio wave sensors. The number of the one or more second radio wave sensors is equal to or smaller than the number of the plurality of first radio wave sensors.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a learning system, an estimation system, a learning method, and an estimation method, and more particularly to a learning system, an estimation system, a learning method, and an estimation method for estimating a respiratory state based on a sensor signal. [Background technology]

[0002] Patent Document 1 discloses a sleep state estimation system that estimates a sleep state using a sleep state estimation model that receives as inputs phase coherence based on the instantaneous phase difference between the instantaneous phase of fluctuations in heartbeat intervals acquired while a person is sleeping and the instantaneous phase of a breathing pattern, and body movement information, heart rate information, or breathing information. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2022-63926 Summary of the Invention [Problem to be solved by the invention]

[0004] However, depending on the posture of a person, a sensor signal obtained by transmitting and receiving radio waves to and from the person may not contain sufficient information indicating the person's respiratory state. Therefore, when generating a learning model for an estimation system that estimates a person's respiratory state using the sensor signal as input, the accuracy of the learning model's estimation of the person's respiratory state may not be sufficiently high.

[0005] An object of the present disclosure is to provide a learning system, an estimation system, a learning method, and an estimation method that can improve the accuracy of estimating a person's respiratory state. [Means for solving the problem]

[0006] A learning system according to one aspect of the present disclosure includes a signal acquisition unit, a state acquisition unit, and a model generation unit. The signal acquisition unit acquires a plurality of sensor signals based on radio waves reflected from a person from a plurality of first radio wave sensors that transmit radio waves toward the person from different directions. The state acquisition unit acquires the person's respiratory state corresponding to the plurality of sensor signals. The model generation unit generates a trained model based on the plurality of sensor signals and the person's respiratory state. The trained model estimates the person's respiratory state based on one or more sensor signals input from one or more second radio wave sensors. The number of the one or more second radio wave sensors is equal to or less than the number of the plurality of first radio wave sensors.

[0007] An estimation system according to one aspect of the present disclosure includes an estimation unit and a second signal acquisition unit different from the first signal acquisition unit, which is the signal acquisition unit. The estimation unit estimates a person's respiratory state using the trained model generated by the learning system. The second signal acquisition unit acquires, from one or more second radio wave sensors that transmit radio waves to the person, one or more sensor signals based on radio waves reflected from the person. The estimation unit inputs the one or more sensor signals acquired by the second signal acquisition unit into the trained model.

[0008] A learning method according to one aspect of the present disclosure is executed by one or more processors. The learning method includes a signal acquisition step, a state acquisition step, and a model generation step. In the signal acquisition step, a plurality of sensor signals based on radio waves reflected from a person are acquired from a plurality of first radio wave sensors that transmit radio waves toward the person from different directions. In the state acquisition step, the respiratory state of the person corresponding to the plurality of sensor signals is acquired. In the model generation step, a trained model is generated based on the plurality of sensor signals and the respiratory state of the person. The trained model estimates the respiratory state of the person based on sensor signals input from one or more second radio wave sensors. The number of the one or more second radio wave sensors is equal to or less than the number of the plurality of first radio wave sensors.

[0009] An estimation method according to one aspect of the present disclosure is executed by one or more processors. The estimation method includes an estimation step and a second signal acquisition step different from the first signal acquisition step, which is the signal acquisition step. In the estimation step, a person's respiratory state is estimated using the trained model generated by the learning method. In the second signal acquisition step, one or more sensor signals based on radio waves reflected from the person are acquired from one or more second radio wave sensors that transmit radio waves to the person. In the estimation step, the one or more sensor signals acquired in the second signal acquisition step are input to the trained model. [Effects of the Invention]

[0010] According to a learning system, an estimation system, a learning method, and an estimation method according to an aspect of the present disclosure, it is possible to improve the accuracy of estimating a person's respiratory state. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a block diagram of a learning system according to the first embodiment. [Figure 2] FIG. 2 is a schematic diagram showing a configuration related to acquisition of a sensor signal in the learning system. [Figure 3] Fig. 3A is a graph showing an example of the waveform of a sensor signal, and Fig. 3B is a graph showing another example of the waveform of a sensor signal. [Figure 4] FIG. 4 is a block diagram of the estimation system according to the first embodiment. [Figure 5] FIG. 5 is a flowchart showing the operation of the learning system according to the first embodiment. [Figure 6] FIG. 6 is a flowchart showing the operation of the estimation system according to the first embodiment. [Figure 7] FIG. 7 is a block diagram showing the configuration of a learning system according to the second embodiment. [Figure 8] FIG. 8 is a block diagram showing the configuration of a learning system according to the third embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0012] Hereinafter, learning systems and estimation systems according to embodiments 1 to 3 will be described with reference to the drawings. However, each diagram described in the following embodiments is a schematic diagram, and the ratios of the sizes and thicknesses of the components do not necessarily reflect the actual dimensional ratios. Note that the configurations described in the following embodiments are merely examples of the present disclosure. The present disclosure is not limited to the following embodiments, and various modifications are possible depending on the design, etc., as long as the effects of the present disclosure can be achieved.

[0013] (Embodiment 1) (1) Learning System The learning system 1 according to the first embodiment includes a signal acquiring unit 10, a state acquiring unit 20, and a model generating unit 30, as shown in FIG.

[0014] (1.1) Signal acquisition section The signal acquiring unit 10 acquires, from each of a plurality of (two in FIG. 2 ) radio wave sensors 100, a sensor signal 15 indicating biometric information of the person 3. More specifically, the signal acquiring unit 10 acquires, from the plurality of radio wave sensors 100, a plurality of sensor signals 15 indicating the breathing state of the person 3 while sleeping.

[0015] The multiple radio wave sensors 100 transmit radio waves 41 to the person 3, receive reflected waves 42, and generate sensor signals 15. The sensor signals 15 exhibit phase changes due to the Doppler effect caused by the movement of the reflection sources of the reflected waves 42. The reflection sources of the reflected waves 42 include the person 3. The movement of the reflection sources includes the movement due to the breathing of the person 3 and the movement due to the body movement of the person 3.

