JP6998518B2 - Motor function estimation information generation device, motor function estimation system, motor function estimation information generation method, motor function estimation method and recording medium - Google Patents

Description

The present disclosure relates to a motor function estimation information generator for estimating motor function, a motor function estimation system, a motor function estimation information generation method, a motor function estimation method, and a recording medium.

In particular, information on motor functions such as walking function is important in determining the rehabilitation content of the user. In recent years, falls due to sarcopenia or deterioration of walking function have become a social problem, and there is an increasing demand for simple and accurate evaluation of the walking ability of users. Walking ability is broadly classified into three abilities: muscle strength, endurance, and balance ability. In order to accurately measure these abilities in physiotherapy, a constant velocity muscular strength measuring device, a cardiopulmonary load test and a body sway meter are used, respectively. Although these devices and tests are accurate, they require large-scale equipment and place a heavy load on the subject, so that they are not suitable for evaluation of daily motor function. To solve this problem, for example, in the technique described in Patent Document 1, the average acceleration in the vertical and anteroposterior directions of an acceleration sensor mounted on the user's waist is calculated as a feature amount, and the lower limb muscle strength is calculated based on the calculated feature amount. I'm estimating. Further, Non-Patent Document 1 describes a technique for extracting features for behavior recognition.

Japanese Patent No. 4696677

Thomas Plotz et al., "Fature Learning for Activity Recognition in Ubiquitous Computing", International Joint Conference on Artificial Intelligence, page 20C

However, in the technique of Patent Document 1, the average value of acceleration, the maximum acceleration value, and the minimum acceleration are used as features, and the characteristics of the measured acceleration waveform are not sufficiently captured. As a result, the technique of Patent Document 1 has a problem that high estimation accuracy of lower limb muscle strength cannot be obtained. There is also a method of generating features for behavior recognition by machine learning, such as the technique of Non-Patent Document 1, but such a conventional technique captures features in short sensor data of about several seconds, and is a number. It is difficult to automatically calculate valid features from long-term sensor data ranging from minutes to hours or more.

Non-limiting and exemplary embodiments of the present disclosure include motor function estimation information generators, motor function estimation systems, motor function estimation information generation methods, and motor function estimation that improve the accuracy of motor functions estimated from long-term sensor data. Method and recording medium.

The motor function estimation information generator according to one aspect of the present disclosure is
A sensor that measures at least one selected from the user's acceleration, heart rate, body temperature, and angular velocity at a given time.
Equipped with a first processing circuit
The first processing circuit is
(A1) At least one sensor value selected from the acceleration, heart rate, body temperature, and angular velocity of the user and the exercise ability value of the user are acquired.
(A2) The characteristics of the sensor value and the time interval within the predetermined time are determined.
(A3) Using the sensor value, a feature vector corresponding to the feature amount of the feature in the time interval is calculated.
(A4) Using the feature vector and the athletic ability value, a first weight vector for estimating the athletic ability value is obtained.
(A5) Using the first weight vector and the athletic ability value, a gradient vector with respect to the feature vector is calculated.
(A6) A new time interval within the predetermined time interval and a new feature based on the new time interval are determined.
(A7) Using the sensor value, a feature candidate vector that is a candidate for the feature vector corresponding to the feature amount of the new feature in the new time interval is calculated.
(A8) Based on the difference between the feature candidate vector and the feature vector, the feature candidate vector satisfying a predetermined condition based on the gradient vector is determined.
(A9) Using the feature candidate vector satisfying the predetermined condition, the first weight vector is modified to a second weight vector.
(A10) The feature and the time interval corresponding to the feature candidate vector satisfying the predetermined condition, and the second weight vector are stored in the storage unit.

The motor function estimation system according to one aspect of the present disclosure is
The motor function estimation information generator and
Equipped with a second processing circuit
The second processing circuit is
Acquires at least one sensor value selected from the user's acceleration, heart rate, body temperature, and angular velocity,
An estimated feature vector is calculated using the feature and time interval stored in the storage unit and the sensor value.
The athletic ability value is estimated by using the second weight vector stored in the storage unit and the estimated feature vector.
The athletic ability value is output.

The method for generating motor function estimation information according to one aspect of the present disclosure is
(B1) At least one sensor value selected from the user's acceleration, heart rate, body temperature, and angular velocity, and the user's athletic ability value are acquired.
(B2) The characteristics of the sensor value and the time interval within a predetermined time are determined.
(B3) Using the sensor value, a feature vector corresponding to the feature amount of the feature in the time interval is calculated.
(B4) Using the feature vector and the athletic ability value, a first weight vector for estimating the athletic ability value is obtained.
(B5) Using the first weight vector and the athletic ability value, a gradient vector with respect to the feature vector is calculated.
(B6) A new time interval within the predetermined time interval and a new feature based on the new time interval are determined.
(B7) Using the sensor value, a feature candidate vector that is a candidate for the feature vector corresponding to the feature amount of the new feature in the new time interval is calculated.
(B8) Based on the difference between the feature candidate vector and the feature vector, the feature candidate vector satisfying a predetermined condition based on the gradient vector is determined.
(B9) Using the feature candidate vector satisfying the predetermined condition, the first weight vector is modified to a second weight vector.
(B10) The feature and the time interval corresponding to the feature candidate vector satisfying the predetermined condition, and the second weight vector are stored in the storage unit.

The motor function estimation method according to one aspect of the present disclosure is
(B1) At least one sensor value selected from the user's acceleration, heart rate, body temperature, and angular velocity, and the user's athletic ability value are acquired.
(B2) The characteristics of the sensor value and the time interval within a predetermined time are determined.
(B3) Using the sensor value, a feature vector corresponding to the feature amount of the feature in the time interval is calculated.
(B4) Using the feature vector and the athletic ability value, a first weight vector for estimating the athletic ability value is obtained.
(B5) Using the first weight vector and the athletic ability value, a gradient vector with respect to the feature vector is calculated.
(B6) A new time interval within the predetermined time interval and a new feature based on the new time interval are determined.
(B7) Using the sensor value, a feature candidate vector that is a candidate for the feature vector corresponding to the feature amount of the new feature in the new time interval is calculated.
(B8) Based on the difference between the feature candidate vector and the feature vector, the feature candidate vector satisfying a predetermined condition based on the gradient vector is determined.
(B9) Using the feature candidate vector satisfying the predetermined condition, the first weight vector is modified to a second weight vector.
(B10) The feature and the time interval corresponding to the feature candidate vector satisfying the predetermined condition, and the second weight vector are stored in the storage unit.
(B11) Further, at least one sensor value selected from the user's acceleration, heart rate, body temperature, and angular velocity is acquired, and the user's acceleration, heart rate, body temperature, and angular velocity are acquired.
(B12) An estimated feature vector is calculated using the feature and time interval stored in the storage unit and the sensor value.
(B13) The athletic ability value is estimated using the second weight vector stored in the storage unit and the estimated feature vector.

The recording medium according to one aspect of the present disclosure is a recording medium provided with a control program for causing a device including a processor to execute a process, the recording medium is non-volatile and can be read by a computer, and the process is performed. ,
(B1) At least one sensor value selected from the user's acceleration, heart rate, body temperature, and angular velocity, and the user's athletic ability value are acquired.
(B2) The feature amount in the sensor value and the time interval within a predetermined time are determined.
(B3) Using the sensor value, a feature vector corresponding to the feature amount of the feature in the time interval is calculated.
(B4) Using the feature vector and the athletic ability value, a first weight vector for estimating the athletic ability value is acquired.
(B5) Using the first weight vector and the athletic ability value, a gradient vector related to the feature vector is calculated.
(B6) A new time interval within the predetermined time interval and a new feature amount based on the new time interval are determined.
(B7) Using the sensor value, a feature candidate vector that is a candidate for the feature vector corresponding to the feature amount of the new feature in the new time interval is calculated.
(B8) Based on the difference between the feature candidate vector and the feature vector, the feature candidate vector satisfying a predetermined condition based on the gradient vector is determined.
(B9) Using the feature candidate vector satisfying the predetermined condition, the first weight vector is modified to the second weight vector.
(B10) The feature and time interval corresponding to the feature candidate vector satisfying the predetermined condition, and the second weight vector are stored in the storage unit.

