JP2022015733A - Fatigue state determination method, learning model generation method, fatigue state determination system, fatigue state determination device, and computer program - Google Patents

Abstract

To provide a fatigue state determination method, a learning model generation method, a fatigue state determination system, a fatigue state determination device, and a computer program for determining presence or absence of fatigue in the lower extremity by using an acceleration sensor.SOLUTION: In a fatigue state determination method, an acceleration rate, an angular speed, and earth magnetism are repeatedly measured by a sensor mounted on the leg of a person while the person is walking. In a case where a history of the acceleration rate, the angular speed, and the earth magnetism while the person is walking is input, the history of the acceleration rate, the angular speed, and the earth magnetism measured by the sensor is input to a learning model for outputting information indicating whether a state of the person is the state before performing a motion accompanied by fatigue in the lower extremity or it is the state after that. According to the information output by the learning model, it is determined whether the state of the person wearing the sensor on the leg is the state before performing the motion or the state after that.SELECTED DRAWING: Figure 1

Description

The present invention relates to a fatigue state determination method for determining the presence or absence of fatigue in the lower limbs, a learning model generation method, a fatigue state determination system, a fatigue state determination device, and a computer program.

Fatigue adversely affects health and can cause accidents or illness. However, it is difficult for a person to be aware of fatigue accurately. Therefore, a technique for objectively determining human fatigue is required. For example, a method using a myoelectric potential measured by an electromyogram is known. The average frequency or the intermediate frequency obtained by performing the frequency analysis of the myoelectric potential decreases as the muscle becomes tired. Therefore, muscle fatigue can be determined based on the average frequency or the intermediate frequency of the myoelectric potential. Patent Document 1 discloses an example of a technique for measuring a myoelectric potential and determining muscle fatigue based on the myoelectric potential.

Japanese Unexamined Patent Publication No. 2017-221647

In the technology using myoelectric potential, it is necessary to attach an electromyogram to the human body. However, it is not easy to keep the EMG attached to the human body for a long time, which causes a burden on the human body. Therefore, a technique for estimating fatigue more easily is required. In addition, movements accompanied by fatigue of the lower limbs such as walking are movements that a person performs on a daily basis, and fatigue of the lower limbs occurs on a daily basis. If the occurrence or recovery of fatigue in the lower limbs can be easily determined, it can be useful for estimating daily fatigue.

The present invention has been made in view of such circumstances, and an object thereof is a fatigue state determination method for determining the presence or absence of fatigue in the lower limbs, a learning model generation method, and a fatigue state using an acceleration sensor. It is an object of the present invention to provide a determination system, a fatigue state determination device, and a computer program.

The fatigue state determination method according to the present invention is a case where the acceleration, angular velocity and geomagnetism during walking are repeatedly measured by a sensor mounted on a human foot, and the history of acceleration, angular velocity and geomagnetism during walking of a person is input. , The history of acceleration, angular velocity and geomagnetism measured by the sensor to a learning model that outputs information indicating whether the person's state is before or after performing a motion accompanied by fatigue of the lower limbs. It is characterized in that it determines whether the state of the person who wears the sensor on the foot is the state before or after the operation according to the input and the information output by the learning model. ..

In the fatigue state determination method according to the present invention, when the learning model is input with the history of acceleration, angular velocity and geomagnetism during walking of a person, and the characteristic amount of myoelectric potential obtained from the foot of the person. Characteristics of myoelectric potential obtained from a person's foot, which has been learned to output information indicating whether the person's state is a state before or after performing a motion accompanied by fatigue of the lower limbs. When the amount is not input and the acceleration, angular velocity, and geomagnetic history of the person during walking are input, the state of the person is the state before or after the movement accompanied by the fatigue of the lower limbs. It is characterized by outputting information indicating the above.

In the fatigue state determination method according to the present invention, the sensor measures acceleration, angular velocity, and geomagnetism in a direction along an axis fixed to the sensor, and the attitude angle of the axis is determined by acceleration, angular velocity, or geomagnetism. It is characterized by calculating based on history and correcting acceleration, angular velocity and geomagnetism based on the attitude angle.

In the fatigue state determination method according to the present invention, the acceleration, the angular velocity and the geomagnetism in the direction along the axis fixed to the sensor are measured by the sensor, and the acceleration, the angular velocity and the geomagnetism are measured at each time point. The attitude angle of the axis fixed to the sensor with respect to the axis fixed to the ground is calculated based on the history of acceleration, angular velocity and geomagnetism, and is fixed to the sensor based on the attitude angle. The acceleration, angular velocity and geomagnetism in the direction along the axis are converted into the acceleration, angular velocity and geomagnetism in the direction along the axis fixed to the ground, and the converted acceleration, angular velocity and geomagnetism are input to the learning model. It is characterized by doing.

In the fatigue state determination method according to the present invention, the learning model is a model that outputs the information when the history of acceleration, angular velocity, and geomagnetism in one walking period with the same foot is input, and is repeatedly measured by the sensor. It is characterized in that the history of acceleration, angular velocity and geomagnetism measured during one walking period is extracted from the acceleration, angular velocity and geomagnetism, and the extracted acceleration, angular velocity and geomagnetism history are input to the learning model. ..

In the fatigue state determination method according to the present invention, the time when the foot to which the sensor is attached touches the ground is specified based on the change in the acceleration in the vertical direction and the change in the angular velocity corresponding to the shaking in the front-back direction, and the foot touches the ground. It is characterized by extracting the history of acceleration, angular velocity and geomagnetism measured during the period from the time when the foot touches the ground to the time when the foot touches the ground as the history of acceleration, angular velocity and geomagnetism measured during one walking period. do.

In the fatigue state determination method according to the present invention, the history of acceleration, angular velocity, and geomagnetism measured in each of a plurality of walking periods is individually input to the learning model, and the learning model is used for the plurality of walking periods. It is characterized in that it is determined whether the state of the person who wears the sensor on the foot is the state before or after the operation according to the plurality of the information output by the user.

The learning model generation method according to the present invention includes acceleration, angular velocity, and geomagnetic history measured during one walk by a sensor attached to a person's foot before performing a motion accompanied by fatigue of the lower limbs, and the person's foot. Acceleration, angular velocity and geomagnetic history measured during one walk by a sensor attached to a person's foot after performing the above motion, and the history of the person's foot, as well as the history of acceleration, angular velocity and geomagnetism obtained from The myoelectricity feature amount obtained from the above was created, and the acceleration, angular velocity, and geomagnetic history measured in each of the plurality of gait periods, the myoelectricity feature amount, and the human state before performing the above-mentioned movement. When the history of acceleration, angular velocity, and geomagnetism during walking of a person is input as training data with information indicating whether the state is the state of the person or the state after the state of the person, the state of the person before the operation is performed. It is characterized by generating a learning model that outputs information indicating whether it is in a certain state or a later state.

The learning model generation method according to the present invention creates a plurality of history of acceleration, angular velocity, and geomagnetism measured during one walking by a sensor attached to a person's foot before performing a motion accompanied by fatigue of the lower limbs. When the history of acceleration, angular velocity and geomagnetism during walking of a person is input by unsupervised learning using the history of acceleration, angular velocity and geomagnetism measured in each walking period as training data, the state of the person is changed. It is characterized by generating a learning model that outputs information indicating the degree of abnormality of the person's state as information indicating whether the state is before or after the operation.

The fatigue state determination system according to the present invention is a sensor that is attached to a person's foot and repeatedly measures acceleration, angular velocity, and geomagnetism during walking of the person, and the state of the person who wears the sensor on the foot causes fatigue of the lower limbs. The fatigue state determination device is provided with a fatigue state determination device for determining whether the state is before or after the accompanying operation, and the fatigue state determination device inputs the history of acceleration, angular velocity, and geomagnetism during walking of a person. In addition, a learning model is provided which outputs information indicating whether the state of the person is a state before or after the operation, and the acceleration, angular velocity, and geomagnetism measured by the sensor are acquired and acquired. Whether the state of the person wearing the sensor on the foot is the state before the operation is performed according to the information output by the learning model when the history of acceleration, angular velocity, and geomagnetism is input to the learning model. It is characterized in that it is determined whether or not it is in the state of.

The fatigue state determination device according to the present invention is a state before or after the person's state is a state before performing an operation accompanied by fatigue of the lower limbs when the history of acceleration, angular velocity and geomagnetism during walking of the person is input. It is equipped with a learning model that outputs information indicating whether or not it is, and acquires acceleration, angular velocity, and geomagnetism during walking measured by a sensor mounted on a human foot, and learns the acquired acceleration, angular velocity, and geomagnetism history. According to the information output by the learning model when input to the model, it is determined whether the state of the person wearing the sensor on the foot is the state before or after the operation. It is a feature.

The computer program according to the present invention is a state before or after the state of the person performing an operation accompanied by fatigue of the lower limbs when the history of acceleration, angular velocity and geomagnetism during walking of the person is input. The history of acceleration, angular velocity, and geomagnetism during walking measured by a sensor mounted on a human foot is input to a learning model that outputs information indicating the above, and the above information is output according to the information output by the learning model. It is characterized in that a computer is made to execute a process of determining whether the state of a person wearing a sensor on his / her foot is a state before or after the operation.

In the present invention, the sensor mounted on the human foot measures the acceleration, the angular velocity and the geomagnetism during walking, and the history of the acceleration, the angular velocity and the geomagnetism is input to the learning model. The learning model shows whether the state of the person wearing the sensor on the foot is the state before or after the movement with fatigue of the lower limbs when the history of acceleration, angular velocity and geomagnetism is input. It is learned to output information. Acceleration, angular velocity, and geomagnetism change due to the effects of fatigue before and after movements involving fatigue of the lower limbs. Therefore, by using a properly trained learning model, it is easy to determine whether the state of the person wearing the sensor on the foot is the state before or after the movement accompanied by the fatigue of the lower limbs. It can be determined.