[0016] Each of the multiple radio wave sensors 100 includes a transceiver 11, a transmitting antenna 12, and a receiving antenna 13. The transmitting antenna 12 and the receiving antenna 13 are installed facing a place 5 where a person 3 sleeps. The place 5 where the person 3 sleeps is, for example, bedding such as a bed. The transceiver 11 transmits radio waves 41 of a predetermined frequency from the transmitting antenna 12 to the place 5. The predetermined frequency is, for example, 24 GHz. The transceiver 11 performs quadrature detection on a received signal based on a reflected wave 42 received by the receiving antenna 13, and generates a sensor signal 15. The sensor signal 15 is, for example, a combination of an in-phase component (I component) and a quadrature component (Q component).

[0017] Here, the multiple radio wave sensors 100 are arranged so that the positions of the transmitting antenna 12 and receiving antenna 13 as seen from the person 3 are different for each radio wave sensor 100. More specifically, the multiple radio wave sensors 100 include radio wave sensor 100a and radio wave sensor 100b. The traveling direction of the radio waves 41 incident on the person 3 from the transmitting antenna 12 of radio wave sensor 100a and the traveling direction of the radio waves 41 incident on the person 3 from the transmitting antenna 12 of radio wave sensor 100b intersect at the person 3. For example, the transmitting antenna 12 and receiving antenna 13 of radio wave sensor 100a are installed in the upper right corner of location 5, and the transmitting antenna 12 and receiving antenna 13 of radio wave sensor 100b are installed in the upper left corner of location 5.

[0018] As described above, the sensor signal 15 exhibits a phase change due to the Doppler effect caused by the movement of the reflection source of the radio waves 41. Therefore, among the components included in the sensor signal 15 that are caused by the movement of the reflection source, the degree of phase change is small for components where the angle between the traveling direction of the radio waves 41 and the reflected waves 42 and the direction of the movement of the reflection source is large. Therefore, the component strength of the component included in the sensor signal 15 that is caused by the breathing of the person 3 changes depending on the relationship between the traveling direction of the radio waves 41 and the reflected waves 42 and the posture of the person 3. The posture of the person 3 is, for example, the posture of the person 3 when sleeping relative to the bedding, and includes, for example, lying on their back, lying on their face, facing right, and facing left.

[0019] More specifically, the breathing movement of the person 3 generally has a large component in the front direction as seen from the person 3 and a small component in the left-right direction as seen from the person 3. Therefore, as will be described later, when the person 3 is in a position with their shoulders turned toward the transmitting antenna 12 and the receiving antenna 13 of the radio wave sensor 100, the angle between the traveling direction of the radio waves 41 and the reflected waves 42 and the front direction of the person 3 becomes large, and therefore the component in the sensor signal 15 caused by the breathing movement of the person 3 tends to become weak.

[0020] Fig. 3A shows a sensor signal 15a acquired from a radio wave sensor 100a. The sensor signal 15a includes an I component 15aI and a Q component 15aQ. Fig. 3B shows a sensor signal 15b acquired from a radio wave sensor 100b. The sensor signal 15b includes an I component 15bI and a Q component 15bQ.

[0021] 3A and 3B, each of the multiple periods T14 corresponds to a period in which the person 3 turns over in their sleep. That is, the sensor signal 15 in each of the multiple periods T14 is a signal acquired when the person 3 turns over in their sleep. Also, in FIGS. 3A and 3B, each of the three periods T11 is a period in which the person 3 is lying on their back. That is, the sensor signal 15 in each of the multiple periods T11 is a signal acquired when the person 3 is lying on their back. Also, in FIGS. 3A and 3B, each of the two periods T12 is a period in which the person 3 is lying on their right. That is, the sensor signal 15 in each of the multiple periods T12 is a signal acquired when the person 3 is lying on their right. Also, in FIGS. 3A and 3B, each of the two periods T13 is a period in which the person 3 is lying on their left. That is, the sensor signal 15 in each of the multiple periods T13 is a signal acquired when the person 3 is lying on their left.

[0022] 3A and 3B, in a period T12 when the person 3 is facing rightward, the amplitude of the sensor signal 15b is smaller than the amplitude of the sensor signal 15a. On the other hand, in a period T13 when the person 3 is facing leftward, as shown in Figures 3A and 3B, the amplitude of the sensor signal 15a acquired from the radio wave sensor 100a is smaller than the amplitude of the sensor signal 15b. The reason for this is that when the person 3 is facing their shoulder toward the transmitting antenna 12 and the receiving antenna 13 of the radio wave sensor 100, the left-right movement of the person 3 is smaller than the forward-backward movement of the person 3, and therefore the amplitude of the sensor signal 15b is smaller.

[0023] However, the direction of travel of radio waves 41 incident on person 3 from transmitting antenna 12 of radio wave sensor 100a and the direction of travel of radio waves 41 incident on person 3 from transmitting antenna 12 of radio wave sensor 100b intersect at person 3. Therefore, it is unlikely that the amplitudes of both sensor signals 15a and 15b will be smaller than the amplitude of sensor signal 15 when transmitting antenna 12 of radio wave sensor 100 is located in front of person 3. In other words, by arranging multiple radio wave sensors 100 as described above, it is highly likely that the component attributable to breathing movement of person 3, contained in at least one of the multiple sensor signals 15, contains information necessary to determine whether or not breathing is abnormal.

[0024] The signal acquiring unit 10 acquires a set of multiple sensor signals 15 generated by multiple radio wave sensors 100 from, for example, a storage medium.

[0025] (1.2) Status acquisition unit The state acquisition unit 20 acquires the respiratory state of the person 3 at the time when the plurality of sensor signals 15 are acquired.

[0026] Here, the respiratory state of the person 3 is information indicating whether or not the breathing of the person 3 is abnormal breathing. More specifically, the respiratory state of the person 3 is information indicating whether or not the breathing of the person 3 while sleeping is abnormal breathing. Abnormal breathing here includes apnea and hypopnea in sleep apnea syndrome. Furthermore, normal breathing here refers to a state that does not correspond to the findings of abnormal breathing in sleep apnea syndrome. More specifically, the respiratory state of the person 3 is information indicating whether the breathing of the person 3 is normal breathing, apnea, or hypopnea at the time when the multiple sensor signals 15 are acquired. The respiratory state of the person 3 includes, for example, the start time and end time of each of the periods of normal breathing, apnea, and hypopnea.