The recording medium according to one aspect of the present disclosure is a recording medium provided with a control program for causing a device including a processor to execute a process, the recording medium is non-volatile and can be read by a computer, and the process is performed. ,
(B1) At least one sensor value selected from the user's acceleration, heart rate, body temperature, and angular velocity, and the user's athletic ability value are acquired.
(B2) The feature amount in the sensor value and the time interval within a predetermined time are determined.
(B3) Using the sensor value, a feature vector corresponding to the feature amount of the feature in the time interval is calculated.
(B4) Using the feature vector and the athletic ability value, a first weight vector for estimating the athletic ability value is acquired.
(B5) Using the first weight vector and the athletic ability value, a gradient vector related to the feature vector is calculated.
(B6) A new time interval within the predetermined time interval and a new feature amount based on the new time interval are determined.
(B7) Using the sensor value, a feature candidate vector, which is a candidate of the feature vector corresponding to the feature amount of the new feature in the new time interval, is calculated.
(B8) Based on the difference between the feature candidate vector and the feature vector, the feature candidate vector satisfying a predetermined condition based on the gradient vector is determined.
(B9) Using the feature candidate vector satisfying the predetermined condition, the first weight vector is modified to the second weight vector.
(B10) The feature and the time interval corresponding to the feature candidate vector satisfying the predetermined condition, and the second weight vector are stored in the storage unit.
(B11) Further, at least one sensor value selected from the user's acceleration, heart rate, body temperature, and angular velocity is acquired.
(B12) An estimated feature vector is calculated using the feature and time interval stored in the storage unit and the sensor value.
(B13) It is included to estimate the athletic ability value by using the second weight vector stored in the storage unit and the estimated feature vector.

It should be noted that these comprehensive or specific embodiments may be realized by an integrated circuit or a computer program, and may be realized by any combination of an apparatus, a system, a method, an integrated circuit, a computer program and a computer-readable recording medium. May be done. Computer-readable recording media include non-volatile recording media such as CD-ROMs (Compact Disc-Read Only Memory).

According to the present disclosure, it is possible to improve the accuracy of the motor function estimated from the sensor data for a long time. Additional benefits and advantages of one aspect of the present disclosure will be apparent from the specification and drawings. This benefit and / or advantage can be provided individually by the various aspects and features disclosed herein and in the drawings, and not all are required to obtain one or more of them.

FIG. 1 is a block diagram showing a configuration of a motor function estimation system according to an embodiment. FIG. 2 is a diagram showing an example of a position where the acceleration sensor and the angular velocity sensor of FIG. 1 are attached to the subject. FIG. 3 is a diagram showing an example of the measured acceleration in the traveling direction when the acceleration sensor attached to the right foot is walking. FIG. 4 is a flowchart showing an example of an operation flow of estimation knowledge generation in the motor function estimation system according to the embodiment. FIG. 5 is a diagram showing an example of a combination of a time interval and an initial value of a feature to be extracted. FIG. 6 is an image showing an example of a method of calculating an average walking waveform. FIG. 7 is a diagram showing an example of a combination of the updated time interval and the feature to be extracted. FIG. 8 is a conceptual diagram of an example of a method for estimating the estimation knowledge of the embodiment. FIG. 9 is a conceptual diagram of an example of an estimation method of estimation knowledge of an embodiment using a neural network. FIG. 10 is a flowchart showing an example of an operation flow of motor function estimation in the motor function estimation system according to the embodiment. FIG. 11 is a diagram showing the correlation between the measured value and the estimated value of the knee extension muscle strength of the subject. FIG. 12 is a diagram comparing the estimation accuracy between the estimation method of the athletic ability value and the other estimation method in the embodiment.

The inventors related to the present disclosure, that is, the present inventors, examined the techniques described in Patent Document 1 and Non-Patent Document 1 mentioned in "Background Techniques", and developed a technique for improving the estimation accuracy of motor function. investigated. In the technique of Patent Document 1, a feature extracted from the sensor data of the sensor attached to the user is used for estimating the motor function of the user. However, in the technique of Patent Document 1, since the extracted feature amount does not sufficiently reflect the measurement data, the estimation accuracy of the motor function becomes low. The present inventors have found that the features extracted from the sensor data affect the estimation accuracy of the motor function, and investigated a technique for finding useful features. Further, the technique of Non-Patent Document 1 describes a technique of generating features by using machine learning, but machine learning is configured so that input data itself is input, and data having a long measurement time can be obtained. Not configured to be entered. The present inventors describe below in order to enable highly accurate motor function estimation even from long-time sensor data by extracting features useful for motor function estimation from long-time sensor data. I found a technology that would be used.

The motor function estimation information generation device according to one aspect of the present disclosure includes a sensor that measures at least one selected from a user's acceleration, heartbeat, body temperature, and angular velocity at a predetermined time, and a first processing circuit. The first processing circuit obtains (a1) at least one sensor value selected from the acceleration, heartbeat, body temperature, and angular velocity of the user, and (a2) the exercise ability value of the user. The initial feature in the sensor value and the time interval within the predetermined time are determined, (a3) the feature vector corresponding to the initial feature in the time interval is calculated using the sensor value, and (a4). ) The feature vector and the athletic ability value are used to obtain a first weight vector for estimating the athletic ability value, and (a5) the first weight vector and the athletic ability value are used. A gradient vector with respect to the feature vector is calculated, (a6) a new feature and a new time interval within the predetermined time are determined, and (a7) the sensor value is used to determine the new feature in the new time interval. A feature candidate vector that is a candidate for a feature vector corresponding to a new feature is calculated, and (a8) the feature candidate satisfying a predetermined condition based on the gradient vector based on the difference between the feature candidate vector and the feature vector. The vector is determined, (a9) the feature candidate vector satisfying the predetermined condition is used, the first weight vector is modified to the second weight vector, and (a10) the feature candidate satisfying the predetermined condition is satisfied. The feature and time interval corresponding to the vector, and the second weight vector are stored in the storage unit.

According to the above aspect, when the feature vector is updated to the more appropriate feature candidate vector, the weight vector is updated to the more appropriate weight vector by using the updated feature candidate vector. Such a weight vector constitutes the estimation knowledge for estimating the athletic ability value. Thereby, the estimation knowledge can extract the optimum feature vector and time interval for the input sensor value, and output the motor ability value based on the extracted feature vector and time interval. Therefore, the motor function estimation information generator can output the optimum motor performance value even for the sensor value having a long measurement time.

In the motor function estimation information generation device according to one aspect of the present disclosure, the weight vector may form a neural network and weight the nodes of the neural network. According to the above aspect, the neural network, which is one of machine learning, can improve the accuracy of the correct answer in the estimation result from the input information. Further, the neural network can keep the amount of input information, that is, the amount of processing for obtaining the estimation result low even if the number of input nodes is large.

In the motor function estimation information generation device according to one aspect of the present disclosure, the gradient vector is an error between the estimated value of the motor ability value estimated by using the first weight vector and the acquired motor ability value. It may be a gradient vector with respect to the feature vector for an index indicating. According to the above aspect, the gradient vector indicates the gradient of the change in the error with respect to the change in the feature vector. For example, by determining a feature candidate vector that changes from the feature vector in the direction of the gradient of the gradient vector so that the error becomes smaller, the second weight vector based on the feature candidate vector is larger than the first weight vector. Highly accurate estimation knowledge can be formed.

In the motor function estimation information generation device according to one aspect of the present disclosure, the feature candidate vector satisfying the predetermined condition is determined based on the difference between the feature candidate vector and the feature vector and the degree of coincidence with the gradient vector. May be done. According to the above aspect, the feature candidate vector satisfying a predetermined condition is determined so that the difference between the feature candidate vector and the feature vector and the gradient vector correlate with each other. The feature candidate vector can be easily determined by the method as described above.

The motor function estimation system according to one aspect of the present disclosure includes the motor function estimation information generation device and a second processing circuit, and the second processing circuit is based on the acceleration, heartbeat, body temperature, and angular velocity of the user. The estimated feature vector is calculated by acquiring at least one selected sensor value, using the feature and time interval stored in the storage unit, and the sensor value, and storing the feature and time interval in the storage unit. The athletic ability value is estimated using the second weight vector and the estimated feature vector, and the athletic ability value is output.