In one embodiment of the present invention, in the learning model, when the acceleration, the angular velocity, the geomagnetic history and the feature amount of the myoelectric potential during walking are input, the human state moves with the fatigue of the lower limbs. It is learned to output information indicating whether the state is before or after the operation. At the time of learning, by probabilistically inactivating the node related to the feature amount of myoelectric potential, learning is performed so that an appropriate determination can be made even when the feature amount of myoelectric potential is not input. At the time of determining the fatigue state, the learning model is in the state before the human state performs the movement accompanied by the fatigue of the lower limbs when the feature amount of the myoelectric potential is not input and the history of acceleration, angular velocity and geomagnetism is input. Outputs information indicating whether the state is present or later. The learning model is properly learned by learning using the feature amount of the myoelectric potential. When determining the fatigue state, it is possible to easily determine whether the human state is a state before or after performing an operation accompanied by fatigue of the lower limbs by performing the judgment without using the feature amount of the myoelectric potential. be able to.

In one embodiment of the invention, a sensor mounted on a human foot measures acceleration, angular velocity and geomagnetism in a direction along an axis fixed to the sensor. The attitude angle of the axis fixed to the sensor is calculated based on the time change of acceleration, angular velocity or geomagnetism, and the acceleration, angular velocity and geomagnetism are corrected based on the attitude angle. For example, the attitude angle of the fixed axis with respect to the sensor with respect to the fixed axis with respect to the ground is calculated. Based on the attitude angle, the acceleration, angular velocity and geomagnetism are converted into acceleration, angular velocity and geomagnetism in the direction of the axis fixed with respect to the ground, and the converted acceleration, angular velocity and geomagnetism are input to the learning model. Acceleration, angular velocity and geomagnetic correction eliminate the effects of sensor movement and clarify the effects of walking. Therefore, it is possible to accurately determine the state of a person.

In one embodiment of the present invention, the history of acceleration, angular velocity, and geomagnetism during one walking with the same foot is input to the learning model, and the state of the person is determined. During the period of one walk, the movement corresponding to one cycle in the periodic movement by one foot is performed, and the history of acceleration, angular velocity and geomagnetism reflects the characteristics of walking. Therefore, the state of a person can be determined based on the history of acceleration, angular velocity, and geomagnetism during one walking period.

In one embodiment of the present invention, the time point at which the foot touches the ground is specified based on the changes in acceleration and angular velocity, and the period from the time when the foot touches the ground to the time when the foot touches the next touches is set as the period of one walk. History of acceleration, angular velocity and geomagnetism over the period is obtained. Since the angular velocity corresponding to the vertical acceleration and the front-back sway changes greatly before and after the foot touches the ground, it is possible to specify the time point when the foot touches the ground.

In one embodiment of the present invention, the acceleration, angular velocity, and geomagnetic history are individually input to the learning model for each of the plurality of walking periods, and the state of the person is determined based on the plurality of information output by the learning model. It is judged. By using a plurality of information output by the learning model, more accurate judgment becomes possible.

In one embodiment of the present invention, the acceleration, angular velocity, geomagnetic history, and myoelectric potential features obtained before performing the motion with fatigue of the lower limbs, and the acceleration obtained after performing the motion with fatigue of the lower limbs. A training model is generated using training data including angular velocity, geomagnetic history, and myoelectric potential features. By learning a learning model using the data obtained before the movement with fatigue of the lower limbs and the data obtained after the movement with fatigue of the lower limbs as training data, the information indicating the human condition can be correctly obtained. A properly trained learning model can be generated for output. Further, by learning the learning model using the features of the myoelectric potential in addition to the acceleration, the angular velocity, and the geomagnetism, the learning model can be learned so that an accurate determination can be made.

In one embodiment of the present invention, the learning model indicates the degree of abnormality in the human condition as information indicating whether the human condition is a state before or after performing a movement accompanied by fatigue of the lower limbs. Output information. In the learning model, learning is performed by unsupervised learning using the history of acceleration, angular velocity, and geomagnetism obtained before performing a motion accompanied by fatigue of the lower limbs as training data. Since it is not necessary to prepare the data obtained after the movement accompanied by the fatigue of the lower limbs as the training data, the training data can be easily created and the learning model can be generated.

INDUSTRIAL APPLICABILITY The present invention has excellent effects such as being able to easily determine whether a person's state is a state before or after performing an operation accompanied by fatigue of the lower limbs.

It is a schematic diagram which shows the structural example of the fatigue state determination system which concerns on Embodiment 1. It is a block diagram which shows the structural example of the inside of a sensor device. It is a block diagram which shows the configuration example inside the terminal apparatus. It is a block diagram which shows the example of the function structure inside the fatigue state determination device. It is a graph which shows the example of the time change of acceleration and angular velocity during walking. It is a conceptual diagram which shows the function of the learning model which concerns on Embodiment 1. FIG. It is a conceptual diagram which shows the structural example of the learning model which concerns on Embodiment 1. FIG. It is a block diagram which shows the configuration example of the learning apparatus which trains a learning model. It is a flowchart which shows the procedure of the process which generates a learning model. It is a flowchart which shows the procedure of the process which the fatigue state determination system acquires acceleration, angular velocity and geomagnetism. It is a flowchart which shows the procedure of the determination process executed by the fatigue state determination apparatus. It is a conceptual diagram which shows the outline of DAGMM. It is a conceptual diagram which shows the structural example of the learning model which concerns on Embodiment 2. It is a conceptual diagram which shows the function of the learning model which concerns on Embodiment 3. It is a schematic diagram which shows the structural example of the fatigue state determination system which concerns on Embodiment 4.

Hereinafter, the present invention will be specifically described with reference to the drawings showing the embodiments thereof.
<Embodiment 1>
FIG. 1 is a schematic diagram showing a configuration example of the fatigue state determination system 100 according to the first embodiment. The fatigue state determination system 100 is a system for determining whether the state of the user 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs as the fatigue state of the user 4. The fatigue state determination system 100 executes the fatigue state determination method. The fatigue state determination system 100 includes a sensor device 3, a terminal device 2, and a fatigue state determination device 1.

The sensor device 3 is attached to the foot of the user 4 and measures the physical quantity of the user 4 during walking. The sensor device 3 transmits data to the terminal device 2. The terminal device 2 receives the data transmitted from the sensor device 3 and communicates with the fatigue state determination device 1 via a communication network 5 such as the Internet. The fatigue state determination device 1 is connected to the communication network 5. The fatigue state determination device 1 performs a process of determining whether the state of the user 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs, based on the measurement result of the sensor device 3. ..

In the present embodiment, as the fatigue state of the user 4, it is determined whether the state of the user 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs. It is known that as the muscle becomes tired, slowing down occurs in which the intermediate frequency and the average frequency of the myoelectric potential shift to the low frequency side. The movement accompanied by fatigue of the lower limbs is a movement of such an intensity that fatigue occurs in the lower limbs, such that the myoelectric potential of the lower limbs is slowed down. The movement accompanied by fatigue of the lower limbs includes any movement as long as the movement is strong enough to cause fatigue in the lower limbs. The state before performing the movement accompanied by fatigue of the lower limbs includes a state in which the movement accompanied by fatigue of the lower limbs is not performed and a state in which the fatigue caused by the movement is recovered. The state after performing the movement accompanied by fatigue of the lower limbs is a state in which the movement accompanied by fatigue of the lower limbs is performed and the fatigue has not been recovered yet. By determining whether the state of the user 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs, it is possible to determine the state of fatigue and recovery from fatigue of the lower limbs of the user 4. ..

FIG. 2 is a block diagram showing an example of the internal configuration of the sensor device 3. The sensor device 3 includes a control unit 31, an acceleration sensor 32 for measuring acceleration, an angular velocity sensor 33 for measuring angular velocity, a geomagnetic sensor 34 for measuring geomagnetism, and a communication unit 35. The control unit 31 controls each part of the sensor device 3. The acceleration sensor 32, the angular velocity sensor 33, and the geomagnetic sensor 34 are all three-axis sensors. That is, the acceleration sensor 32, the angular velocity sensor 33, and the geomagnetic sensor 34 measure the components of the acceleration, the angular velocity, and the geomagnetism in the directions along the three axes fixed to the sensor device 3 and orthogonal to each other. It is desirable that the accelerometer 32, the angular velocity sensor 33, and the geomagnetic sensor 34 have the same three axes.

The communication unit 35 transmits data including acceleration, angular velocity, and geomagnetic values measured by the angular velocity sensor 33 and the geomagnetic sensor 34 to the outside of the sensor device 3 by wireless communication. The control unit 31 causes the communication unit 35 to transmit data including acceleration, angular velocity, and geomagnetic values to the terminal device 2. The acceleration sensor 32, the angular velocity sensor 33, and the geomagnetic sensor 34 repeatedly measure acceleration, angular velocity, and geomagnetism.

The sensor device 3 is attached to the foot of the user 4 by attaching a belt, winding the belt around the foot of the user 4, and fixing the belt by a hook-and-loop fastener provided on the belt. Alternatively, the sensor device 3 is provided with a hook-and-loop fastener, and may be fixed to the clothes of the user 4 by the hook-and-loop fastener. The sensor device 3 may be attached to the user 4's foot by being housed in the bag and the bag being fixed to the user 4's foot by a fixing tool such as a belt, a hook-and-loop fastener, or a button. The sensor device 3 may be attached to the foot of the user 4 by being stored in a storage portion such as a pocket of the clothes of the user 4. The sensor device 3 may be attached to the socks or shoes of the user 4.