[0027] Apnea refers to a state in which a person 3 stops breathing. Apnea includes central apnea and obstructive apnea. Central apnea refers to a state in which a person 3 is not breathing and the thorax and other parts of the body are not moving. Obstructive apnea refers to a state in which a blockage exists in the person 3's airway, and although breathing occurs, the ventilation volume is zero. Hypopnea refers to a state in which the person 3's ventilation volume is 50% or less compared to normal breathing. Like apnea, hypopnea also includes central hypopnea, in which the person 3's breathing is insufficient, and obstructive hypopnea, in which the ventilation volume is reduced due to a narrowed area in the person 3's airway.

[0028] The state acquisition unit 20 acquires, for example, the respiratory state of the person 3 prepared in advance by a medical professional such as a doctor. The respiratory state of the person 3 is, for example, a time chart showing the respiratory state of the person 3 at the time when the plurality of sensor signals 15 is acquired.

[0029] Specifically, a medical professional puts a person 3, who is a patient with sleep apnea syndrome, to sleep at a location 5 and performs a test to determine the breathing state of the person 3. This test is, for example, a polysomnography (PSG) test. When acquiring data for the learning system, a plurality of sensor signals 15 are generated by transmitting a plurality of radio waves 41 to the location 5 at the same time as the test is performed, and the signals are acquired by a signal acquisition unit 10. The medical professional analyzes the results of the PSG test to identify periods in which abnormal breathing occurred. Information on these periods of abnormal breathing, together with other periods of normal breathing, forms a time chart showing the breathing state of the person 3. A state acquisition unit 20 acquires this time chart as the breathing state of the person 3.

[0030] (1.3) Model Generation The model generation unit 30 generates a trained model 31 that estimates the respiratory state of the person 3 based on the plurality of sensor signals 15 and the respiratory state of the person 3. As described above, the plurality of sensor signals 15 are acquired from the plurality of radio wave sensors 100 that transmit radio waves 41 to the person 3 from mutually different directions. The trained model 31 is used in the estimation system 2 (see FIG. 4) described below, and estimates the person's respiratory state based on one or more (one in FIG. 4) sensor signals 15. As described above, the plurality of sensor signals 15 are based on radio waves (reflected waves) 42 reflected from the sleeping person 3. Furthermore, the respiratory state of the person 3 is the respiratory state of the person 3 while sleeping.

[0031] Here, the number of radio wave sensors 100 (corresponding to the second radio wave sensors in the present disclosure) that generate the sensor signals 15 used as input by the trained model 31 is less than or equal to the number of radio wave sensors 100 (corresponding to the first radio wave sensors in the present disclosure) that generate the multiple sensor signals 15 acquired by the signal acquisition unit 10.

[0032] The model generation unit 30 creates training data using the multiple sensor signals 15 and the respiratory state of the person 3. The training data is a pair of one or more (one in the first embodiment) input sensor signals 15 and the correct respiratory state of the person 3, and the trained model 31 is generated by machine learning so as to output an estimated value as close as possible to the correct answer of the training data when the training data is input. More specifically, the model generation unit 30 receives the sensor signal 15a acquired from the radio wave sensor 100a as input and creates training data in which the person's respiratory state is the correct answer. The model generation unit 30 also receives the sensor signal 15b acquired by the radio wave sensor 100b as input and creates training data in which the person's respiratory state is the correct answer. That is, the model generation unit 30 receives each of the sensor signals 15 acquired from each of the multiple radio wave sensors 100 as input and creates training data in which the person's respiratory state is the correct answer. In this way, the model generation unit 30 receives one sensor signal 15 as input and generates a trained model 31 that estimates the person's respiratory state.

[0033] The model generation unit 30 performs machine learning using the training data to generate a trained model 31. Any machine learning method such as a support vector machine or a neural network can be used as the machine learning method.

[0034] The model generation unit 30 may verify the trained model 31 by using a portion of the training data, which is a pair of the input sensor signal 15 and the correct respiratory state of the person 3, as training data for the model, and by using the remaining training data as verification data. That is, the model generation unit 30 prevents the training data from including a pair of the input and correct answer of the verification data. This makes it possible to prevent the prediction accuracy of the trained model 31 from being overly highly evaluated when verifying the prediction accuracy of the trained model 31. This is because when the verification data includes training data, the correct answer for the verification data is known information for the trained model 31, and therefore the accuracy rate for the verification data is high regardless of the prediction accuracy of the trained model 31.

[0035] The model generation unit 30 is, for example, a combination of a processor, a memory, and a program. The processor is, for example, a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit).

[0036] (2) Estimation system As shown in FIG. 4, the estimation system 2 according to the first embodiment includes a signal acquisition unit 40 and an estimation unit 50.

[0037] (2.1) Signal acquisition unit The signal acquisition unit 40 acquires sensor signals 15 indicating biometric information of the person 3 from one or more (one in FIG. 4 ) radio wave sensors 100. Here, the number of sensor signals 15 acquired by the signal acquisition unit 40 is equal to or less than the number of sensor signals 15 acquired by the signal acquisition unit 10 of the learning system 1. The number of sensor signals 15 acquired by the signal acquisition unit 40 is the same as the number of sensor signals 15 used as input by the trained model 31.

[0038] Furthermore, the transmitting antenna 12 and the receiving antenna 13 of one radio wave sensor 100 are installed, for example, directly above a person in a bedroom.

[0039] (2.2) Estimation part The estimation unit 50 estimates the respiratory state of the person 3 from the sensor signal 15, using the trained model 31 generated by machine learning in the model generation unit 30. As described above, the respiratory state of the person 3 is a time chart that corresponds to the time chart of the sensor signal and indicates whether the breathing of the person 3 is abnormal or not.

[0040] The estimation unit 50 inputs the sensor signal 15 acquired by the signal acquisition unit 40 into the trained model 31, and calculates the respiratory state of the person 3, which is the output, to estimate the respiratory state of the person 3.

[0041] The estimation unit 50 is, for example, a combination of a processor, a memory, and a program. The processor is, for example, a central processing unit (CPU) or a graphics processing unit (GPU).