According to the above aspect, the motor function estimation system can output the optimum motor ability value with respect to the input sensor value by using the estimation knowledge of the motor function estimation information generator.

In the motor function estimation information generation method according to one aspect of the present disclosure, (b1) at least one sensor value selected from the acceleration, heartbeat, body temperature, and angular velocity of the user and the exercise ability value of the user are acquired. (B2) The initial feature in the sensor value and the time interval within a predetermined time are determined, and (b3) the feature vector corresponding to the initial feature in the time interval is calculated using the sensor value. , (B4) the first weight vector for estimating the athletic ability value is obtained by using the feature vector and the athletic ability value, and (b5) the first weight vector and the athletic ability value are obtained. Using, the gradient vector for the feature vector is calculated, (b6) the new feature and the new time interval within the predetermined time are determined, and (b7) the sensor value is used to determine the new time. A feature candidate vector that is a candidate for the feature vector corresponding to the new feature is calculated in the section, and (b8) a predetermined condition based on the gradient vector is satisfied based on the difference between the feature candidate vector and the feature vector. The feature candidate vector is determined, and (b9) the feature candidate vector satisfying the predetermined condition is used to modify the first weight vector to a second weight vector.

According to the above aspect, the same effect as that of the motor function estimation information generation device according to one aspect of the present disclosure can be obtained.

In the motor function estimation information generation method according to one aspect of the present disclosure, the weight vector may form a neural network and weight the nodes of the neural network.

In the motor function estimation information generation method according to one aspect of the present disclosure, the gradient vector is an error between the estimated value of the motor ability value estimated by using the first weight vector and the acquired motor ability value. It may be a gradient vector with respect to the feature vector for an index indicating.

In the motor function estimation information generation method according to one aspect of the present disclosure, the feature candidate vector satisfying the predetermined condition is determined based on the difference between the feature candidate vector and the feature vector and the degree of coincidence with the gradient vector. May be done.

In the motor function estimation method according to one aspect of the present disclosure, (b1) at least one sensor value selected from the acceleration, heartbeat, body temperature, and angular velocity of the user and the motor ability value of the user are acquired, and (b2). ) The initial feature amount in the sensor value and the time interval within a predetermined time are determined, and (b3) the feature vector corresponding to the initial feature amount in the time interval is calculated using the sensor value. , (B4) the first weight vector for estimating the athletic ability value is obtained by using the feature vector and the athletic ability value, and (b5) the first weight vector and the athletic ability value are obtained. Using, a gradient vector for the feature vector is calculated, (b6) a new feature quantity and a new time interval within the predetermined time are determined, and (b7) the sensor value is used to determine the new feature amount. A feature candidate vector that is a candidate for the feature vector corresponding to the new feature amount is calculated in the time interval, and (b8) a predetermined condition based on the gradient vector based on the difference between the feature candidate vector and the feature vector. The feature candidate vector satisfying the condition is determined, (b9) the feature candidate vector satisfying the predetermined condition is used, the first weight vector is modified to the second weight vector, and (b10) the predetermined condition is satisfied. The feature amount and time interval corresponding to the feature candidate vector satisfying the condition, and the second weight vector are stored in the storage unit, and (b11) further, at least selected from the user's acceleration, heartbeat, body temperature, and angular velocity. One sensor value is acquired, an estimated feature vector is calculated using (b12) the feature amount and time interval stored in the storage unit, and the sensor value, and (b13) is stored in the storage unit. The athletic ability value is estimated using the second weight vector and the estimated feature vector.

According to the above aspect, the same effect as the motor function estimation system according to one aspect of the present disclosure can be obtained.

The recording medium according to one aspect of the present disclosure is a recording medium provided with a control program for causing a device including a processor to execute a process, the recording medium is non-volatile, can be read by a computer, and the process is performed. , (B1) At least one sensor value selected from the user's acceleration, heartbeat, body temperature, and angular velocity, and the user's athletic ability value are acquired, and (b2) the initial feature of the sensor value and a predetermined value. The time interval in time is determined, (b3) the sensor value is used to calculate the feature vector corresponding to the initial feature in the time interval, and (b4) the feature vector and the athletic ability value are calculated. The first weight vector for estimating the athletic ability value is obtained by using (b5), and the gradient vector related to the feature vector is calculated by using the first weight vector and the athletic ability value (b5). b6) A new feature and a new time interval within the predetermined time are determined, and (b7) the sensor value is used as a candidate for a feature vector corresponding to the new feature in the new time interval. A certain feature candidate vector is calculated, (b8) the feature candidate vector satisfying a predetermined condition based on the gradient vector is determined based on the difference between the feature candidate vector and the feature vector, and (b9) the predetermined feature vector is determined. Using the feature candidate vector satisfying the condition, the first weight vector is modified to the second weight vector, and (b10) the feature and the time interval corresponding to the feature candidate vector satisfying the predetermined condition, and It includes storing the second weight vector in the storage unit.

According to the above aspect, the same effect as that of the motor function estimation information generation device according to one aspect of the present disclosure can be obtained.

In the recording medium according to one aspect of the present disclosure, the weight vector may form a neural network and weight the nodes of the neural network.

In the recording medium according to one aspect of the present disclosure, the gradient vector is an index indicating an error between the estimated value of the athletic ability value estimated by using the first weight vector and the acquired athletic ability value. It may be a gradient vector related to the feature vector.

In the recording medium according to one aspect of the present disclosure, the feature candidate vector satisfying the predetermined condition may be determined based on the difference between the feature candidate vector and the feature vector and the degree of coincidence with the gradient vector. ..

The recording medium according to one aspect of the present disclosure is a recording medium provided with a control program for causing a device including a processor to execute a process, the recording medium is non-volatile, can be read by a computer, and the process is performed. , (B1) At least one sensor value selected from the user's acceleration, heartbeat, body temperature, and angular velocity, and the user's athletic ability value are acquired, and (b2) the initial feature in the sensor value and a predetermined. The time interval in time is determined, (b3) the sensor value is used to calculate the feature vector corresponding to the initial feature in the time interval, and (b4) the feature vector and the athletic ability value are calculated. The first weight vector for estimating the athletic ability value is obtained by using (b5), and the gradient vector related to the feature vector is calculated by using the first weight vector and the athletic ability value (b5). b6) A new feature and a new time interval within the predetermined time are determined, and (b7) the sensor value is used as a candidate for a feature vector corresponding to the new feature in the new time interval. A certain feature candidate vector is calculated, (b8) the feature candidate vector satisfying a predetermined condition based on the gradient vector is determined based on the difference between the feature candidate vector and the feature vector, and (b9) the predetermined feature vector is determined. Using the feature candidate vector satisfying the condition, the first weight vector is modified to the second weight vector, and (b10) the feature and the time interval corresponding to the feature candidate vector satisfying the predetermined condition, and The second weight vector is stored in the storage unit, (b11) further, at least one sensor value selected from the user's acceleration, heartbeat, body temperature, and angular velocity is acquired, and (b12) is stored in the storage unit. The estimated feature vector is calculated using the feature and the time interval, and the sensor value, and (b13) the second weight vector stored in the storage unit and the estimated feature vector are used. , Includes estimating athletic performance values.

According to the above aspect, the same effect as the motor function estimation system according to one aspect of the present disclosure can be obtained.

Hereinafter, the motor function estimation system and the like according to the embodiment will be described with reference to the drawings. The embodiments described below are comprehensive or specific examples. The numerical values, shapes, materials, components, arrangement positions and connection forms of the components, steps (processes), order of steps, etc. shown in the following embodiments are examples, and are not intended to limit the present disclosure. .. Further, among the components in the following embodiments, the components not described in the independent claim indicating the highest level concept are described as arbitrary components. Further, in the following description of the embodiment, expressions with "abbreviations" such as substantially parallel and substantially orthogonal may be used. For example, substantially parallel means not only completely parallel, but also substantially parallel, that is, including a difference of, for example, about several percent. The same applies to other expressions with "abbreviations".