FIG. 3 is a block diagram showing an example of the internal configuration of the terminal device 2. The terminal device 2 is a portable computer such as a smartphone or a tablet computer. For example, the terminal device 2 is the property of the user 4. The terminal device 2 includes a calculation unit 21, a memory 22, a storage unit 23, an operation unit 24, a display unit 25, and a communication unit 26. The arithmetic unit 21 is configured by using, for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), or a multi-core CPU. The arithmetic unit 21 may be configured by using a quantum computer. The memory 22 stores temporary data generated by the calculation. The memory 22 is, for example, a RAM (Random Access Memory). The storage unit 23 is non-volatile, and is, for example, a hard disk or a non-volatile semiconductor memory. The operation unit 24 accepts the input of information such as text by accepting the operation from the user 4. The operation unit 24 is, for example, a touch panel. The display unit 25 displays an image. The display unit 25 is, for example, a liquid crystal display or an EL display (Electroluminescent Display). The communication unit 26 receives the data transmitted from the sensor device 3 by wireless communication. Further, the communication unit 26 communicates with the fatigue state determination device 1 via the communication network 5 by wireless communication.

The storage unit 23 stores the computer program 231. The computer program 231 is downloaded using the communication unit 26 and stored in the storage unit 23. For example, the computer program 231 is downloaded from the fatigue state determination device 1. The computer program 231 may be stored in the storage unit 23 in advance. Alternatively, the computer program 231 may be stored in the memory 22 instead of the storage unit 23. For example, the computer program 231 may be downloaded and stored in the memory 22 when the process related to the fatigue state determination system 100 is executed, and may be erased from the memory 22 when the process related to the fatigue state determination system 100 is completed. The arithmetic unit 21 executes necessary processing according to the computer program 231.

FIG. 4 is a block diagram showing an example of the internal functional configuration of the fatigue state determination device 1. The fatigue state determination device 1 is configured by using a computer such as a server device. The fatigue state determination device 1 communicates with a calculation unit 11, a memory 12 for storing temporary data generated by the calculation, a storage unit 13, and a drive unit 14 for reading information from a recording medium 10 such as an optical disk. It is provided with a unit 15. The arithmetic unit 11 is configured by using, for example, a CPU, a GPU, or a multi-core CPU. The arithmetic unit 11 may be configured by using a quantum computer. The memory 12 is, for example, a RAM. The storage unit 13 is non-volatile, for example, a hard disk. The communication unit 15 is connected to the communication network 5. The communication unit 15 communicates with the terminal device 2 via the communication network 5.

The calculation unit 11 causes the drive unit 14 to read the computer program 131 recorded on the recording medium 10, and stores the read computer program 131 in the storage unit 13. The calculation unit 11 executes the processing required for the fatigue state determination device 1 according to the computer program 131. The computer program 131 may be downloaded from the outside of the fatigue state determination device 1 by the communication unit 15. In this case, the fatigue state determination device 1 does not have to include the drive unit 14.

The fatigue state determination device 1 is used to determine whether the state of the user 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs based on the measurement result of the sensor device 3. It is equipped with a learning model 16. In the learning model 16, the state of the user 4 is the state before or after the movement accompanied by the fatigue of the lower limbs when the acceleration, the angular velocity, the history of the geomagnetism, and the feature amount of the myoelectric potential of the lower limbs are input. It is learned to output information indicating whether or not it is. Further, the learning model 16 learns while stochastically inactivating the nodes related to the myoelectric potential features, so that the myoelectric potential features of the lower limbs are not input and the acceleration, angular velocity, and geomagnetic history are input. When this is done, it is configured to output information indicating whether the state of the user 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs.

The learning model 16 is realized by the arithmetic unit 11 executing information processing according to the computer program 131. The storage unit 13 stores data necessary for realizing the learning model 16. The learning model 16 may be configured by using hardware. For example, the learning model 16 may be composed of hardware including a processor and a memory for storing necessary programs and data. Alternatively, the learning model 16 may be realized by using a quantum computer.

A method of determining whether the state of the user 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs based on the measurement result of the sensor device 3 will be described. The sensor device 3 measures the acceleration and the angular velocity generated when the user 4 who wears the sensor device 3 on his / her foot walks, and measures the geomagnetism when the user 4 walks. After the movement with fatigue of the lower limbs, the movement of walking changes due to fatigue as compared with the movement with fatigue of the lower limbs. For example, the height at which the foot is swung up, the direction of the foot, the speed at which the foot is moved, and the like change. Therefore, the acceleration, the angular velocity, and the geomagnetism measured by the sensor device 3 mounted on the foot change before and after the movement accompanied by the fatigue of the lower limbs. Therefore, based on the measurement result of the sensor device 3, it can be determined whether the state of the user 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs.

FIG. 5 is a graph showing an example of changes in acceleration and angular velocity during walking over time. The horizontal axis of FIG. 5 indicates time, and the vertical axis indicates acceleration and angular velocity. Acceleration and angular velocity have positive or negative values depending on the orientation. The thick line in FIG. 5 indicates the acceleration in the Z direction, and the thin line indicates the angular velocity in the X direction. The Z direction is the vertical direction. The X direction is one of the X and Y directions orthogonal to each other included in the plane orthogonal to the Z direction. For example, the X-direction, the Y-direction, and the Z-direction are directions along the X-axis, the Y-axis, and the Z-axis, which are fixed to the ground and orthogonal to each other. In the example shown in FIG. 5, the angular velocity in the X direction corresponds to the sway of the foot of the user 4 in the anteroposterior direction. The X direction and the Y direction may be set so that the angular velocity in the Y direction corresponds to the shaking of the foot of the user 4 in the front-back direction.

FIG. 5 shows an example of acceleration and angular velocity measured by the sensor device 3 mounted on one leg. When walking, the foot moves periodically, and the acceleration and angular velocity change periodically. Immediately before the foot touches the ground, the foot moves downward, and immediately after that, it moves upward. Therefore, the acceleration in the Z direction greatly changes from downward to upward before and after touchdown. Immediately before the foot touches the ground, the foot swings forward to step on, and immediately after that, the foot swings backward to kick the ground backward. Therefore, the angular velocity in the X direction changes direction before and after touchdown, and changes greatly. Therefore, it is possible to specify the time when the foot touches the ground based on the change of the acceleration in the Z direction and the acceleration in the X direction with respect to time. For example, when the acceleration in the Z direction exceeds the positive first threshold value and the angular velocity in the X direction is lower than the negative second threshold value within the immediately preceding predetermined period, the acceleration in the Z direction exceeds the first threshold value. The time when is exceeded is specified as the time of grounding.

In the present embodiment, the period from the time when the same foot touches the ground to the time when the same foot touches the ground next is defined as the period of one walking with one foot. The period of one walk with one foot corresponds to the period of two steps with both feet. Since the time when the foot touches the ground can be specified, the period of one walk can be specified. In addition, the history of acceleration, angular velocity, and geomagnetism measured multiple times during one walking can be extracted from the measurement results of acceleration, angular velocity, and geomagnetism that are repeatedly measured. FIG. 5 shows the time when the foot touches the ground with a broken line, and shows the period of each walking. The duration of one walk is not constant. Therefore, the number of measurements of acceleration, angular velocity, and geomagnetism during one walking period is not constant.

In the present embodiment, the determination is made based on the history of acceleration, angular velocity, and geomagnetism during one walking period. During the period of one walk, the movement corresponding to one cycle in the periodic movement by one foot is performed, and the acceleration, the angular velocity and the geomagnetism cause the change corresponding to one cycle in the periodic change. Therefore, the history of acceleration, angular velocity, and geomagnetism during one walking period reflects the walking characteristics of the user 4 and can be information for determining the state of the user 4.

Further, in the present embodiment, the myoelectric potential is also used for learning the learning model 16. The action potential is an action potential generated when a muscle is active, and is measured by an electromyogram. By performing frequency analysis of the measured myoelectric potential, the feature amount of the myoelectric potential can be obtained. For example, frequency analysis is performed by Fourier transforming the measurement result of the electromyogram. The characteristic amount of myoelectric potential is a statistic of myoelectric potential that represents the characteristic of myoelectric potential and shows a change depending on the presence or absence of muscle fatigue. For example, the feature amount of the myoelectric potential is an intermediate frequency or an average frequency of the myoelectric potential. The intermediate frequency of the myoelectric potential is a frequency that bisects the area of the power spectrum of the myoelectric potential obtained by frequency analysis. The average frequency of the myoelectric potential is the average of the frequencies of the myoelectric potential obtained by the frequency analysis.

Due to the slow wave, the higher the degree of muscle fatigue, the lower the intermediate frequency and average frequency of the myoelectric potential. On the contrary, when the muscle is not tired or the muscle is recovered from fatigue, the intermediate frequency and the average frequency of the myoelectric potential are high. Therefore, the feature amount of the myoelectric potential obtained from the lower limbs of the user 4 is an amount depending on whether the state of the user 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs. Become. Therefore, by learning the learning model 16 using the feature amount of the myoelectric potential, it is possible to accurately determine whether the state of the user 4 is the state before or after the movement accompanied by the fatigue of the lower limbs. The learning model 16 can be trained so that it can be determined.