[0042] (3) Operation (3.1) Operation of the learning system As shown in FIG. 5, the signal acquisition unit 10 of the learning system 1 acquires multiple sensor signals 15 (step S11). More specifically, the multiple radio wave sensors 100 transmit radio waves 41, receive reflected waves 42, and generate sensor signals 15. For example, the transceiver 11 of each of the multiple radio wave sensors 100 transmits radio waves 41 from the transmitting antenna 12 to a person 3 wearing a PSG sensor on their body. Furthermore, the transceiver 11 of each of the multiple radio wave sensors 100 performs quadrature detection on a received signal based on the radio waves 42 received by the receiving antenna 13, and generates a sensor signal 15. The signal acquisition unit 10 of the learning system 1 acquires the multiple sensor signals 15 generated by the multiple radio wave sensors 100.

[0043] Next, the state acquisition unit 20 of the learning system 1 acquires the respiratory state of the person 3 corresponding to the plurality of sensor signals 15 (step S12). More specifically, the state acquisition unit 20 acquires the respiratory state of the person 3 generated based on the result of the PSG.

[0044] Next, the model generation unit 30 of the learning system 1 creates training data (step S13). More specifically, in step S13, the model generation unit 30 creates training data for each of the multiple sensor signals 15, with the person's respiratory state being the correct answer. Specifically, the model generation unit 30 receives the sensor signal 15a acquired from the radio wave sensor 100a as an input and creates training data with the respiratory state of person 3 being the correct answer. The model generation unit 30 also receives the sensor signal 15b acquired from the radio wave sensor 100b as an input and creates training data with the respiratory state of person 3 being the correct answer.

[0045] Next, the model generation unit 30 of the learning system 1 performs machine learning based on the training data to generate a trained model 31 (step S14). More specifically, the model generation unit 30 performs machine learning based on the training and verification training data generated in step S13. As a result, if the sensor signal 15 input to the estimation system 2 is similar to either the sensor signal 15a or the sensor signal 15b included in the training data, the trained model 31 can accurately estimate the respiratory state of the person 3.

[0046] (3.2) Operation of the estimation system As shown in FIG. 6, the signal acquisition unit 40 of the estimation system 2 acquires the sensor signal 15 indicating the biological information of the person 3 from the radio wave sensor 100 (step S21).

[0047] Next, the estimation unit 50 of the estimation system 2 inputs the sensor signal 15 to the trained model 31 and estimates the respiratory state of the person 3 as an output (step S22).

[0048] (4) Effects The learning system 1 according to the first embodiment includes a signal acquiring unit 10, a state acquiring unit 20, and a model generating unit 30. The signal acquiring unit 10 acquires a plurality of sensor signals 15 based on radio waves 42 reflected from the person 3 from a plurality of radio wave sensors 100 that transmit radio waves 41 toward the person 3 from different directions. The state acquiring unit 20 acquires the respiratory state of the person 3 corresponding to the plurality of sensor signals 15. The model generating unit 30 generates a trained model 31 based on the plurality of sensor signals 15 and the respiratory state of the person 3. The trained model 31 estimates the respiratory state of the person 3 based on the sensor signals 15 input from the radio wave sensors 100. The number of the radio wave sensors 100 is equal to or less than the number of the plurality of radio wave sensors 100. As a result, the trained model 31 is generated based on training data that receives as input signals based on the plurality of sensor signals 15. This improves the accuracy of the estimation result of the respiratory state of the person 3 by the trained model 31.

[0049] Furthermore, in the learning system 1 according to the first embodiment, the signal acquiring unit 10 acquires a plurality of sensor signals 15 based on radio waves 42 reflected from the sleeping person 3. The state acquiring unit 20 acquires the breathing state of the person 3 while sleeping as the breathing state of the person 3. This enables the trained model 31 to estimate the breathing state of the person while sleeping, allowing the person to notice abnormalities that are not subjectively noticeable.

[0050] Furthermore, in the learning system 1 according to the first embodiment, the model generation unit 30 uses training data in which the respiratory state of person 3 is the correct answer for each of the plurality of sensor signals 15 to generate a trained model 31 that receives the sensor signals 15 as input and outputs an estimated value of the respiratory state of person 3. As a result, the trained model 31 is generated based on the sensor signal 15 that more strongly indicates the movement of the body surface of person 3 among the plurality of sensor signals 15. This improves the accuracy of the estimation result of the respiratory state of person 3 by the trained model 31 that receives the sensor signals 15 as input.

[0051] The estimation system 2 according to the first embodiment also includes an estimation unit 50 and a signal acquisition unit 40. The estimation unit 50 estimates the respiratory state of the person using the trained model 31 generated by the learning system 1. The signal acquisition unit 40 acquires one or more sensor signals 15 based on radio waves 42 reflected from the person 3 from one or more radio wave sensors 100 that transmit radio waves 41 to the person 3. The estimation unit 50 inputs the one or more sensor signals 15 acquired by the signal acquisition unit 40 to the trained model 31. As a result, in the estimation system 2, because the trained model 31 has high estimation accuracy, it is possible to estimate the respiratory state of the person 3 with high accuracy even if the number of the one or more sensor signals 15 is small.

[0052] The learning method according to the first embodiment is executed by one or more processors. The learning method includes a signal acquisition step, a state acquisition step, and a learning step. In the signal acquisition step, a plurality of sensor signals 15 based on radio waves 42 reflected from the person 3 are acquired from a plurality of radio wave sensors 100 that transmit radio waves 41 toward the person 3 from different directions. In the state acquisition step, the respiratory state of the person 3 corresponding to the plurality of sensor signals 15 is acquired. In the learning step, a trained model is generated based on the plurality of sensor signals 15 and the respiratory state of the person 3. The trained model estimates the respiratory state of the person 3 based on the sensor signals 15 input from one or more radio wave sensors 100. The number of the one or more radio wave sensors 100 is smaller than the number of the plurality of radio wave sensors 100. As a result, the trained model 31 is generated based on training data that inputs signals based on the plurality of sensor signals 15. This improves the accuracy of the estimation result of the respiratory state of the person 3 by the trained model 31.

[0053] The estimation method according to the first embodiment is executed by one or more processors. The estimation method includes an estimation step and a signal acquisition step. In the estimation step, the respiratory state of the person 3 is estimated using a trained model 31 generated by the learning method according to the first embodiment. In the signal acquisition step, one or more sensor signals 15 are acquired from one or more radio wave sensors 100 that transmit and receive radio waves to the person 3. In the estimation step, the one or more sensor signals 15 acquired in the signal acquisition step are input to the trained model 31. This enables the estimation system 2 to estimate the respiratory state of the person 3 with high accuracy based on the high accuracy of the trained model 31.