[Embodiment]
[1. Configuration of motor function estimation system according to the embodiment]
First, with reference to FIG. 1, the configuration of the motor function estimation system 1000 according to the embodiment will be described. Note that FIG. 1 is a block diagram showing a configuration of the motor function estimation system 1000 according to the embodiment. The motor function estimation system 1000 estimates the state of the user's motor function, for example, the motor ability value from the measurement result of the sensor attached to the user. In this specification, the user may be referred to as a subject.

As shown in FIG. 1, the motor function estimation system 1000 includes an estimation information generation device 100, a first storage unit 200, a second storage unit 210, an estimation device 300, a sensor 400, and a display unit 500. .. The motor function estimation system 1000 may be composed of one device or a plurality of devices. The motor function estimation system 1000 may be incorporated in all or part of the device and form a part of the device. At least the estimation information generation device 100 of the estimation information generation device 100, the first storage unit 200, the second storage unit 210, and the estimation device 300 may form the motor function estimation device.

The estimation information generation device 100 builds estimation knowledge for estimating an ability value indicating a user's motor function from the measurement result of the sensor 400. For example, the estimation information generation device 100 constructs estimation knowledge using machine learning. In the present embodiment, the estimation information generation device 100 extracts the time interval used for estimation and the feature to be calculated as the estimation knowledge for estimating the user's motor function from the user's behavior sensor value at a predetermined time. , Calculate the relationship between the calculated characteristics and the athletic ability value. The user's behavior sensor value includes the measurement result of the sensor 400 attached to the user. For example, the behavior sensor value may include the measured value associated with the measured time. The estimation information generation device 100 includes a data acquisition unit 110, an initial feature determination unit 120, an initial feature calculation unit 130, a weight adjustment unit 140, a feature gradient calculation unit 150, a feature candidate calculation unit 160, and a feature search unit. It has 170, a determination unit 180, and an estimation knowledge output unit 190.

The estimation device 300 estimates the user's motor function from the behavior sensor value acquired by the sensor 400 using the estimation knowledge constructed by the estimation information generation device 100. The estimation device 300 includes a sensor value acquisition unit 310, a feature acquisition unit 320, a feature amount calculation unit 330, an estimation knowledge acquisition unit 340, a motor function estimation unit 350, and an output unit 360.

The estimation information generation device 100 and the estimation device 300 may be configured by hardware such as a circuit or an integrated circuit including each of the above-mentioned components, or may be realized by software such as a program executed on a computer. good. For example, the above-mentioned components may be configured by a computer system (not shown) including a CPU (Central Processing Unit), a RAM (Random Access Memory), a ROM (Read-Only Memory), and the like. Some or all of the functions of the above components may be achieved by the CPU using the RAM as working memory to execute a program recorded in the ROM. The program may be provided as an application by communication via a communication network such as the Internet, communication according to a mobile communication standard, or the like.

Such an estimation information generation device 100 and an estimation device 300 may form one component or may form separate components. The estimation information generation device 100 and the estimation device 300 include, for example, a computer device, an MPU (Micro Processing Unit), a CPU, a processor, a circuit such as an LSI (Large Scale Integration), an IC card (Integrated Circuit Card), or an IC card. It may be a single module or the like.

The estimation information generation device 100 and the estimation device 300 may be arranged in the band 400a shown in FIG. 2 together with the sensor 400, and may perform wired communication or wireless communication with a display unit 500 or a device including the display unit 500. Alternatively, the estimation information generation device 100 and the estimation device 300 may be arranged at a position away from the sensor 400, and may perform wired communication or wireless communication with the sensor 400. In this case, the estimation information generation device 100 and the estimation device 300 may independently form a module, or may be incorporated in another device such as a computer device. Any existing wired communication may be applied to the above wired communication. Any existing wireless communication may be applied to the wireless communication. For example, a wireless LAN (Local Area Network) such as Wi-Fi (registered trademark) (Wireless Fidelity) may be applied to the wireless communication, and a short distance such as Bluetooth (registered trademark) or ZigBee (registered trademark) may be applied. Wireless communication may be applied.

(Sensor 400)
The sensor 400 is attached to the user, and as the measurement result, the user's action value is acquired and output. The example of the action value is the value of at least one parameter selected from the parameters such as acceleration, heart rate, body temperature and angular velocity during the user's daily activities or during the 6-minute walking test which is an example of the exercise test. The 6-minute walk test (6-minute walk test: 6MWT) is a quantitative load test in which the person walks at a constant speed. Specifically, the subject who is the user walks as fast as possible for 6 minutes, and the exercise tolerance of the subject is evaluated based on the walking distance obtained as a result.

The sensor 400 may be configured to output the user's ID and the user's action value in association with each other as a measurement result. For example, the sensor 400 may have an input unit and may be configured to acquire a user ID input by the user to the input unit. For example, the input unit may be a touch panel or a character and / or a number key. The sensor 400 may be configured to communicate with the estimation information generation device 100 and the estimation device 300 by wire or wirelessly, and may transmit the measured value of the sensor 400 (hereinafter, also referred to as a sensor value) to these. Any existing wired communication may be applied to the above wired communication. Any existing wireless communication may be applied to the wireless communication. For example, a wireless LAN such as Wi-Fi (registered trademark) may be applied to the wireless communication, or short-range wireless communication such as Bluetooth (registered trademark) or ZigBee (registered trademark) may be applied.

In this embodiment, the sensor 400 includes an acceleration sensor 401, a heart rate sensor 402, a temperature sensor 403, and an angular velocity sensor 404. However, the sensor 400 may include some of the accelerometer 401, heart rate sensor 402, temperature sensor 403 and angular velocity sensor 404, and other sensors such as humidity sensors, barometric pressure sensors and the like for detecting the user and its surroundings. May include.

The acceleration sensor 401 is attached to a portion of the user's body to be measured and measures the acceleration of the measurement target portion. The acceleration sensor 401 may be a uniaxial acceleration sensor that measures acceleration in one direction, may be a biaxial acceleration sensor that measures acceleration in two orthogonal directions, or may be a biaxial acceleration sensor that measures acceleration in three orthogonal directions. It may be an axial acceleration sensor. Accelerometers with two or more axes can detect acceleration in a desired direction regardless of the arrangement direction of the acceleration sensors. A plurality of uniaxial acceleration sensors may be used to detect accelerations of two or more axes. In the present embodiment, the acceleration sensor 401 is attached to, for example, the user's foot and measures the acceleration of the foot in the traveling direction of the user, although not limited to the present invention. More specifically, the accelerometer 401 measures the acceleration of the user's foot during daily life or during an exercise test. Exercise here includes walking or running. The traveling direction of the user is the same as the walking direction or the traveling direction of the user. In the present embodiment, a specific example of the acceleration sensor value, which is the measured value of the acceleration sensor 401, is the acceleration of the three axes at the left ankle, the right ankle, and the waist.

The accelerometer 401 is attached to the user's foot or waist during exercise. More specifically, the accelerometer 401 is arranged at a position near the user's ankle. The location near the user's ankle may include the ankle, instep and sole of the foot. For example, FIG. 2 shows an example of the position of the accelerometer 401 worn by the user. In this way, the accelerometer 401 is attached to at least one ankle of the user. For example, the accelerometer 401 is placed on a band 400a that is wrapped around the user's ankle and worn. Alternatively, the accelerometer 401 may be placed inside the user's shoes.

For example, referring to FIG. 3, an example of the result of measuring the acceleration of the user's foot during exercise using the acceleration sensor 401 is shown in a graph. Note that FIG. 3 shows an example of the measurement result of the accelerometer 401 in the traveling direction during walking in the exercise test, and specifically, a part of the time section in the 6-minute walking test, which is one of the exercise tests. An example of the measurement result of the acceleration sensor 401 is shown. Specifically, FIG. 3 is measurement data of acceleration over a period of 6 seconds. The vertical axis of FIG. 3 is acceleration (unit: mG), and the horizontal axis is time (unit: sec [seconds]). It should be noted that 1G = 9.80665 m / s ² . In the graph shown in FIG. 3, the acceleration forward in the traveling direction of the user is a positive value.

The heart rate sensor 402 measures the user's heart rate. The heart rate sensor 402 measures the user's heart rate during daily activities or exercise tests. The heart rate sensor 402 is placed, for example, on the user's wrist, ankle, or chest.