FIG. 6 is a conceptual diagram showing the function of the learning model 16 according to the first embodiment. The learning model 16 includes a history of accelerations in the X, Y, and Z directions during one walk, a history of angular velocities in the X, Y, and Z directions during one walk, and an X during one walk. The history of geomagnetism in the direction, Y direction, and Z direction is input. The history of acceleration in the period of one walk is the value of a plurality of accelerations measured sequentially in the period of one walk. The accelerations in the X direction, the Y direction, and the Z direction are the accelerations in the X direction, the Y direction, and the Z direction, respectively. The same applies to the angular velocity and the geomagnetism. Further, the feature amount of the myoelectric potential is input to the learning model 16. The learning model 16 is a state before the state of the user 4 performs an operation accompanied by fatigue of the lower limbs when the acceleration, the angular velocity, the geomagnetic history, and the feature amount of the myoelectric potential in one walking period are input. It is learned to output state information indicating whether it is a later state. Further, the learning model 16 learns while probabilistically inactivating the node related to the feature amount of the myoelectric potential, so that the feature amount of the myoelectric potential of the lower limbs is not input, and the acceleration, the angular velocity, and the angular velocity during one walking period are not input. It is configured to output state information when geomagnetic history is input.

FIG. 7 is a conceptual diagram showing a configuration example of the learning model 16 according to the first embodiment. FIG. 7 shows an example in which the learning model 16 is configured using a fully coupled neural network including an input layer 161 and a plurality of intermediate layers 1621, 1622, ..., 162n and an output layer 163. n is the number of intermediate layers. Circles in FIG. 7 indicate nodes. The input layer 161 has a plurality of nodes, each of which is input with a plurality of values of acceleration in the X direction during one walking period. Further, the input layer 161 has a plurality of nodes in which a plurality of values of acceleration in the Y direction and a plurality of values of acceleration in the Z direction are input, respectively. Further, the input layer 161 has a plurality of nodes for inputting a plurality of values of the angular velocities in the X direction, the Y direction and the Z direction, and a plurality of values of the geomagnetism in the X direction, the Y direction and the Z direction, respectively. Further, the input layer 161 has a node to which the feature amount of the myoelectric potential is input. For example, acceleration, angular velocity, and geomagnetism are measured m (m is a natural number) times during one walking period, and (9m + 1) values are input to the input layer 161.

The learning model 16 has an intermediate layer of n layers. The first intermediate layer 1621 has a plurality of nodes. Each node of the input layer 161 outputs a signal value to a plurality of nodes of the first intermediate layer 1621. Each node receives a signal value from the node of the input layer 161 and calculates the signal value using a parameter, and outputs the calculation result data to a plurality of nodes included in the second intermediate layer 1622. The node included in each intermediate layer receives data from a plurality of nodes in the previous intermediate layer, calculates the received data using parameters, and outputs the data to the node in the later intermediate layer. For example, the node has f (Σ (w * x), where x is the value of the data received from each node in the previous layer, w is the weight corresponding to each node, b is the bias value, and f () is the activation function. ) + B) is performed, and the data of the calculation result is output to a plurality of nodes in the subsequent layer.

The output layer 163 of the learning model 16 has a single node. The plurality of nodes included in the nth intermediate layer 162n output data to the nodes included in the output layer 163. The node of the output layer 163 receives data from a plurality of nodes, calculates the received data using parameters, and outputs state information. For example, the state information is a value of 0 or 1, a value of 0 indicates that the state of the user 4 is a state before performing an operation accompanied by fatigue of the lower limbs, and a value of 1 is a state of the user 4. Indicates that is the state after performing an action accompanied by fatigue of the lower limbs. The correspondence between the value of 0 or 1 and the state of the user 4 may be reversed. The state information may be a value other than 0 or 1. Alternatively, the state information is a value corresponding to the probability that the state of the user 4 is the state before the movement accompanied by the fatigue of the lower limbs, or the probability that the state is the state after the movement accompanied by the fatigue of the lower limbs. May be. For example, the state information takes either a positive or negative value, and the absolute value of the negative value corresponds to the probability that the state of the user 4 is a state before performing an operation accompanied by fatigue of the lower limbs, and is positive. The absolute value of the value of may correspond to the probability that the state of the user 4 is the state before performing the movement accompanied by the fatigue of the lower limbs.

As the learning model 16, a convolutional neural network (CNN) or a recurrent neural network (RNN) may be used as the neural network. The learning model 16 may be in a form in which image data is input as a history of acceleration, angular velocity, and geomagnetism during one walking period. The image data is, for example, the image data of a graph showing the time change of the acceleration, the angular velocity, and the geomagnetic value in the X direction, the Y direction, and the Z direction during one walking period.

The learning of the learning model 16 is performed using a computer. FIG. 8 is a block diagram showing a configuration example of a learning device 6 that learns the learning model 16. The learning device 6 executes the learning model generation method. The learning device 6 is a computer such as a server device. The learning device 6 includes a calculation unit 61, a memory 62, a storage unit 63, an operation unit 64, a display unit 65, and an input unit 66. The arithmetic unit 61 is configured by using, for example, a CPU, a GPU, or a multi-core CPU. The arithmetic unit 61 may be configured by using a quantum computer. The memory 62 is, for example, a RAM. The storage unit 63 is non-volatile, for example, a hard disk. The storage unit 63 stores the computer program 631. The arithmetic unit 61 executes the process according to the computer program 631. The operation unit 64 accepts the input of information such as text by accepting the operation from the user. The operation unit 64 is, for example, a keyboard or a touch panel. The display unit 65 is, for example, a liquid crystal display or an EL display. The input unit 66 is an interface that accepts data input from the outside.

FIG. 9 is a flowchart showing the procedure of the process of generating the learning model 16. Hereinafter, the step is abbreviated as S. The arithmetic unit 61 executes the following processing according to the computer program 631. The calculation unit 61 is attached to the foot of the person and the history of acceleration, angular velocity, and geomagnetism measured during one walking by the sensor device 3 attached to the foot of the person before performing the movement accompanied by the fatigue of the lower limbs. Pre-operation data including the feature amount of myoelectric potential measured by the electromyogram is generated (S11). In S11, the learning device 6 receives data representing the measurement result of the sensor device 3 in the input unit 66, and stores the data in the storage unit 63. The calculation unit 61 performs a process of generating pre-operation data according to an instruction received by the operation unit 64. For example, the calculation unit 61 extracts the history of acceleration, angular velocity, and geomagnetism in one walking period from the history of long-term acceleration, angular velocity, and geomagnetism as shown in FIG.

The learning device 6 receives a feature amount such as an intermediate frequency or an average frequency of the myoelectric potential in the input unit 66, and the calculation unit 61 includes the feature amount of the myoelectric potential in the pre-operation data. Alternatively, the learning device 6 receives the measurement result of the myoelectric potential by the electromyogram, the calculation unit 61 performs frequency analysis, calculates the feature amount of the myoelectric potential, and includes the feature amount of the myoelectric potential in the pre-operation data. You may. The learning device 6 may receive the power spectrum of the myoelectric potential, and the calculation unit 61 may calculate the feature amount of the myoelectric potential and include the feature amount of the myoelectric potential in the pre-operation data. The myoelectric potential features included in the pre-operation data do not show slowing.

The calculation unit 61 generates a plurality of pre-operation data. The calculation unit 61 may generate a plurality of pre-motion data corresponding to a plurality of walks by one subject. The calculation unit 61 may generate a plurality of pre-motion data corresponding to walking by a plurality of subjects. The calculation unit 61 may generate a part of a plurality of pre-operation data by simulation. The calculation unit 61 stores a plurality of pre-operation data in the storage unit 63.

Next, the calculation unit 61 includes a history of acceleration, angular velocity, and geomagnetism measured during one walk by a sensor device 3 attached to a person's foot after performing a specific motion accompanied by fatigue of the lower limbs. Post-operation data including the feature amount of myoelectric potential measured by an electromyogram attached to a human foot is generated (S12). For example, a specific motion is a heel walk with a predetermined number of steps such as 600 steps. Heel walk is walking in a position where the toes do not touch the ground. The heel walk places a heavy burden on the leg muscles such as the tibialis anterior muscle, and after the heel walk, the degree of fatigue increases and the change in walking movement becomes remarkable. Therefore, the differences in acceleration, angular velocity, and geomagnetism before and after the heel walk become remarkable. By using the data obtained from the subjects who performed the heel walk, post-movement data that characterizes the post-motion state with fatigue of the lower limbs can be obtained. The feature amount of myoelectric potential included in the post-operation data indicates the occurrence of slow wave. For example, the intermediate frequency or the average frequency, which is a feature amount, is lower than a predetermined value.

In S12, the learning device 6 receives data representing the measurement result of the sensor device 3 at the input unit 66. The learning device 6 receives the feature amount of the myoelectric potential, the measurement result of the myoelectric potential by the electromyogram, or the power spectrum of the myoelectric potential at the input unit 66. The calculation unit 61 performs a process of generating data after the operation. The calculation unit 61 generates a plurality of post-operation data. The calculation unit 61 stores a plurality of post-operation data in the storage unit 63. As described later, the acceleration, angular velocity, and geomagnetic values included in the pre-operation data and the post-operation data are the acceleration, angular velocity, and geomagnetism in the direction along the axis fixed to the sensor device 3 in the X direction and Y. It is a value after conversion into acceleration, angular velocity and geomagnetism in the direction and Z direction.

Next, the calculation unit 61 is a state in which the state of the subject whose acceleration, angular velocity, geomagnetism, and myoelectric potential are measured with respect to the pre-motion data, the post-motion data, and each data is the state before the motion accompanied by the fatigue of the lower limbs. Processing for generating the learning model 16 is performed using the state information indicating whether the state is the later state as training data (S13). In S13, the calculation unit 61 inputs the acceleration, angular velocity, geomagnetic history, and myoelectric potential features included in the pre-operation data or the post-operation data to the input layer of the learning model 16. The calculation unit 61 associates the pre-motion data with state information indicating that the state is before the motion accompanied by fatigue of the lower limbs, and the post-motion data is the state after the motion accompanied by fatigue of the lower limbs. Associate the status information indicating that. The learning model 16 outputs state information from the node of the output layer 163. The calculation unit 61 calculates the error of the state information by the error function having the state information associated with the input pre-operation data or the post-operation data and the state information output from the node of the output layer 163 as variables, and makes an error. Adjust the calculation parameters of each node of the learning model 16 so that That is, when the pre-motion data is input, state information close to the information indicating that the motion is before the motion with fatigue of the lower limbs is output, and when the post-motion data is input, the motion with fatigue of the lower limbs is output. The parameters are adjusted so that state information close to the information indicating that the state has been performed is output. For example, the arithmetic unit 61 adjusts the parameters by the error back propagation method. The arithmetic unit 61 may adjust the parameters by a learning algorithm other than the error back propagation method.