[0054] (Embodiment 2) (1) Composition As shown in FIG. 7, the learning system 1a according to the second embodiment further includes a posture estimation unit 60 and a model generation unit 30a instead of the model generation unit 30 in addition to the configuration of the learning system 1 according to the first embodiment.

[0055] The posture estimation unit 60 estimates the posture of the person 3 based on the plurality of sensor signals 15. Specifically, the posture of the person 3 is the posture of the person 3 while sleeping. Specifically, the posture estimation unit 60 detects a change in the posture of the person 3 based on a change in the amplitude of the plurality of sensor signals 15. The posture estimation unit 60 also estimates the posture of the person 3 by comparing the amplitudes of the plurality of sensor signals 15. The posture estimation unit 60 calculates a state reliability of the respiratory state of the person 3 at each time for each of the plurality of sensor signals 15.

[0056] Specifically, the posture estimation unit 60 estimates the posture of the person 3 as follows.

[0057] For example, when person 3 changes his / her posture, the person 3 turns over before and after the change, which is a body movement of person 3, and therefore the amplitude of sensor signal 15 temporarily increases as in period T14 shown in Figures 3A and 3B. Therefore, posture estimation unit 60 detects the amplitude of sensor signal 15 and estimates the period during which the amplitude of sensor signal 15 increases as the person 3 turning over. The period during which the amplitude of sensor signal 15 increases is, for example, a period during which the amplitude of sensor signal 15 exceeds a threshold continuously for a predetermined period of time or more.

[0058] When the person 3 changes his / her posture, the person 3 turns over (body movement), and therefore the posture of the person 3 remains constant during the period between two periods T14. Therefore, the posture estimation unit 60 estimates the posture of the person 3 based on the amplitude of the sensor signal 15 for each period between the multiple periods T14 (each of the periods T11 to T13 in FIGS. 3A and 13B).

[0059] 3A and 3B, the difference in amplitude between the two sensor signals 15 is small. Therefore, the posture estimation unit 60 can estimate that there is a small difference between the angle between the radio wave 41 from the radio wave sensor 100a and the front of the person 3 and the angle between the radio wave 41 from the radio wave sensor 100b and the front of the person 3. Therefore, if the radio wave sensor 100a is located to the upper right of the person 3 and the radio wave sensor 100b is located to the upper left of the person, the posture estimation unit 60 estimates that the posture of the person 3 is facing up or down.

[0060] 3A and 3B, the amplitude of sensor signal 15a is greater than the amplitude of sensor signal 15b. Therefore, the posture estimation unit 60 can estimate that the angle between radio waves 41 from radio wave sensor 100a and the front of person 3 is smaller than the angle between radio waves 41 from radio wave sensor 100b and the front of person 3. In other words, the posture estimation unit 60 can estimate that, as viewed from person 3, radio wave sensor 100a is located in a direction close to the front or back, and radio wave sensor 100b is located in a direction close to the right or left. Therefore, if radio wave sensor 100a is located to the upper right of person 3 and radio wave sensor 100b is located to the upper left of person 3, the posture estimation unit 60 estimates that the posture of person 3 is facing right.

[0061] 3A and 3B, the amplitude of sensor signal 15b is greater than the amplitude of sensor signal 15a. Therefore, the posture estimation unit 60 can estimate that the angle between radio waves 41 from radio wave sensor 100a and the front of person 3 is greater than the angle between radio waves 41 from radio wave sensor 100b and the front of person 3. In other words, the posture estimation unit 60 can estimate that, as viewed from person 3, radio wave sensor 100a is located closer to the right or left, and radio wave sensor 100b is located closer to the front or back. Therefore, if radio wave sensor 100a is located to the upper right of person 3 and radio wave sensor 100b is located to the upper left of person 3, the posture estimation unit 60 estimates that the posture of person 3 is facing leftward.

[0062] The posture estimation unit 60 calculates the state reliability of the respiratory state of the person 3 based on the posture of the person 3 and changes in the posture of the person 3. More specifically, the posture estimation unit 60 first estimates the posture of the person 3 based on changes in the posture of the person 3. Then, the posture estimation unit 60 calculates the state reliability of the respiratory state of the person 3 for each of the plurality of sensor signals 15 based on the estimated posture of the person 3. Here, the state reliability of the respiratory state of the person 3 is an index indicating to what extent the sensor signal 15 indicates the respiratory state of the person 3. Specifically, if the radio wave sensor 100 that generated the sensor signal 15 is located in a direction that hits the front or back of the person 3, the state reliability of the respiratory state of the person 3 is estimated to be high. Furthermore, if the radio wave sensor 100 that generated the sensor signal 15 is located in a direction that hits the right or left of the person 3, the state reliability of the respiratory state of the person 3 is estimated to be low.

[0063] Specifically, the posture estimation unit 60 estimates that the state reliability of the respiratory state of the person 3 in each of the multiple periods T11 and T12 for the sensor signal 15a shown in Fig. 3A is high. On the other hand, the posture estimation unit 60 estimates that the state reliability of the respiratory state of the person 3 in each of the multiple periods T13 for the sensor signal 15a shown in Fig. 3A is low. Similarly, the posture estimation unit 60 estimates that the state reliability of the respiratory state of the person 3 in each of the multiple periods T11 and T13 for the sensor signal 15b shown in Fig. 3B is high. On the other hand, the posture estimation unit 60 estimates that the state reliability of the respiratory state of the person 3 in each of the multiple periods T12 for the sensor signal 15b shown in Fig. 3B is low.

[0064] The model generation unit 30a creates teacher data based on the respiratory state of the person 3 output by the state acquisition unit 20 and the state reliability output by the posture estimation unit 60. Similar to the model generation unit 30 according to the first embodiment, the model generation unit 30a uses each of the multiple sensor signals 15 as an input in the teacher data. Furthermore, when the state reliability of the respiratory state of the sensor signals 15 is high, the model generation unit 30a uses the respiratory state of the person 3 as a correct answer for the input in the teacher data. On the other hand, when the state reliability of the respiratory state of the sensor signals 15 is low, the model generation unit 30a uses the state of "unknown" as a correct answer for the input in the teacher data.