The temperature sensor 403 measures the user's body temperature. The temperature sensor 403 measures the body temperature of the user during daily activities or exercise tests. The temperature sensor 403 may be arranged in the same place as the acceleration sensor 401 or the heart rate sensor 402, or may be arranged in another place.

The angular velocity sensor 404 measures the user's angular velocity. The angular velocity sensor 404 measures the angular velocity of the user during daily life or an exercise test. The angular velocity sensor 404 may be arranged together at the same position as the acceleration sensor 401, or may be arranged at another place. A specific example of the angular velocity sensor value of such an angular velocity sensor is the angular velocity of three axes in the left ankle, the right ankle, and the waist.

(Estimated information generator 100)
The data acquisition unit 110 acquires one or more learning data in which a behavior sensor value and an athletic ability value at a predetermined time are set. The behavior sensor value includes a measurement result measured by a sensor 400 or the like over a predetermined time. Specifically, the behavior sensor value includes the measured value associated with the measured time. The correspondence between the measured time and the measured value may be performed by, for example, the sensor 400 or the data acquisition unit 110 when the measured value is acquired, and the data acquisition in the estimation information generation device 100 may be performed. It may be done by other components of unit 110. The measured value may be a measured value directly acquired via the sensor 400, or may be a measured value stored in a storage means such as the first storage unit 200.

The exercise ability value is an ability value related to muscle strength, endurance, balance ability, etc., and is an exercise ability value corresponding to a behavior sensor value. For example, the exercise ability value is exemplified by the muscle strength value, the maximum oxygen uptake, the balance ability value, etc., which will be described later. The behavior sensor value and the motor ability value may be prepared in advance for constructing the estimation knowledge of the estimation information generation device 100. The behavior sensor value and the motor ability value may be specific information about a specific user, or may be general-purpose information about an unspecified number of users. The combination of the behavior sensor value and the exercise ability value is a combination of the behavior sensor value acquired from the user through the sensor 400 by performing an exercise test or the like and the exercise ability value detected from the user using various devices. There may be. That is, both the behavior sensor value and the athletic ability value may be actually measured values instead of estimated values. The data acquisition unit 110 acquires a set of a behavior sensor value and an athletic ability value as learning data in the estimation information generation device 100.

From the behavior sensor values acquired by the data acquisition unit 110, the initial feature determination unit 120 determines a specific time interval, which is a time interval used for estimating motor ability for learning, and a feature to be extracted in the specific time interval. do. Specifically, the initial feature determination unit 120 determines a feature to be extracted in a specific time interval, and calculates a feature amount obtained by quantifying the feature. Then, the initial feature determination unit 120 sets the calculated feature amount as the initial value of the feature amount in the extracted feature which is the extracted feature. The predetermined time, which is the measurement period of the behavior sensor value, is divided into a plurality of time intervals, and these divided time intervals are referred to as time intervals. Further, the feature is a feature relating to the measurement parameter of the sensor 400. The features may relate to, for example, indicators such as mean value, variance value, entropy, correlation coefficient, PCA (principal component analysis, Restricted Boltzmann Machine), and RBM (Restricted Boltzmann Machine) of measured values of parameters. .. Then, the feature amount may be a numerical value calculated from the measured values of the parameters in the time interval using the above index. The initial feature determination unit 120 may calculate feature quantities of a plurality of features.

For example, when the feature to be extracted is an acceleration feature, that is, the above index, the initial feature determination unit 120 determines the extraction feature of the specific time section and the acceleration by a preset method, and the measured value of the acceleration in the specific time section. Therefore, the feature amount of the extraction feature of the acceleration may be calculated, and the calculated feature amount may be used as the initial value of the feature amount. Alternatively, the initial feature determination unit 120 arbitrarily determines the extraction feature of the specific time section and the acceleration, calculates the feature amount of the acceleration extraction feature from the measured value of the acceleration in the specific time section, and features the calculated feature amount. It may be the initial value of the quantity. Alternatively, the initial feature determination unit 120 may use preset specific time intervals and initial values of the feature amounts of the extracted features. The time interval is preferably long so that the feature amount does not exhibit two or more different behaviors within the time interval. For example, the time interval is preferably long so as not to include the measured values indicating two or more actions in the time interval, in other words, the tendency of the measured values in the time interval does not change. It is preferably a length. As will be described later, the length of the time interval may be, for example, several seconds or several tens of seconds, specifically, 1 second, 5 seconds, 10 seconds, 15 seconds, 20 seconds, 25 seconds, 30 seconds, etc. May be.

The initial feature calculation unit 130 calculates a feature vector corresponding to the feature amount of the specific time interval and the extracted feature determined by the initial feature determination unit 120 from the behavior sensor value acquired by the data acquisition unit 110. Specifically, the feature vector calculated by the initial feature calculation unit 130 includes the feature amount of the extracted feature calculated from the measured value of the behavior sensor value included in the specific time interval as an element. The feature vector may include the feature amount of each of the plurality of extracted features as a plurality of elements.

The weight adjusting unit 140 adjusts the weight vector so that the exercise ability value acquired by the data acquisition unit 110 can be estimated by the estimation formula including the feature vector calculated by the initial feature calculation unit 130 and the weight vector. The weight vector is a vector that weights each element of the feature vector. Estimating the athletic ability value from the feature vector In the above estimation formula, the weight vector weights each element of the feature vector in the process of estimation calculation so that an appropriate result of the athletic ability value can be obtained. do.

The feature gradient calculation unit 150 obtains a gradient related to the feature vector from the exercise ability value acquired by the data acquisition unit 110, the feature vector calculated by the initial feature calculation unit 130, and the weight vector calculated by the weight adjustment unit 140. The gradient with respect to the feature vector may be, for example, the gradient of the element with respect to the error of the estimation formula with respect to the feature vector. Further, the gradient is a vector, and is also referred to as a feature gradient vector below. The details of the feature gradient vector will be described later, but the feature gradient vector may be, for example, a partial derivative of the error of the estimation formula with respect to the feature vector.

The feature candidate calculation unit 160 determines a time interval when the current feature vector is calculated, that is, a time interval different from the specific time interval, as a candidate for a new time interval. For example, the feature candidate calculation unit 160 determines a plurality of new time intervals as candidates. The new time interval may be arbitrarily selected within the measurement period of the behavior sensor value as long as it is a time interval different from the specific time interval which is the time interval corresponding to the current feature vector. The feature candidate calculation unit 160 determines a new time interval different from the specific time interval for the feature corresponding to the feature amount included in the current feature vector, and uses the measured value included in the determined time interval. , Calculate the feature amount of the above features. Then, the feature candidate calculation unit 160 calculates a feature candidate vector including the calculated feature amount as an element. The quantity of the feature candidate vector is the same as the quantity of the time interval determined as a candidate. The details of the feature candidate vector will be described later.

From the plurality of feature candidate vectors calculated by the feature candidate calculation unit 160, the feature search unit 170 selects the feature candidate vector having the highest degree of agreement between the feature candidate vector and the current feature vector with respect to the feature gradient vector. Explore. Specifically, the feature search unit 170 searches for a feature candidate vector whose sign of the element of the difference between the feature candidate vector and the current feature vector best matches the sign of the element of the feature gradient vector. As will be described later, in the present embodiment, the above difference is a difference obtained by subtracting the feature candidate vector from the current feature vector, but the difference is not limited to this.

The determination unit 180 determines the end of learning based on the search result of the feature candidate vector by the feature search unit 170. Specifically, when the feature candidate vector is extracted by the search by the feature search unit 170, the determination unit 180 continues learning, that is, adjustment of the weight vector or the like by using the extracted feature candidate vector as a new feature vector. When the feature candidate vector is not extracted by the search, the determination unit 180 ends the learning and determines the existing information such as the existing weight vector as an appropriate component of the estimation knowledge of the athletic ability value. The weight adjustment unit 140, the feature gradient calculation unit 150, the feature candidate calculation unit 160, the feature search unit 170, and the determination unit 180 constitute a learning unit that allows machine learning to be performed on the estimated knowledge of the motor ability value. , Let the estimation knowledge of the athletic ability value learn the weighting of the estimation process.