The calculation unit 61 repeats the process using the plurality of pre-operation data and the plurality of post-operation data, and adjusts the parameters of each node of the learning model 16 to perform machine learning of the learning model 16. In performing machine learning, the arithmetic unit 61 probabilistically inactivates the nodes related to the feature amount of myoelectric potential, repeats the process, and adjusts the parameters of each node. For example, the calculation unit 61 inactivates the node related to the feature amount of the myoelectric potential with a predetermined probability. The node related to the feature amount of myoelectric potential is a node in which the feature amount of myoelectric potential is input or a node that performs calculation using the feature amount of myoelectric potential. The inactivation of a node means that the node is not used, and in the state where a certain node is inactivated, the arithmetic unit 61 performs a process of learning a neural network that does not include the node. Each time the process is repeated, the inactivated node is stochastically changed. The probability does not have to be fixed. The calculation unit 61 stores the learned data in which the adjusted final parameters are recorded in the storage unit 63. In this way, the trained learning model 16 is generated. By using the pre-operation data and the post-operation data as training data, it is possible to generate an appropriately trained learning model 16. After the end of S13, the arithmetic unit 61 ends the process. The learning model 16 included in the fatigue state determination device 1 is manufactured based on the trained data. For example, the learning model 16 is manufactured by writing the parameters recorded in the learned data to the storage unit 13.

Next, the process executed by the fatigue state determination system 100 will be described. FIG. 10 is a flowchart showing a procedure of a process in which the fatigue state determination system 100 acquires acceleration, angular velocity, and geomagnetism. The sensor device 3 mounted on the foot of the user 4 measures the acceleration, the angular velocity, and the geomagnetism by the acceleration sensor 32, the angular velocity sensor 33, and the geomagnetic sensor 34 (S21). The acceleration sensor 32, the angular velocity sensor 33, and the geomagnetic sensor 34 measure the acceleration, the angular velocity, and the geomagnetism in the directions along the three axes perpendicular to each other fixed to the sensor device 3. The directions along the x-axis, y-axis, and z-axis fixed to the sensor device 3 are the x-direction, y-direction, and z-direction. Generally, the x-direction, y-direction, and z-direction are different from the X-direction, Y-direction, and Z-direction. Since the sensor device 3 moves when the user 4 walks, the x-direction, the y-direction, and the z-direction change with respect to the ground.

Each time the acceleration, the angular velocity and the geomagnetism are measured, the sensor device 3 transmits the acceleration, the angular velocity and the geomagnetism to the fatigue state determination device 1 via the terminal device 2 (S22). In S22, the control unit 31 causes the communication unit 35 to transmit data including the values of the acceleration, the angular velocity, and the geomagnetism to the terminal device 2 each time the acceleration, the angular velocity, and the geomagnetism are measured. The terminal device 2 receives the data in the communication unit 26, and the calculation unit 21 causes the communication unit 26 to transmit the received data to the fatigue state determination device 1. The fatigue state determination device 1 receives data in the communication unit 15, and the calculation unit 11 stores the acceleration, angular velocity, and geomagnetic values included in the received data in the storage unit 13 (S23). The storage unit 13 stores repeatedly measured acceleration, angular velocity, and geomagnetic values.

Next, the calculation unit 11 calculates the attitude angles of the x-axis, y-axis, and z-axis with respect to the X-axis, Y-axis, and Z-axis with respect to the received acceleration, angular velocity, and geomagnetism (S24). The Z-axis is an axis along the vertical direction, and the X-axis and the Y-axis are two axes orthogonal to each other included in a plane orthogonal to the Z-axis. The X-axis, Y-axis and Z-axis are fixed to the ground. For example, the Z direction along the Z axis is the vertical direction, the X direction along the X axis is the east-west direction, and the Y direction along the Y axis is the north-south direction. Alternatively, the X-axis and the Y-axis may be in a fixed direction with respect to the user's body. For example, the X direction along the X axis may be the left-right direction with respect to the user 4, and the Y direction along the Y axis may be the front-back direction with respect to the user 4. The posture angle indicates how much the x-axis, y-axis, and z-axis are tilted with respect to the X-axis, Y-axis, and Z-axis. The posture angle is expressed in three dimensions.

In S24, the calculation unit 11 determines whether or not the sensor device 3 is stationary. For example, when the latest predetermined number of stored acceleration, angular velocity, and geomagnetic values have not changed, the calculation unit 11 determines that the sensor device 3 is stationary. When the sensor device 3 is stationary, the calculation unit 11 specifies the Z direction assuming that the direction of the measured acceleration is the direction of gravity, and the X direction and the orthogonal X directions included in the plane orthogonal to the Z direction are included. Specify the Y direction. The arithmetic unit 11 may use the geomagnetism when specifying the X direction and the Y direction. When the sensor device 3 is not stationary, the calculation unit 11 calculates the attitude angle by utilizing the fact that the integral of the angular velocity is an angle. The calculation is based on the acceleration, angular velocity and geomagnetic history stored so far. More specifically, the calculation unit 11 calculates the attitude angle by inputting acceleration, angular velocity and geomagnetism to the Madgwick filter and performing an operation to acquire the output of the Madgwick filter. The calculation unit 11 may calculate the posture angle by a method other than the method using the Madgwick filter. The calculation unit 11 may calculate the attitude angle by using only a part of the information in the acceleration, the angular velocity, and the geomagnetism, such as calculating the attitude angle based on the history of the angular velocity.

The calculation unit 11 converts acceleration, angular velocity, and geomagnetism into acceleration, angular velocity, and geomagnetism in the X, Y, and Z directions based on the attitude angle (S25). The acceleration, angular velocity and geomagnetism measured by the sensor device 3 are acceleration in the x-direction, y-direction and z-direction, angular velocity and geomagnetism. Since the attitude angles in the x, y, and z directions with respect to the X, Y, and Z directions have been obtained, the acceleration, angular velocity, and geomagnetism in the x, y, and z directions are measured in the X direction based on the attitude angles. , Y and Z directions of acceleration, angular velocity and geomagnetism can be converted. The acceleration, angular velocity and geomagnetism measured by the sensor device 3 are corrected by the conversion to X-direction, Y-direction and Z-direction acceleration, angular velocity and geomagnetism. By obtaining the acceleration, angular velocity and geomagnetism in the X, Y and Z directions, the influence of the movement of the sensor device 3 on the acceleration, angular velocity and geomagnetism is removed, and the user 4 walks on the acceleration, angular velocity and geomagnetism. The effect of is clarified.

The calculation unit 11 stores the converted acceleration, angular velocity, and geomagnetism in the storage unit 13 (S25). After S25 is completed, the calculation unit 11 ends the process of acquiring acceleration, angular velocity, and geomagnetism. The storage unit 13 stores the history of acceleration, angular velocity, and geomagnetism in the X, Y, and Z directions.

The processes of S21 to S26 are repeatedly executed each time the sensor device 3 measures acceleration, angular velocity, and geomagnetism. The sensor device 3 does not transmit data each time the acceleration, angular velocity, and geomagnetism are measured in S22, but transmits data summarizing the results of measuring the acceleration, angular velocity, and geomagnetism a plurality of times. May be good. In this embodiment, the processes S22 to S26 are executed for a plurality of measurement results of acceleration, angular velocity, and geomagnetism transmitted collectively.

The fatigue state determination system 100 may be in a form in which the processes of S23 to S25 are executed by the terminal device 2. In this embodiment, the terminal device 2 receives data including acceleration, angular velocity, and geomagnetic values in the communication unit 26, and the calculation unit 21 stores the acceleration, angular velocity, and geomagnetic values in the storage unit 23, and S24 and S25. Executes the processing of. After that, the calculation unit 21 causes the communication unit 26 to transmit data including the values of acceleration, angular velocity, and geomagnetism in the X, Y, and Z directions to the fatigue state determination device 1. The fatigue state determination device 1 receives data in the communication unit 15, and the calculation unit 11 executes the process of S26. Alternatively, the fatigue state determination system 100 may be in a form of transmitting data from the sensor device 3 to the fatigue state determination device 1 without going through the terminal device 2.

The fatigue state determination system 100 may be in a form in which the processing up to S25 is executed by the sensor device 3. In this embodiment, the sensor device 3 does not execute the processing of S22, but executes the processing of S21 and S23 to S25 by itself. After that, the sensor device 3 transmits data to the fatigue state determination device 1 without going through the terminal device 2. The sensor device 3 may transmit data to the fatigue state determination device 1 without going through the terminal device 2.

In a state where the history of acceleration, angular velocity, and geomagnetism in the X, Y, and Z directions is stored in the storage unit 13, the fatigue state determination device 1 performs an operation in which the state of the user 4 is accompanied by fatigue of the lower limbs. It is determined whether it is in the state of or after. FIG. 11 is a flowchart showing a procedure of determination processing executed by the fatigue state determination device 1. The calculation unit 11 extracts the history of acceleration, angular velocity, and geomagnetism during one walking period from the history of acceleration, angular velocity, and geomagnetism stored in the storage unit 13 (S31).