[0065] Specifically, in the teacher data, for an input that is the sensor signal 15a in the period T11, the model generation unit 30a determines the respiratory state of the person 3 in the period T11 as the correct answer. Furthermore, in the teacher data, for an input that is the sensor signal 15b in the period T11, the model generation unit 30a determines the respiratory state of the person 3 in the period T11 as the correct answer. Furthermore, in the teacher data, for an input that is the sensor signal 15a in the period T12, the model generation unit 30a determines the respiratory state of the person 3 in the period T12 as the correct answer. Furthermore, in the teacher data, for an input that is the sensor signal 15b in the period T13, the model generation unit 30a determines the respiratory state of the person 3 in the period T13 as the correct answer.

[0066] On the other hand, the model generation unit 30a determines that the state of "unknown" is the correct answer for the input that is the sensor signal 15b in the period T12 in the teacher data. Also, the model generation unit 30a determines that the state of "unknown" is the correct answer for the input that is the sensor signal 15a in the period T13 in the teacher data.

[0067] As a result, the trained model 31 generated by the model generation unit 30a does not learn, among the multiple sensor signals 15, portions of the respiratory state of the person 3 that have low state reliability as information for determining whether the respiratory state corresponds to abnormal breathing or normal breathing, and therefore, in estimation using the trained model 31, the estimation accuracy of the respiratory state of the person 3 is improved.

[0068] (2) Effects The learning system 1a according to the second embodiment differs from the learning system 1 according to the first embodiment in that it further includes a posture estimation unit 60 that estimates the posture of the person 3 based on a plurality of sensor signals 15. The posture estimation unit 60 calculates a state reliability of the respiratory state of the person 3 for each of the plurality of sensor signals 15 based on the posture of the person 3. The model generation unit 30a creates training data based on the plurality of sensor signals 15 and the state reliability, and generates a trained model 31 that receives the sensor signals 15 as input and outputs the respiratory state of the person 3. As a result, the trained model 31 is generated based on the sensor signals 15 that are estimated to indicate the movement of the body surface of the person 3, out of the plurality of sensor signals 15. This improves the accuracy of the estimation result of the respiratory state of the person 3 by the trained model 31 that receives the sensor signals 15 as input.

[0069] (Variation 1) In the learning system 1a according to the first modification of the second embodiment, the model generating unit 30a generates training data by using some of the multiple sensor signals 15 as input.

[0070] The training data generated by the model generation unit 30a uses each of the plurality of sensor signals 15 as an input, and sets the respiratory state of the person 3 corresponding to the input as a correct answer. At this time, the model generation unit 30a excludes, from the input of the training data, a portion of the plurality of sensor signals 15 corresponding to a period estimated by the posture estimation unit 60 as having low state reliability.

[0071] Specifically, in the teacher data, for an input that is the sensor signal 15a in the period T11, the model generation unit 30a determines the respiratory state of the person 3 in the period T11 as the correct answer. Furthermore, in the teacher data, for an input that is the sensor signal 15b in the period T11, the model generation unit 30a determines the respiratory state of the person 3 in the period T11 as the correct answer. Furthermore, in the teacher data, for an input that is the sensor signal 15a in the period T12, the model generation unit 30a determines the respiratory state of the person 3 in the period T12 as the correct answer. Furthermore, in the teacher data, for an input that is the sensor signal 15b in the period T13, the model generation unit 30a determines the respiratory state of the person 3 in the period T13 as the correct answer.

[0072] On the other hand, the model generating unit 30a does not use the sensor signal 15b in the period T12 and the sensor signal 15a in the period T13 as inputs in the training data.

[0073] This improves the accuracy of estimating the respiratory state of the person 3 in estimation using the trained model 31.

[0074] With the above configuration, as with the learning system 1a according to embodiment 2, in machine learning, the portion of the sensor signal 15 in which the amplitude of the component indicating the body surface movement due to breathing of the person 3 is small is not used for learning abnormal breathing, and the portion that clearly indicates the body surface movement due to breathing of the person 3 is used for learning abnormal breathing. Therefore, the estimation accuracy of the trained model 31 is improved.

[0075] (Variation 2) In the learning system 1a according to the second modification of the second embodiment, the posture estimation unit 60 acquires information indicating the posture of the person 3.

[0076] For example, when the signal acquisition unit 10 of the learning system 1a generates the plurality of sensor signals 15 acquired, an acceleration sensor is attached to the person 3 as a PSG sensor. Then, the posture estimation unit 60 acquires time series data indicating the posture of the person 3 based on the acceleration sensor as a result of the PSG. The time series data indicating the posture of the person 3 is, for example, data indicating each of the periods in which the person 3 is lying on their back, facing right, facing left, and lying face down during the acquisition period of the plurality of sensor signals 15. Note that the acceleration sensor attached to the person 3 does not need to be part of the PSG sensors. Furthermore, the time series data indicating the posture of the person 3 does not need to be information based on an acceleration sensor as long as it is data based on something other than the plurality of sensor signals 15.

[0077] With the above configuration, similarly to the learning system 1a according to the second embodiment, the trained model 31 is generated based on the sensor signal 15 that is estimated to represent the movement of the body surface of the person 3 among the multiple sensor signals 15. Therefore, the accuracy of the estimation result of the respiratory state of the person 3 by the trained model 31 that uses the sensor signal 15 as an input is improved.

[0078] (Embodiment 3) (1) Composition The learning system 1b according to the third embodiment further includes a signal processing unit 70 in addition to the configuration of the learning system 1 according to the first embodiment, and includes a model generation unit 30b instead of the model generation unit 30. The signal processing unit 70 converts a plurality of (e.g., three) sensor signals 15 into one or more (e.g., two) sensor signals. Here, the number of one or more sensor signals is equal to or less than the number of the plurality of sensor signals 15.

[0079] The signal processing unit 70 converts the plurality of sensor signals 15 into one or more sensor signals by dimensionality reduction. Specifically, the signal processing unit 70 converts the plurality of sensor signals 15 into one or more sensor signals by using, for example, principal component analysis.

[0080] More specifically, the signal processing unit 70 converts the multiple sensor signals 15 into one or more sensor signals so as to maximize the data variance. For example, the signal processing unit 70 generates a covariance matrix from the multiple sensor signals 15 and calculates the eigenvalues and eigenvectors of the covariance matrix. Then, the signal processing unit 70 extracts one or more eigenvectors as principal component directions, the number of which is equal to the number of sensor signals, starting with the eigenvector with the largest corresponding eigenvalue. The signal processing unit 70 generates one or more sensor signals that correspond one-to-one to the one or more eigenvectors by linearly combining the multiple sensor signals 15. This makes it possible to reduce the number of sensor signals 15 to one or more while maintaining the correlation between the multiple sensor signals 15.