The estimation knowledge output unit 190 stores in the first storage unit 200 a combination of a new time interval corresponding to the feature candidate vector obtained by the feature search unit 170 and a feature corresponding to the feature amount included in the feature candidate vector. Then, the weight vector calculated by the weight adjusting unit 140 is stored in the second storage unit 210.

(First storage unit 200 and second storage unit 210)
The first storage unit 200 and the second storage unit 210 are configured so that various information can be stored and the stored information can be taken out. Specifically, the first storage unit 200 stores information such as a combination of time intervals and features to be extracted from the behavior sensor value. The first storage unit 200 may store the behavior sensor value. In this case, the first storage unit 200 may store only the measured value of the behavior sensor value, or may store the measured value of the behavior sensor value and the measured time in association with each other. The second storage unit 210 stores the construction information of the estimation knowledge such as the weight vector constituting the estimation knowledge. The first storage unit 200 and the second storage unit 210 may be composed of, for example, a hard disk or a semiconductor memory. The first storage unit 200 and the second storage unit 210 may form separate components, or may form one component together.

(Estimator 300)
The estimation device 300 estimates the motor ability value from the input behavior sensor value by using the estimation knowledge generated by the estimation information generation device 100. For example, the estimation device 300 estimates the exercise ability value of the exercise corresponding to the input parameter by inputting the measurement result of the sensor 400 for the exercising user.

The sensor value acquisition unit 310 acquires a behavioral sensor value including a measurement result over a predetermined time from a sensor 400 mounted on a predetermined one or more places on the user's body.

The feature acquisition unit 320 acquires a combination of time intervals and features to be extracted from the behavior sensor value from the first storage unit 200 in order to extract the feature used for motor function estimation from the behavior sensor value. As described above, the above combination is a combination of time intervals and features calculated by the estimation information generation device 100 and effective for estimating the motor function. Specifically, it is a combination of the time interval corresponding to the feature candidate vector stored in the first storage unit 200 by the estimation knowledge output unit 190 and the feature corresponding to the feature amount included in the feature candidate vector.

The feature amount calculation unit 330 calculates the feature vector corresponding to the time interval and the feature acquired by the feature acquisition unit 320 from the behavior sensor value acquired by the sensor value acquisition unit 310. Specifically, the feature amount calculation unit 330 extracts the time interval corresponding to the acquired time interval from the time intervals in the behavior sensor value. Further, the feature amount calculation unit 330 calculates the feature amount of the acquired feature by using the measured value included in the extracted time interval. For example, when there are a plurality of features corresponding to the acquired time intervals, the feature amount calculation unit 330 calculates a plurality of feature amounts. Then, the feature amount calculation unit 330 calculates the feature vector using the extracted time interval and the calculated feature amount.

The estimation knowledge acquisition unit 340 acquires a weight vector from the second storage unit 210 as estimation knowledge for estimating the motor ability value from the feature vector.

The motor function estimation unit 350 estimates the motor ability value using the feature vector calculated by the feature amount calculation unit 330 and the weight vector acquired by the estimation knowledge acquisition unit 340. For example, the motor function estimation unit 350 estimates the user's motor ability value using the weight vector and the feature vector according to the estimation formula used by the weight adjustment unit 140 for adjusting the weight vector.

The output unit 360 outputs the athletic ability value estimated by the motor function estimation unit 350. The output unit 360 may output the exercise ability value to the display unit 500, or may output it to an external device different from the exercise function estimation system 1000.

(Display unit 500)
The display unit 500, also called a notification unit, outputs various information input from the estimation device 300 visually and / or audibly. Specifically, the display unit 500 notifies the exercise ability value estimated by the estimation device 300. An example of the display unit 500 is a display or a speaker. For example, the display unit 500 may be a liquid crystal panel, a display including a display panel such as an organic or inorganic EL (Electroluminescence), a speaker, or a combination thereof. The display unit 500 may be a part of a computer device equipped with the motor function estimation system 1000, or may be a part of a device such as a computer device not equipped with the motor function estimation system 1000, and may be a single device. There may be. The display unit 500 may be a part of a user's mobile terminal such as a smartphone, a smart watch, or a tablet. Alternatively, the display unit 500 may be arranged in the band 400a shown in FIG. 2 together with the sensor 400.

[2. Operation of the motor function estimation system according to the embodiment]
Next, the operation of the motor function estimation system 1000 according to the embodiment will be described. Specifically, the operation of the estimation information generation device 100 and the operation of the estimation device 300 will be mainly described.

[2-1. Processing operation for creating estimated information by the estimated information generator]
With reference to FIG. 4, the processing operation of the estimation information creation by the estimation information generation device 100 will be described. Note that FIG. 4 is a flowchart showing an example of an operation flow of estimation information generation in the motor function estimation system 1000 according to the embodiment.

(Step S10)
The data acquisition unit 110 acquires the user's motor ability value and the user's behavior sensor value. The user's behavior sensor value includes the measured value measured by the sensor 400 mounted on the user and the measured time corresponding to the measured value, that is, the time when the measured value is measured. In the present embodiment, the acquired athletic ability value and behavior sensor value are not the estimated information but the information actually measured by the user. The behavior sensor value has data over a predetermined time. As described above, the behavior sensor value is information in which the data and the measurement time of the data are associated with each other. As an example of the data, as shown in FIG. 3, the acceleration value for a certain period of time is included. The exercise ability value of the user is, for example, a muscle strength value such as a lower limb muscle strength value, a maximum oxygen uptake, a balance ability value, and the like.

The lower limb muscle strength value is, for example, the average value of the knee extension muscle strength. For example, the lower limb muscle strength value can be measured by using a isokinetic muscle strength measuring method that controls the angular velocity of flexion and extension of the knee joint to a constant speed and measures the muscle strength exerted at that time, that is, the joint torque. be. The balance ability is, for example, the ability of the user to maintain balance under gravity. More specifically, the balance ability is the ability of the user to maintain equilibrium by maintaining the center of gravity of the body within the support basal plane. Alternatively, the balance ability is the ability to maintain equilibrium by returning the center of gravity of the body to the inside of the support basal plane when the center of gravity of the body deviates from the inside of the support basal plane. The balance ability value uses the functional reach method that measures the maximum distance that the arm can be extended forward, the open-eye or one-leg standing time method that measures the time that one leg can be continued in the open or closed state. It is measurable. VO2 max is an example of an indicator of endurance. Maximum oxygen uptake can be detected by measuring the subject's respiratory volume, oxygen consumption and carbon dioxide emissions as the load is increased stepwise using a treadmill or bicycle ergometer.

For example, the data acquisition unit 110 may acquire the user's behavior sensor value and the user's athletic ability value recorded in the database. For example, the user ID, the behavior sensor value, and the athletic ability value may be combined and recorded in the database. Alternatively, the data acquisition unit 110 may acquire the behavior sensor value from the sensor 400 and acquire the athletic ability value from the database. The data acquisition unit 110 may acquire the athletic ability value recorded in the database by referring to the user ID associated with the behavior sensor value.

In other words, the data acquisition unit 110 receives a plurality of data in which the behavior sensor value for a predetermined time measured by the sensor 400 and the motor ability value are set as learning data for generating the motor function estimation knowledge. get. The behavior sensor value and the athletic ability value may be composed of the measured values of a specific user, or may be composed of the measured values of an unspecified number of users. By using the behavior sensor value and the athletic ability value composed of the measured values of a specific user, the estimation knowledge specialized for the specific user can be obtained, and the behavior sensor value composed of the measured values of an unspecified number of users can be obtained. And by using the athletic ability values, general-purpose estimation knowledge can be obtained.

(Step S20)
When the initial feature determination unit 120 determines a plurality of features to be extracted based on the behavior sensor value, for example, based on the parameters of the measured values included in the behavior sensor value, and further calculates the feature amount of the feature. Determine the section. For example, the initial feature determination unit 120 has a feature determination database, and records a plurality of selection criterion data in the feature determination database. The feature determination database may be stored in the first storage unit 200.