As described above with reference to FIG. 5, the time point at which the foot touches the ground is specified based on the time change of the acceleration in the vertical direction and the time change of the angular velocity in the direction corresponding to the shaking of the foot in the front-back direction of the user 4. can do. In S31, the calculation unit 11 specifies a time point when the foot of the user 4 touches the ground, for example, based on the time change of the acceleration in the Z direction and the angular velocity in the X direction. The calculation unit 11 sets the history of acceleration, angular velocity, and geomagnetism included in the period from the time when the foot touches the ground to the time when the foot touches the ground from the stored acceleration, angular velocity, and geomagnetic history for one walking period. It is extracted as a history of acceleration, angular velocity and geomagnetism measured in. At this time, the calculation unit 11 extracts the history of acceleration, angular velocity, and geomagnetism in each of the X direction, the Y direction, and the Z direction. The arithmetic unit 11 may specify three or more time points when the foot touches the ground, and may extract the history of acceleration, angular velocity, and geomagnetism for each of the plurality of walking periods. The method of specifying the time of touchdown is not limited to the method of using the acceleration in the Z direction and the angular velocity in the X direction. The arithmetic unit 11 may use other accelerations or angular velocities to specify the time of touchdown, or may use three or more accelerations or angular velocities.

Next, the calculation unit 11 inputs the history of acceleration, angular velocity, and geomagnetism in the X, Y, and Z directions during one walking period into the learning model 16 (S32). In S32, the calculation unit 11 inputs the history of acceleration, angular velocity, and geomagnetism to the learning model 16, while causing the learning model 16 to perform the calculation without inputting the feature amount of the myoelectric potential. For example, among the nodes of the learning model 16, the calculation unit 11 includes a node in which the feature amount of the myoelectric potential is input and a node that performs the calculation using the feature amount of the myoelectric potential without using acceleration, angular velocity, and geomagnetism. Inactivates. As a result, the learning model 16 functions to output state information when the history of acceleration, angular velocity, and geomagnetism is input without inputting the feature amount of myoelectric potential. As described above, the learning model 16 in which the history of acceleration, angular velocity, and geomagnetism is input performs the calculation of the neural network, and the state of the user 4 is the state before or after the operation accompanied by the fatigue of the lower limbs. Outputs state information indicating whether it is a state. Since the influence of the user 4's walking clearly appears on the acceleration, angular velocity and geomagnetism in the X, Y and Z directions, accurate use of the acceleration, angular velocity and geomagnetism in the X, Y and Z directions State information can be obtained.

When the history of acceleration, angular velocity and geomagnetism is extracted for each of a plurality of walking periods, the history of acceleration, angular velocity and geomagnetism is input to the learning model 16 and the state information is output from the learning model 16. Is performed individually for each of the periods of multiple walks. State information is output for each of a plurality of one-walking periods, and a plurality of state information can be obtained. Since the duration of one walk is not constant, the number of acceleration, angular velocity and geomagnetic values measured during the duration of one walk is not constant. The number of acceleration, angular velocity, and geomagnetic values input to the learning model 16 does not have to be constant. The calculation unit 11 may adjust the number of acceleration, angular velocity, and geomagnetic values input to the learning model 16 to be constant by thinning out or complementing. Alternatively, the arithmetic unit may average the history of acceleration, angular velocity, and geomagnetism during a plurality of walking periods, and input the averaged acceleration, angular velocity, and geomagnetic history into the learning model 16.

Next, the calculation unit 11 determines whether the state of the user 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs, according to the state information output by the learning model 16 (. S33). In S33, the calculation unit 11 makes a determination according to the value of the state information. In the form in which the state information represents a probability, in S33, the calculation unit 11 makes a determination according to the probability. For example, the state information is a value corresponding to the probability that the state of the user 4 is a state after performing an operation accompanied by fatigue of the lower limbs. When the value of the state information is less than the predetermined value, the calculation unit 11 determines that the state of the user 4 is the state before performing the operation accompanied by the fatigue of the lower limbs, and the value of the state information is equal to or more than the predetermined value. In a certain case, it is determined that the state of the user 4 is a state after performing an operation accompanied by fatigue of the lower limbs. Based on the state information, the calculation unit 11 determines whether the state of the user 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs without using the myoelectric potential. In addition, based on the state information, the calculation unit 11 has a probability that the state of the user 4 is a state before performing an operation accompanied by fatigue of the lower limbs, or after the state of the user 4 performs an operation accompanied by fatigue of the lower limbs. The probability of being in the state of may be determined.

When a plurality of state information is obtained, the calculation unit 11 makes a determination based on the plurality of state information. For example, the calculation unit 11 averages a plurality of state information and makes a determination based on the average value of the state information. For example, the calculation unit 11 makes a determination based on each of the plurality of state information, and makes a final determination by adopting a determination result that occupies a majority of the plurality of determination results. By using a plurality of state information, more accurate judgment becomes possible.

Next, the calculation unit 11 outputs the determination result (S34). In S34, the calculation unit 11 causes the communication unit 15 to transmit information indicating the determination result to the terminal device 2. The terminal device 2 receives the information transmitted from the fatigue state determination device 1 in the communication unit 26. The calculation unit 21 causes the display unit 25 to display the determination result indicated by the received information. The user 4 can confirm the determination result by using the terminal device 2. In S34, it may be output whether the state of the user 4 is the state before or after the operation accompanied by the fatigue of the lower limbs, and the state of the user 4 performs the operation accompanied by the fatigue of the lower limbs. The probability of being in the previous state or the probability that the state of the user 4 is in the state after performing an operation accompanied by fatigue of the lower limbs may be output. After the end of S34, the fatigue state determination device 1 ends the process of determining whether the state of the user 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs.

As described in detail above, the user 4 walks with the sensor device 3 attached to the foot, and the sensor device 3 measures the acceleration and angular velocity generated by walking, and the geomagnetism. The fatigue state determination device 1 uses the learning model 16 to determine whether the state of the user 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs from the history of acceleration, angular velocity, and geomagnetism. judge. After the motion with fatigue of the lower limbs, the walking motion changes due to fatigue, and the acceleration, angular velocity, and geomagnetism measured by the sensor device 3 change as compared with before the motion with fatigue of the lower limbs. Therefore, by using the appropriate learning model 16, the state of the user 4 is before or after the operation accompanied by the fatigue of the lower limbs without the user 4 continuing to wear the EMG for a long time. It is possible to easily determine whether or not the state is.

Not limited to specific movements such as heel walks, if the movement is strong enough to cause fatigue in the lower limbs, such as slowing of the myoelectric potential of the lower limbs, when the user 4 performs the movement. Is determined to be the state after the operation is performed. Therefore, no matter what kind of movement the user 4 performs in daily life, if the movement is accompanied by fatigue of the lower limbs, the state of the user 4 is performed with the fatigue of the lower limbs. It is determined that the state is after the operation. The result of determining whether the state of the user 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs can be useful for estimating the fatigue of the lower limbs of the user 4, and further, use. It can be useful for estimating the daily fatigue of person 4.

In this embodiment, an example in which the learning model 16 is a neural network is shown, but the fatigue state determination system 100 uses a model other than the neural network, such as a support vector machine or a random forest, as the learning model 16. May be. In the present embodiment, a heel walk of a predetermined number of steps is exemplified as a specific motion accompanied by fatigue of the lower limbs for creating post-motion data, but a specific motion for creating post-motion data causes fatigue of the lower limbs. If it is an accompanying operation, it may be another operation. The specific movement may be a heel walk that is continued for a predetermined time, or may be a walk other than the heel walk, such as a normal walk, a toe walk, or a fast walk. The specific movement may be walking of a predetermined number of steps that does not limit the walking method, or walking that is continued for a predetermined time. Specific movements may include movements other than walking, such as running or movements in daily life.

<Embodiment 2>
In the first embodiment, the learning model 16 is learned by supervised learning. In the second embodiment, the learning model 16 is an abnormality detection model that is learned by unsupervised learning and outputs the degree of abnormality of the state of the user 4 as state information. The degree of abnormality is a value indicating the degree of abnormality in the state of the user 4, assuming that the state before performing an operation accompanied by fatigue of the lower limbs is a normal state. When the degree of abnormality is large, it is determined that the state of the user 4 is an abnormal state, that is, a state after performing an operation accompanied by fatigue of the lower limbs. An example of using DAGMM (Deep Autoencoding Gaussian Mixture Model) as an abnormality detection model is shown. DAGMM is described in the following literature.

Bo Zong, Qi Song, Martin Renqiang Min, Wei Cheng, Cristian Lumezanu, Daeki Cho, and Haifeng Chen, "Deep Autoencoding Gaussian Mixture Model for Unsupervised Anomaly Detection", International Conference on Learning Representations, 2018.

FIG. 12 is a conceptual diagram showing an outline of DAGMM. It is assumed that the state of the user 4 is represented by the vector z. For the vector z, the probability of belonging to each of the K clusters represented by GMM (Gaussian mixture model) is obtained. The history of acceleration, angular velocity and geomagnetism in the X, Y and Z directions during a walking period is represented by the vector x. The vector x is input to the compression network 71. The compression network 71 is an autoencoder. The compression network 71 outputs the reconstruction vector x'in response to the input of the vector x. The reconstructed vector x'is a reconstructed vector x by a self-encoder. Further, the compression network 71 outputs the latent variable z _c . The latent variable z _c is a low-dimensional feature representation of the vector x.

From the vector x and the reconstruction vector x', the feature amount _zr derived from the reconstruction error is calculated. Let f (x, x') be the error function between the vector x and the reconstruction vector x', and z _r = f (x, x'). The vector z representing the state of the user 4 is obtained by combining the latent variable z _c and the feature quantity z _r . The vector z is represented by z = [z _c , z _r ]. The vector z is input to the estimation network 72. The estimation network 72 is a neural network. The estimation network 72 outputs a vector γ indicating the probability that the vector z belongs to each cluster in response to the input of the vector z. The vector γ has K components. Assuming that k is a natural number of 1 or more and K or less, the component γ _k included in the vector γ is the probability that the vector z belongs to the kth cluster. The vector γ is obtained as the output of the softmax function.