[0081] The model generation unit 30b creates training data based on one or more sensor signals output by the signal processing unit 70 and the respiratory state of the person 3. Specifically, the model generation unit 30b receives one or more sensor signals as input and creates training data that uses the respiratory state of the person 3 as a correct answer. This performs machine learning based on one or more sensor signals in which information related to the breathing of the person 3 is extracted from the multiple sensor signals 15. Therefore, by performing dimensionality reduction while minimizing the loss of information contained in the multiple sensor signals 15, it is possible to generate a trained model 31 with high estimation accuracy.

[0082] (2) Effects The learning system 1b according to the third embodiment further includes a signal processing unit 70 that generates one or more sensor signals based on a plurality of sensor signals 15. The model generation unit 30 generates a trained model 31 based on the one or more sensor signals and the person's respiratory state. This reduces the reduction in the amount of information contained in the plurality of sensor signals 15, making it possible to generate a trained model 31 with high estimation accuracy.

[0083] Furthermore, in the learning system 1b according to the third embodiment, the signal processing unit 70 generates one or more sensor signals based on the plurality of sensor signals 15 by dimensionality reduction. This makes it possible to reduce the number of the plurality of sensor signals 15 to one or more sensor signals while maintaining the correlation between the signals.

[0084] (Aspect) A learning system (1; 1a; 1b) according to a first aspect includes a signal acquisition unit (10), a state acquisition unit (20), and a model generation unit (30; 30a; 30b). The signal acquisition unit (10) acquires a plurality of sensor signals (15) based on radio waves (42) reflected from the person (3) from a plurality of first radio wave sensors (100) that transmit radio waves (41) toward the person (3) from different directions. The state acquisition unit (20) acquires a respiratory state of the person (3) corresponding to the plurality of sensor signals (15). The model generation unit (30; 30a; 30b) generates a trained model (31) based on the plurality of sensor signals (15) and the respiratory state of the person (3). The trained model (31) estimates the respiratory state of the person (3) based on one or more sensor signals (15) input from one or more second radio wave sensors (100). The number of the one or more second radio wave sensors (100) is equal to or less than the number of the plurality of first radio wave sensors (100).

[0085] According to the learning system (1; 1a; 1b) of the above aspect, the trained model (31) is generated based on training data that receives as input signals based on a plurality of sensor signals (15), the number of which is greater than one or more sensor signals (15). Therefore, the accuracy of the estimation result of the respiratory state of the person (3) by the trained model (31) is improved.

[0086] In the learning system (1; 1a; 1b) according to the second aspect, in the first aspect, the signal acquiring unit (10) acquires a plurality of sensor signals (15) based on radio waves (42) reflected from a sleeping person (3). The state acquiring unit (20) acquires the breathing state of the person (3) while sleeping as the breathing state of the person (3).

[0087] According to the learning system (1; 1a; 1b) of the above-described embodiment, the trained model (31) can estimate the breathing state of the person (3) during sleep, thereby enabling the person (3) to notice abnormalities that are not subjectively noticeable.

[0088] In the learning system (1; 1a) according to the third aspect, in the first aspect, the model generation unit (30; 30a) uses training data in which the respiratory state of the person (3) is the correct answer for each of the plurality of sensor signals (15), to generate a trained model (31) that receives one or more sensor signals (15) as input and outputs the respiratory state of the person (3).

[0089] According to the learning system (1; 1a) of the above aspect, the trained model (31) is generated based on the sensor signal (15) that more strongly indicates the movement of the body surface of the person (3) among the plurality of sensor signals (15). Therefore, the accuracy of the estimation result of the respiratory state of the person (3) by the trained model (31) that uses the sensor signal (15) as an input is improved.

[0090] A learning system (1a) according to a fourth aspect is the learning system (1a) of any of the first to third aspects, further including a posture estimation unit (60). The posture estimation unit (60) estimates the posture of the person based on a plurality of sensor signals (15). The posture estimation unit (60) calculates a state reliability of the respiratory state of the person (3) for each of the plurality of sensor signals (15) based on the posture of the person. A model generation unit (30a) creates training data based on the plurality of sensor signals (15) and the state reliability, and generates a trained model (31) that receives one or more sensor signals (15) as input and outputs the respiratory state of the person (3).

[0091] According to the learning system (1a) of the above aspect, the trained model (31) is generated based on the sensor signal (15) estimated to indicate the movement of the body surface of the person (3) among the plurality of sensor signals (15). Therefore, the accuracy of the estimation result of the respiratory state of the person (3) by the trained model (31) using the sensor signal (15) as an input is improved.

[0092] In the learning system (1a) according to the fifth aspect, in the fourth aspect, the posture estimation unit (60) calculates the state reliability based on the posture of the person (3) and a change in the posture of the person (3).

[0093] According to the learning system (1a) of the above aspect, the posture estimation unit (60) can reduce the load of estimating the posture of the person (3) and can improve the accuracy of posture estimation by reducing erroneous recognition of posture changes.

[0094] The learning system (1b) according to the sixth aspect is the learning system (1b) of the first aspect, further including a signal processing unit (70) that generates one or more sensor signals based on a plurality of sensor signals (15). The model generating unit (30b) generates a trained model (31) based on the one or more sensor signals and the respiratory state of the person (3).

[0095] According to the learning system (1b) of the above aspect, it is possible to reduce the reduction in the amount of information contained in the multiple sensor signals (15) and generate a trained model (31) with high estimation accuracy.

[0096] In the learning system (1b) according to the seventh aspect, in the sixth aspect, the signal processing unit (70) generates one or more sensor signals based on the plurality of sensor signals (15) by dimension reduction.

[0097] According to the learning system (1b) of the above aspect, it is possible to reduce the number of multiple sensor signals (15) to one or more sensor signals while maintaining the correlation between the signals.