For example, one selection criterion data may include one parameter of the measured value, a feature valid for the parameter, and a time interval useful for calculating the feature amount of the feature as information. One selection criterion data may be data in which a parameter, a feature effective for the parameter, and a time interval for calculating the feature amount of the feature are associated with each other. One selection criterion data may be data that independently includes a feature valid for the parameter and a time interval for calculating the feature amount of the feature. In one selection criterion data, a plurality of features may be associated with one parameter. A plurality of time intervals may be associated with one feature. The selection criterion data includes the parameter of the measured value, one or more features effective for expressing the feature of the parameter, and the time interval effective for calculating the feature amount of each feature in association with each other. I'm out.

The initial feature determination unit 120 refers to a plurality of selection reference data recorded in the feature determination database, and determines a feature effective for the parameter of the measured value and a time interval for calculating the feature amount of the feature. The time interval determined by the initial feature determination unit 120 is the initial time interval, and the feature amount of the feature calculated by the initial feature determination unit 120 within the time interval is the initial value of the feature amount. Alternatively, the initial feature determination unit 120 may randomly determine the time interval and the feature within a predetermined time interval.

When one selection criterion data includes a plurality of features, the initial feature determination unit 120 randomly selects at least two features from the plurality of features and determines the selected features as extraction features. Examples of features are mean, variance, entropy, correlation coefficient, PCA (Principal Component Analysis), RBM (Restricted Boltzmann Machine) and the like. The initial feature determination unit 120 sets the time interval effective for the determined feature and the feature amount of the feature calculated by the measured value of the time section as the initial value of the time interval and the feature amount, respectively, and such an initial value. Are prepared for the number of determined features.

Referring to FIG. 5, an example of a feature and a time interval determined by the initial feature determination unit 120 according to the selection criterion data is shown. FIG. 5 shows an example of features and time intervals extracted when PCA is used. In this example, in the selection criterion data, the feature corresponding to the parameter of the measured value included in the behavior sensor value is "the coefficient of the principal component when the behavior sensor value is expressed by the principal component vector of PCA", which is a feature. The time interval valid for the coefficient of a certain principal component is the "time interval from 0 seconds to 30 seconds". In this example, by performing analysis by PCA on one behavior sensor value, 32 features by the coefficients of the first principal component to the 32nd principal component are extracted. The value of the coefficient of each principal component is the feature amount. The initial feature determination unit 120 extracts 32 features and their feature quantities from each of the plurality of behavior sensor values acquired in step S10. Feature numbers n (1 ≦ n ≦ 32) are sequentially attached to the 32 features.

In this example, the behavior sensor value includes the measured value for 6 minutes. In the 6 minutes, which is the measurement period of the behavior sensor value, the initial feature determination unit 120 sets the start time nt of the time interval (in this example, the length of the time interval is 30 seconds) and the PCA for each feature of the feature number n. The coefficient of the principal component of the principal component number kn is determined. The principal component number kn is an ordinal number of the principal component in FIG. In this example, the initial feature determination unit 120 determines that the start time tun = 0 and the number kn = n. At this time, the features of the feature numbers 1 to 32 are the first principal component to the 32nd principal component when the behavior sensor values from 0 seconds to 30 seconds in the corresponding time interval are expressed by the principal component vector of the PCA. It is a coefficient. The initial feature determination unit 120 may arbitrarily determine the start time nt and the number kn.

In PCA, when the behavior sensor value in a certain time interval is _VA and the principal component vectors containing the principal component elements of the same principal component number are _{VP1 to VP32 , respectively} , _VA = _VP1 and C1 + _VP2 . -C ₂ + ... + _VP32 -C ₃₂ . Here, C ₁ to C ₃₂ each represent the coefficients of the principal components having the same principal component number. The characteristics of the behavior sensor value can be expressed by the coefficients C ₁ to C ₃₂ .

Referring to FIG. 6, another example of the procedure for extracting the feature amount according to the selection criterion data is shown. This example shows an example of a feature amount and a time interval calculated when an average value is used as a feature. In this example, the selection criterion data includes "normalization and addition averaging of the behavior sensor value" as a feature corresponding to the parameter of the measured value included in the behavior sensor value. This feature is a calculation result obtained by dividing the behavior sensor values in the time interval from 0 seconds to 6 seconds, normalizing each behavior sensor value after the division, and averaging the values. The effective time interval for the feature is about 6/7 seconds, which is the time of one cycle of walking.

FIG. 6 includes acceleration data for 6 seconds during walking, as in FIG. FIG. 6 shows an image of an example of a method of calculating an average walking waveform of acceleration as a feature from an acceleration waveform for 6 seconds. In the example of FIG. 6, the initial feature determination unit 120 extracts seven acceleration values from the entire time interval (6 seconds) for each time interval (about 6/7 seconds) for one cycle of walking. Acceleration values for about 6/7 seconds are extracted respectively. Such an acceleration value forms a walking waveform representing one step of walking. The initial feature determination unit 120 normalizes each of the seven walking waveforms, adds and averages the normalized seven walking waveforms, and calculates the average acceleration that is an input when calculating the feature amount.

(Step S30)
Next, the initial feature calculation unit 130 calculates the feature vector using the feature quantities of the plurality of features determined by using the behavior sensor values in step S20. The feature quantities of the plurality of features determined in step S20 are referred to as initial feature quantities. The feature vector contains a plurality of initial features as elements. For example, in the case of the example of FIG. 5, the feature vector V _X is the coefficient vector _VC . The coefficient vector VC may be ( _{C 1} , C ₂ , ..., C ₃₂ ).

(Step S40)
Further, the weight adjusting unit 140 calculates and adjusts the weight vector for estimating the athletic ability value acquired in step S10 by using the calculated feature vector.

Here, the athletic ability value y is estimated by an estimation formula f (V _X , V _W ) including a feature vector V _X and a weight vector V _W corresponding to the feature vector V _X. As a specific estimation method, a known learning method for estimating a correct answer value from a feature vector and a weight vector, such as linear regression, neural network, and logistic regression, is used.

The weight vector is V _W'using the gradient of the weight of the error function based on the error function E = (y-f (V _X , V _W )) ² representing the square of the estimation error of the athletic ability value. = V _W -a · ∂E / ∂V _W can be adjusted. V _W indicates the weight vector before adjustment, and V _W'indicates the weight vector after adjustment. ∂E / _∂VW is a partial derivative of the error function with respect to the weight vector, and indicates the weight gradient of the error function. a represents a constant. That is, when the exercise capacity value estimated by the estimation formula using the feature vector V _X and the weight vector V _W has an error with respect to the exercise capacity value acquired in step S10, the weight adjustment unit 140 has an error. The weight vector V _W is adjusted using the error function of. Here, the error function E is an example of an index showing an error between the estimated value of the athletic ability value estimated by using the weight vector and the acquired athletic ability value.

(Step S50)
The feature gradient calculation unit 150 uses the athletic ability value acquired in step S10, the feature vector V _X calculated in step S30, and the weight vector V _W'adjusted in step S40 to obtain an error function related to the feature vector. The gradient ∂E / ∂V _X (also called the feature gradient vector) of E is calculated. In the gradient ∂E / ∂V _X , V _X is a feature vector. Further, the gradient ∂E / _∂VX is an example of a gradient vector relating to a feature vector for an index showing an error between the estimated value of the athletic ability value estimated using the weight vector and the acquired athletic ability value. ..

Using this gradient ∂E / ∂V _X , it is conceivable to update the feature vector V _X as shown by V _X '= V _X −b · ∂E / ∂V _X. V _X'indicates the updated feature vector and b is a constant. If the feature vector can be changed based on the above equation as in the weight vector, the error between the estimated value of the athletic ability value and the acquired athletic ability value can be reduced. However, since the feature amount, which is an element of the feature vector, is a value calculated from the behavior sensor value, it cannot be directly changed like the weight vector. Therefore, the feature vector is changed by changing the time interval for calculating the feature vector and generating the feature vector having the feature amount calculated in the changed time interval as an element. When the feature vector is changed in this way, the feature vector is updated in the direction of the gradient _∂E / ∂VX so that the above error becomes small. Here, the conventional machine learning is a configuration in which only the weight vector is learned and updated so that an appropriate result can be obtained. However, in the estimation information generation device 100, the adjustment of the weight vector for improving the estimation accuracy and the search for the feature vector described later in step S70 can be realized at the same time by the above-mentioned processing. This is not possible with conventional machine learning. Then, the adjustment of the weight vector and the realization of the search of the feature vector are particularly useful when the estimation is performed assuming the input of long-time sensor data or the like whose useful features are not known in advance.