One vector x is obtained from one pre-operation data. The pre-motion data consists of a history of acceleration, angular velocity, and geomagnetism measured during one walk by a sensor device 3 attached to a person's foot before performing a motion accompanied by fatigue of the lower limbs. A plurality of vectors x are obtained from the plurality of pre-operation data, and a vector γ is obtained for each vector x. Let N be the number of vectors x. Let i be a natural number of 1 or more and N or less, let z _i be the i-th vector z, and let γ _ik be the probability that the vector z _i belongs to the k-th cluster. The mixing ratio φ _k of the k-th cluster is expressed by the following equation (1), the average μ _k of the k-th cluster is expressed by the following equation (2), and the covariance matrix Σ _k of the k-th cluster is It is expressed by the following equation (3).

Here, the negative log-likelihood E (z) can be calculated by the following equation (4). The negative log-likelihood E (z) indicates the degree to which the vector z is separated from the K clusters.

Let x _i be the i-th vector x, and let x'i be the _i -th reconstruction vector x'. The objective function J of DAGMM is expressed by the following equation (5).

L (x _i , _x'i ) included in equation (5) is a loss function that characterizes the reconstruction error. For example, using the L2 norm, L (x _i , _x'i ) = || x _{i --x'i} || ₂ ² . The dimension of the vector z. (Σ _k ) _jj is the jth diagonal element of the covariance matrix Σ _k . | ・ | Is a determinant. λ ₁ and λ ₂ are constants, for example, λ ₁ = 0.1 and λ ₂ = 0.005. The compression network 71 and the estimation network 72 are optimized by adjusting the parameters of the compression network 71 and the estimation network 72 so as to minimize the objective function J. Further, the hyperparameters of the compressed network 71 and the estimated network 72 can be optimized by a method such as grid search or Bayesian optimization.

The learning device 6 learns the compression network 71 and the estimation network 72 using the plurality of pre-operation data. The arithmetic unit 11 performs a process of generating a learning model 16 by learning the compression network 71 and the estimation network 72. In S11, the calculation unit 11 generates a plurality of pre-operation data, and stores the plurality of pre-operation data in the storage unit 63. The arithmetic unit 11 does not execute the process of S12. In S13, in order to generate the learning model 16, the arithmetic unit 11 performs unsupervised learning of the compression network 71 and the estimation network 72 using a plurality of pre-operation data as training data. The arithmetic unit 11 generates a plurality of vectors x from a plurality of pre-operation data, and uses the compression network 71 and the estimation network 72 to obtain the plurality of vectors z, the mixing ratio φ _k of each cluster, and the average μ _k of each cluster. , The covariance matrix Σ _k of each cluster, and the objective function J are calculated. Then, the arithmetic unit 11 adjusts the parameters of the compression network 71 and the estimation network 72 so that the objective function J becomes as small as possible.

The calculation unit 61 stores the learned data in which the adjusted final parameters are recorded in the storage unit 63. The trained data includes at least the parameters of the trained compressed network 71. In addition, the trained data includes the mixing ratio φ _k of each cluster after training, the average μ _k of each cluster, and the covariance matrix Σ _k of each cluster. The learning model 16 included in the fatigue state determination device 1 is manufactured based on the trained data. For example, the learning model 16 is manufactured by writing the parameters recorded in the learned data to the storage unit 13.

FIG. 13 is a conceptual diagram showing a configuration example of the learning model 16 according to the second embodiment. The learning model 16 has a compression network 164. The compression network 164 is a self-encoder equivalent to the learned compression network 71. The vector x representing the history of acceleration, angular velocity and geomagnetism in the X, Y and Z directions during one walking period is input to the compression network 164. The compression network 164 outputs the reconstruction vector x'and the latent variable z _c in response to the input of the vector x. The learning model 16 has an error calculation unit 165. The error calculation unit 165 calculates the feature amount _zr derived from the reconstruction error from the vector x and the reconstruction vector x'. The vector z is obtained from the latent variable z _c and the feature amount z _r .

The learning model 16 has an abnormality degree calculation unit 166. The anomaly degree calculation unit 166 uses the vector z and the mixing ratio φ _k , the average μ _k , and the covariance matrix Σ _k obtained by learning, and uses the negative log-likelihood E (4) as the anomaly degree. z) is calculated. The learning model 16 outputs the calculated abnormality degree E (z) as state information. The mixing ratio φ _k , the average μ _k , and the covariance matrix Σ _k are stored in the storage unit 13.

When the fatigue state determination device 1 executes the determination process, the calculation unit 11 uses S32 to obtain a vector x composed of accelerations, angular velocities, and geomagnetic histories in the X, Y, and Z directions during one walking period. , Input to the learning model 16. The learning model 16 to which the vector x is input calculates the vector z by performing the calculation of the compression network 164 and the error calculation unit 165. Further, the learning model 16 performs the calculation of the abnormality degree calculation unit 166 and outputs the abnormality degree E (z) as the state information. The degree of abnormality E (z) represents how far the state of the user 4 is from the normal state. The larger the value of the state information, the higher the probability that the state of the user 4 is an abnormal state. It is possible to determine that the abnormal state is a state after performing an action accompanied by fatigue of the lower limbs.

When a plurality of vectors x are obtained, the arithmetic unit 11 individually executes the input of the vector x into the learning model 16 and the output of the state information from the learning model 16 for each of the plurality of vectors x. State information is output for each of the plurality of vectors x, and a plurality of state information can be obtained.

In S33, when the value of the state information is equal to or less than a predetermined threshold value, the calculation unit 11 determines that the state of the user 4 is a state before performing an operation accompanied by fatigue of the lower limbs, and the value of the state information is set. When the predetermined threshold value is exceeded, it is determined that the state of the user 4 is the state after performing an operation accompanied by fatigue of the lower limbs. For example, the threshold is set based on a plurality of pre-operation data for testing that is different from the training data. A threshold value is set so that a plurality of vectors x obtained from a plurality of pre-operation data are input to the learning model 16, a plurality of state information is output from the learning model 16, and most of the plurality of state information is equal to or less than the threshold value. Will be done. For example, the threshold value is set so that 95% of the plurality of state information is equal to or less than the threshold value. The threshold value is stored in the storage unit 13. Alternatively, the threshold may be set so that most of the state information obtained from the plurality of post-operation data for testing exceeds the threshold. When the value of the state information is less than the predetermined threshold value, the calculation unit 11 determines that the state of the user 4 is the state before performing the operation accompanied by the fatigue of the lower limbs, and the value of the state information is equal to or more than the predetermined threshold value. If this is the case, it may be determined that the state of the user 4 is the state after performing an operation accompanied by fatigue of the lower limbs. When a plurality of state information is obtained, the calculation unit 11 makes a determination based on the plurality of state information.

The learning model 16 does not use the pre-stored mixing ratio φ _k , average μ _k , and covariance matrix Σ _k when calculating the anomaly degree E (z), but rather the anomaly degree E (z). It may be in the form of calculating the mixing ratio φ _k , the average μ _k , and the covariance matrix Σ _k each time the calculation is performed. In this form, the training model 16 includes an estimated network, which calculates the vector γ from the vector z using the estimated network, and obtains the mixing ratio φ _k , the mean μ _k , and the covariance matrix Σ _k from the vector γ and the training data. Calculate and calculate the degree of anomaly E (z).

Also in the second embodiment, the learning model 16 appropriately outputs the state information by performing appropriate learning. Also in the third embodiment, by using the appropriate learning model 16, the state of the user 4 before the user 4 performs an operation accompanied by fatigue of the lower limbs without the user 4 continuing to wear the EMG for a long time. It is possible to easily determine whether the state is the same or the later state. Since the learning model 16 is learned by unsupervised learning using the pre-motion data as training data, the learning model 16 can be generated without using the post-motion data. It is not necessary to prepare post-operation data as training data, and it is not necessary to impose a burden on the subject and create post-operation data. Therefore, the training data can be easily created and the learning model 16 can be generated. In the second embodiment, an example in which DAGMM is used as an abnormality detection model is shown, but the learning model 16 may be another abnormality detection model.

<Embodiment 3>
In embodiments 1 and 2, accelerations in the x, y and z directions, angular velocities and geomagnetism are converted into accelerations in the X, Y and Z directions, angular velocities and geomagnetism in the X, Y and Z directions. The acceleration, angular velocity and geomagnetism were input to the learning model 16. In the third embodiment, the learning model 16 is input with accelerations, angular velocities and geomagnetism in the x-direction, y-direction and z-direction. FIG. 14 is a conceptual diagram showing the function of the learning model 16 according to the third embodiment. The learning model 16 includes a history of accelerations in the x-direction, y-direction, and z-direction during one walk, a history of angular velocities in the x-direction, y-direction, and z-direction during one walk, and x in the period of one walk. The history of geomagnetism in the directions, y-directions and z-directions, the posture angles in the x-directions, y-directions and z-directions with respect to the X-direction, Y-direction and Z-direction, and the feature amount of the myoelectric potential of the lower limbs are input.

The learning model 16 is used when the history of acceleration, angular velocity and geomagnetism in the x-direction, y-direction and z-direction, the history of the posture angle, and the feature amount of the myoelectric potential of the lower limbs are input during one walking period. It is learned to output state information indicating whether the state of the person 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs. Further, the learning model 16 is learned while stochastically inactivating the nodes related to the feature amount of the myoelectric potential, so that the feature amount of the myoelectric potential of the lower limbs is not input, and the history and posture of acceleration, angular velocity and geomagnetism are not input. It is configured to output information indicating whether the state of the user 4 is a state before or after performing an operation accompanied by fatigue of the lower limbs when the history of the corner is input.