[0098] An estimation system (2) according to an eighth aspect includes an estimation unit (50) and a second signal acquisition unit (40) different from the first signal acquisition unit (10) which is a signal acquisition unit. The estimation unit (50) estimates a person's respiratory state using a trained model (31) generated by the learning system (1) according to any one of the first to seventh aspects. The second signal acquisition unit (40) acquires, from one or more second radio wave sensors (100) which transmit radio waves (41) to the person (3), one or more sensor signals (15) based on radio waves (42) reflected from the person (3). The estimation unit (50) inputs the one or more sensor signals (15) acquired by the second signal acquisition unit (40) to the trained model (31).

[0099] According to the estimation system (2) of the above aspect, the learned model (31) has high estimation accuracy, so that even if the number of sensor signals (15) is small, the respiratory state of the person (3) can be estimated with high accuracy.

[0100] A learning method according to a ninth aspect is executed by one or more processors. The learning method includes a signal acquiring step, a state acquiring step, and a learning step. In the signal acquiring step, a plurality of sensor signals (15) based on radio waves (42) reflected from the person (3) are acquired from a plurality of first radio wave sensors (100) that transmit radio waves (41) toward the person (3) from mutually different directions. In the state acquiring step, a respiratory state of the person (3) corresponding to the plurality of sensor signals (15) is acquired. In the model generating step, a trained model (31) is generated based on the plurality of sensor signals (15) and the respiratory state of the person (3). The trained model (31) estimates the respiratory state of the person (3) based on one or more sensor signals (15) input from one or more second radio wave sensors (100). The number of the one or more second radio wave sensors (100) is equal to or less than the number of the plurality of first radio wave sensors (100).

[0101] According to the learning method of the above aspect, the trained model (31) is generated based on training data that receives as input signals based on a plurality of sensor signals (15), the number of which is greater than one or more sensor signals (15). Therefore, the accuracy of the estimation result of the respiratory state of the person (3) by the trained model (31) is improved.

[0102] An estimation method according to a tenth aspect is executed by one or more processors. The estimation method includes an estimation step and a second signal acquisition step different from the first signal acquisition step, which is a signal acquisition step. In the estimation step, a respiratory state of a person (3) is estimated using a trained model (31) generated by the learning method according to the ninth aspect. In the second signal acquisition step, one or more sensor signals (15) based on radio waves (42) reflected from the person (3) are acquired from one or more second radio wave sensors (100) that transmit radio waves (41) to the person (3). In the estimation step, the one or more sensor signals (15) acquired in the second signal acquisition step are input to the trained model (31).

[0103] According to the estimation method of the above aspect, the learned model (31) has high estimation accuracy, so that even if the number of sensor signals (15) is small, the respiratory state of the person (3) can be estimated with high accuracy. [Explanation of symbols]

[0104] 1, 1a, 1b Learning System 2. Estimation System 10. Signal acquisition unit (first signal acquisition unit) 20 Status acquisition unit 30, 30a, 30b Model generation unit 40 signal acquisition unit (second signal acquisition unit) 41 Radio Waves 42 Received waves (radio waves) 50 Estimation part 60 Posture estimation section 70 Signal Processing Section 3 people 15 Sensor Signal 31 Trained Models 100 Radio wave sensor (first radio wave sensor, second radio wave sensor)

Claims

1. a signal acquisition unit that acquires a plurality of sensor signals based on radio waves reflected from a person from a plurality of first radio wave sensors that transmit radio waves to the person from different directions; a state acquisition unit that acquires a respiratory state of the person corresponding to the plurality of sensor signals; a model generation unit that generates a trained model that estimates the respiratory state of the person based on one or more sensor signals input from one or more second radio wave sensors, based on the plurality of sensor signals and the respiratory state of the person; Equipped with the number of the one or more second radio wave sensors is equal to or less than the number of the plurality of first radio wave sensors; Learning system.

2. the signal acquisition unit acquires a plurality of sensor signals based on radio waves reflected from the sleeping person; the state acquisition unit acquires a respiratory state of the person while sleeping as the respiratory state of the person. The learning system of claim 1 .

3. the model generation unit generates the trained model using training data that sets the person's respiratory state as a correct answer for each of the plurality of sensor signals, with one or more sensor signals as an input and the person's respiratory state as an output. The learning system of claim 1 .

4. a posture estimation unit that estimates a posture of the person based on the plurality of sensor signals; the posture estimation unit calculates a state reliability of the person's respiratory state for each of the plurality of sensor signals based on the posture of the person; the model generation unit creates training data based on the plurality of sensor signals, the respiratory state of the person, and the state reliability, and generates the trained model that receives the one or more sensor signals as input and outputs the respiratory state of the person. The learning system of claim 1 .

5. the posture estimation unit calculates the state reliability based on the posture of the person and a change in the posture of the person. The learning system according to claim 4 .

6. a signal processing unit that generates one or more sensor signals based on the plurality of sensor signals; The model generation unit generates the trained model based on the one or more sensor signals and the respiratory state of the person. The learning system of claim 1 .

7. The signal processing unit generates one or more sensor signals based on the plurality of sensor signals by dimensionality reduction. The learning system of claim 6.

8. an estimation unit that estimates a respiratory state of a person using the trained model generated by the learning system according to claim 1; a second signal acquisition unit different from the first signal acquisition unit that is the signal acquisition unit and that acquires, from one or more second radio wave sensors that transmit radio waves to a person, one or more sensor signals based on radio waves reflected from the person; The estimation unit inputs the one or more sensor signals acquired by the second signal acquisition unit into the trained model. Estimation system.

9. Executed by one or more processors, a signal acquiring step of acquiring, from a plurality of first radio wave sensors that transmit radio waves toward the person from different directions, a plurality of sensor signals based on radio waves reflected from the person; a state acquiring step of acquiring a respiratory state of the person corresponding to the plurality of sensor signals; a model generation step of generating a trained model that estimates the respiratory state of the person based on the sensor signals input from one or more second radio wave sensors, based on the plurality of sensor signals and the respiratory state of the person; the number of the one or more second radio wave sensors is equal to or less than the number of the plurality of first radio wave sensors; How to learn.

10. Executed by one or more processors, an estimation step of estimating a respiratory state of a person using the trained model generated by the training method according to claim 9; a second signal acquisition step different from the first signal acquisition step, which is the signal acquisition step, of acquiring one or more sensor signals based on radio waves reflected from the person from one or more second radio wave sensors that transmit radio waves to the person; In the estimation step, the one or more sensor signals acquired in the second signal acquisition step are input to the trained model. Estimation method.

Images