In the third embodiment, the fatigue state determination system 100 does not execute S25 and S26 after executing S24, and the fatigue state determination device 1 stores the posture angle calculated in S24. In S31, the calculation unit 11 extracts the history of the posture angle in one walking period as well as the history of acceleration, angular velocity, and geomagnetism. In S32, the calculation unit 11 inputs the history of the attitude angle to the learning model 16 in addition to the history of acceleration, angular velocity, and geomagnetism. Further, the calculation unit 11 causes the learning model 16 to perform the calculation without inputting the feature amount of the myoelectric potential into the learning model 16. For example, among the nodes of the learning model 16, the calculation unit 11 includes a node in which the feature amount of the myoelectric potential is input and a node that performs the calculation using the feature amount of the myoelectric potential without using acceleration, angular velocity, and geomagnetism. Inactivates. The learning model 16 functions to output state information when the history of acceleration, angular velocity, geomagnetism and the history of posture angle are input without inputting the feature amount of myoelectric potential. Also in the third embodiment, the learning model 16 appropriately outputs the state information by performing appropriate learning. Also in the third embodiment, by using the appropriate learning model 16, the state of the user 4 before the user 4 performs an operation accompanied by fatigue of the lower limbs without the user 4 continuing to wear the EMG for a long time. It is possible to easily determine whether the state is the same or the later state.

<Embodiment 4>
FIG. 15 is a schematic diagram showing a configuration example of the fatigue state determination system 100 according to the fourth embodiment. The fatigue state determination system 100 includes a sensor device 3 and a terminal device 2. The terminal device 2 includes a learning model 16 and functions as a fatigue state determination device. For example, in the fourth embodiment, the learning model 16 is realized by the arithmetic unit 21 executing information processing according to the computer program 231. The terminal device 2 executes the processes of S23 to S26 and the processes of S31 to S34. Also in the fourth embodiment, by using the appropriate learning model 16, the state of the user 4 before the user 4 performs an operation accompanied by fatigue of the lower limbs without the user 4 continuing to wear the EMG for a long time. It is possible to easily determine whether the state is the same or the later state.

The fatigue state determination system 100 may have a form in which the sensor device 3 functions as a fatigue state determination device. In this embodiment, the sensor device 3 executes the processes of S21, S23 to S26, and S31 to S34 without executing S22. In S34, the sensor device 3 transmits the determination result to the terminal device 2.

In the first to fourth embodiments, the x direction, the y direction and the z direction, and the X direction, the Y direction and the Z direction are assumed to be three directions orthogonal to each other, but two intersecting at an angle other than the right angle. The direction may be included. In the first to fourth embodiments, the sensor device 3 is attached to one leg of the user 4, but the fatigue state determination system 100 has the sensor device 3 attached to both legs of the user 4. May be. The fatigue state determination device 1 may determine the state of the user 4 for each of both feet of the user 4, or may determine the state of the user 4 based on both the measurement results of the two sensor devices 3. good.

The present invention is not limited to the contents of the above-described embodiment, and various modifications can be made within the scope of the claims. That is, an embodiment obtained by combining technical means appropriately modified within the scope of the claims is also included in the technical scope of the present invention.

1 Fatigue state judgment device 10 Recording medium 100 Fatigue state judgment system 131 Computer program 16 Learning model 2 Terminal device 3 Sensor device 4 User 5 Communication network 6 Learning device

Claims (12)

Acceleration, angular velocity and geomagnetism during walking are repeatedly measured by sensors attached to human feet.
Learning to output information indicating whether the person's state is a state before or after performing a movement accompanied by fatigue of the lower limbs when the history of acceleration, angular velocity, and geomagnetism during walking of the person is input. Input the history of acceleration, angular velocity and geomagnetism measured by the sensor into the model.
Fatigue state determination, which is characterized by determining whether the state of a person wearing the sensor on his / her foot is a state before or after performing the operation according to the information output by the learning model. Method. The learning model is
When the history of acceleration, angular velocity and geomagnetism during walking of a person and the feature amount of myoelectric potential obtained from the person's foot are input, the state of the person before the movement accompanied by fatigue of the lower limbs is performed. It has been trained to output information indicating whether it is in a state or a later state.
When the feature amount of myoelectric potential obtained from a person's foot is not input and the history of acceleration, angular velocity and geomagnetism during walking of the person is input, the person's state performs an action accompanied by fatigue of the lower limbs. The fatigue state determination method according to claim 1, wherein information indicating whether the state is a previous state or a later state is output. The sensor measures acceleration, angular velocity and geomagnetism in the direction along an axis fixed to the sensor.
The attitude angle of the axis is calculated based on the history of acceleration, angular velocity or geomagnetism.
The fatigue state determination method according to claim 1 or 2, wherein the acceleration, the angular velocity, and the geomagnetism are corrected based on the posture angle. The sensor measures acceleration, angular velocity and geomagnetism in the direction along an axis fixed to the sensor.
At each time when the acceleration, angular velocity and geomagnetism were measured, the attitude angle of the axis fixed to the sensor with respect to the axis fixed to the ground was calculated based on the history of acceleration, angular velocity and geomagnetism.
Based on the attitude angle, the acceleration, angular velocity and geomagnetism in the direction along the axis fixed to the sensor are converted into the acceleration, angular velocity and geomagnetism in the direction along the axis fixed to the ground.
The fatigue state determination method according to claim 1 or 2, wherein the converted acceleration, angular velocity, and geomagnetism are input to the learning model. The learning model is a model that outputs the information when the history of acceleration, angular velocity, and geomagnetism in one walking period with the same foot is input.
From the acceleration, angular velocity and geomagnetism measured repeatedly by the sensor, the history of acceleration, angular velocity and geomagnetism measured during one walking period is extracted.
The fatigue state determination method according to any one of claims 1 to 4, wherein the extracted acceleration, angular velocity, and geomagnetic history are input to the learning model. Based on the changes in acceleration and angular velocity, the time when the foot on which the sensor is attached touches the ground is specified.
The history of acceleration, angular velocity and geomagnetism measured during the period from the time when the foot touches the ground to the time when the foot touches the ground is extracted as the history of acceleration, angular velocity and geomagnetism measured during one walking period. 5. The method for determining a fatigue state according to claim 5. The history of acceleration, angular velocity and geomagnetism measured in each of a plurality of walking periods is individually input to the learning model.
It is determined whether the state of the person wearing the sensor on the foot is the state before or after the operation according to the plurality of information output by the learning model for the period of the plurality of walks. The fatigue state determination method according to claim 5 or 6, wherein the method is to be performed. Acceleration, angular velocity, and geomagnetic history measured during one walk by a sensor attached to a person's foot before performing movements involving lower limb fatigue, and the characteristic amount of myoelectric potential obtained from the person's foot. make,
Acceleration, angular velocity and geomagnetic history measured during one walk by a sensor attached to the person's foot after performing the above movement, and the feature amount of myoelectric potential obtained from the person's foot were created.
It shows the history of acceleration, angular velocity, and geomagnetism measured in each of a plurality of walking periods, the feature amount of myoelectric potential, and whether the human state is a state before or after the above movement. Using the information as training data, when the history of acceleration, angular velocity, and geomagnetism during walking of a person is input, information indicating whether the state of the person is a state before or after the operation is output. A learning model generation method characterized by generating a learning model. Creates a history of acceleration, angular velocity and geomagnetism measured during one walk by a sensor attached to a person's foot before performing a motion with fatigue of the lower limbs.
When the history of acceleration, angular velocity and geomagnetism during walking of a person is input by unsupervised learning using the history of acceleration, angular velocity and geomagnetism measured in each of a plurality of walking periods as training data, the said person A learning model characterized by generating a learning model that outputs information indicating the degree of abnormality of the person's state as information indicating whether the state is a state before or after the operation. Generation method. A sensor that is attached to a person's foot and repeatedly measures acceleration, angular velocity, and geomagnetism while the person is walking.
It is provided with a fatigue state determination device for determining whether the state of the person wearing the sensor on the foot is a state before or after performing an operation accompanied by fatigue of the lower limbs.
The fatigue state determination device is
It is equipped with a learning model that outputs information indicating whether the state of the person is a state before or after the operation when the history of acceleration, angular velocity, and geomagnetism during walking of the person is input.
Acquires the acceleration, angular velocity and geomagnetism measured by the sensor,
The state of the person wearing the sensor on his / her foot is the state before the operation according to the information output by the learning model when the acquired acceleration, angular velocity, and geomagnetic history are input to the learning model. A fatigue state determination system characterized by determining whether the condition is after or after. Learning to output information indicating whether the person's state is a state before or after performing a movement accompanied by fatigue of the lower limbs when the history of acceleration, angular velocity, and geomagnetism during walking of the person is input. Equipped with a model,
Acquires acceleration, angular velocity and geomagnetism during walking measured by sensors attached to human feet,
The state of the person wearing the sensor on his / her foot is the state before the operation according to the information output by the learning model when the acquired acceleration, angular velocity, and geomagnetic history are input to the learning model. A fatigue state determination device characterized by determining whether the state is after or after. Learning to output information indicating whether the person's state is a state before or after performing a movement accompanied by fatigue of the lower limbs when the history of acceleration, angular velocity, and geomagnetism during walking of the person is input. Enter the history of acceleration, angular velocity and geomagnetism during walking measured by a sensor attached to the human foot into the model.
According to the information output by the learning model, the computer is made to execute a process of determining whether the state of the person wearing the sensor on the foot is the state before or after the operation. Characterized computer program.

Images