(Step S60)
The feature candidate calculation unit 160 calculates a feature candidate vector using a new time interval and a new feature corresponding to the new time interval. Here, the new time interval and the new feature mean that the time interval and the feature determined by the initial feature determination unit 120 and at least one of the time interval and the feature are different features and time intervals. Specifically, the feature candidate calculation unit 160 determines a new time interval in which the time interval of at least one feature is changed, and calculates the feature amount of the feature using the behavior sensor value in the determined time interval. .. Then, the feature candidate calculation unit 160 calculates a new feature vector including the calculated feature amount as an element as a feature candidate vector. The feature candidate calculation unit 160 may arbitrarily determine a new time interval.

For example, in the example of FIG. 5, the feature candidate calculation unit 160 determines a candidate for a new time interval using the feature of the feature number n (1 ≦ n ≦ 32). For example, the feature candidate calculation unit 160 may use the feature number n as a candidate for a new time interval, which is a time interval in which the current start time nt and the end time nt + 30 are shifted by a fixed time. For example, assuming that the time interval for shifting from the current start time tun of the feature number n is 15 seconds, "tn-30 to tun", "tun-15 to tun + 15", "tun to tun + 30", and "tn + 15". Time intervals such as "tn + 45" and "tn + 30 to tn + 60" can be candidates.

The feature candidate calculation unit 160 analyzes the PCA for the behavior sensor value of the new time interval for each of the new time intervals determined as candidates. Then, the feature candidate calculation unit 160 calculates the coefficients of the first principal component to the 32nd principal component when the plurality of behavior sensor values in the new time interval are expressed by the principal component vector of the PCA. Assuming that the number of candidates for the new time interval is M, M × 32 values are calculated as the coefficient of the feature number n, that is, as a new candidate value for the feature amount. Specifically, since the feature quantity of the feature of the feature number n can take any of the coefficients of the first principal component to the 32nd principal component, 32 kinds of coefficients can be taken. Further, the 32 feature quantities of the feature number n are calculated for M new time intervals. Therefore, M × 32 values are calculated as new candidate values for the feature amount.

The feature candidate calculation unit 160 generates a feature candidate vector for the feature number n by arranging the values of these coefficients, that is, the feature quantities, for all the behavior sensor values acquired by the data acquisition unit 110. do. Since the number of new candidate values for the feature amount is M × 32, M × 32 feature candidate vectors can be obtained. Then, when the number of behavior sensor values acquired by the data acquisition unit 110 is N, the feature candidate vector has N elements. The feature candidate calculation unit 160 calculates such a feature candidate vector for all feature numbers (feature numbers 1 to 32 in this example).

(Step S70)
The feature search unit 170 searches for the most suitable feature candidate vector for moving the current feature vector in the direction of the feature gradient vector by using the plurality of feature candidate vectors calculated in step S60. The feature search unit 170 searches for a feature candidate vector that satisfies a predetermined condition based on the feature gradient vector. Specifically, when the feature search unit 170 searches for the feature candidate vector of the nth feature, the current nth behavior sensor value acquired by the data acquisition unit 110 is analyzed by PCA. An N-dimensional vector in which the features that are the values of the coefficients are arranged, that is, a feature vector is used. Then, the feature search unit 170 has a positive / negative sign of each element of the subtracted vector formed after the subtraction when the feature candidate vector is subtracted from the feature vector, and a positive / negative sign of each corresponding element of the feature gradient vector. Judge the degree of agreement with. The feature search unit 170 selects the feature candidate vector used for calculating the subtracted vector whose positive and negative signs of each element of the subtracted vector best match the positive and negative signs of the corresponding elements of the feature gradient vector.

Then, the feature search unit 170 updates, that is, replaces the feature amount and the time interval of the current feature vector by using the feature amount and the time interval corresponding to the feature amount and the feature amount of the selected feature candidate vector. If there are only feature candidate vectors that do not match the positive and negative signs of both elements between the subtracted vector and the feature candidate vector, the feature search unit 170 will perform the feature search unit 170. , It is determined that the feature amount that can be appropriately updated, that is, the feature candidate vector is not found, and the update process is not performed. The feature search unit 170 performs the above processing for all feature numbers (1 to 32 in this embodiment). Therefore, the feature search unit 170 updates the feature amount and the time interval of the current feature vector for each of the features of the feature numbers 1 to 32, using the time interval corresponding to the feature amount and the feature amount of the selected feature candidate vector. do.

For example, referring to FIG. 7, the feature search unit 170 shows an example of the time interval and the feature updated from the time interval and the feature amount shown in FIG. For example, before and after the update, the feature of feature number 1 maintains the coefficient of the first principal component. In the feature of feature number 1, the time interval is not changed. Further, the feature of feature number 2 changes from the coefficient of the second principal component to the coefficient of the fifth principal component. In the feature of feature number 2, the time interval is shifted by 15 seconds, and the time interval of 0 to 30 seconds is changed to the time interval of 15 seconds to 45 seconds.

In this way, the difference between the feature candidate vector and the feature vector, that is, the difference obtained by subtracting the feature candidate vector from the feature vector is calculated, and the feature candidate vector satisfying the condition that the degree of coincidence between this difference and the feature gradient vector is the highest. Is elected. In other words, a feature candidate vector satisfying a predetermined condition based on the gradient vector is determined based on the difference between the feature candidate vector and the feature vector. As will be described later, the weight vector is adjusted based on the determined feature candidate vector, and the weight vector based on such a feature candidate vector is larger than the weight vector based on the feature vector before the change to the feature candidate vector. , The estimation error of the athletic ability value can be reduced.

(Step S80)
The determination unit 180 determines whether or not to end the learning for generating the motor function estimation knowledge. Specifically, the determination unit 180 determines whether or not the feature amount has been updated in step S70, and if it is updated (No in step S80), determines that learning has not been completed and performs the process of step S40. return. As a result, the estimation information generation device 100 readjusts the weight vector using the updated feature amount and time interval. If the feature amount is not updated in step S70 (Yes in step S80), the determination unit 180 determines that learning has been completed, and proceeds to the process of step S90. The determination unit 180 may also use the number of repetitions of the determination in the preset step S80 for the determination of the end of learning, and the feature candidate vector that changes the feature vector in the direction of the feature gradient vector in step S70. If is not found, it may be considered that the learning is completed.

(Step S90)
The estimation knowledge output unit 190 stores the new time interval and the new feature quantity, that is, the new feature vector calculated in step S70 in the first storage unit 200. Specifically, the estimation knowledge output unit 190 stores in the first storage unit 200 what kind of feature the new feature is, that is, the feature corresponding to the new feature. For example, in the example of FIG. 7, for the second feature, the time interval of 15 seconds to 45 seconds and the feature of the coefficient of the fifth principal component are stored in combination. At this time, the estimation knowledge output unit 190 stores the new time interval and the new feature by associating the parameters of the measured values of the behavior sensor values used for calculating the new time interval and the new feature amount. May be good. Further, the estimation knowledge output unit 190 stores the weight vector acquired by the weight adjustment unit 140 in the second storage unit 210. This weight vector is the newest weight vector updated in step S40. This weight vector may be associated with the new time interval and new feature described above. If the weight vector stored in the second storage unit 210 is only the newest weight vector, the above correspondence may not be necessary.

Through the above steps S10 to S90, the estimation information generation device 100 calculates the time interval, the feature, and the weight vector required for estimation as estimation knowledge in order to estimate the motor function from the behavior sensor value at a predetermined time. be able to.

As a result, as shown in the conceptual diagram shown in FIG. 8, it is possible to realize learning including a long-time behavior sensor value, a time interval and a feature of interest. Note that FIG. 8 is a conceptual diagram of an example of the estimation method of the estimation knowledge of the embodiment, and the motor ability value y _i of the person i is estimated by inputting the time interval and the feature amount calculated from the time interval. The model to be used is shown. In FIG. 8, the input feature amount is indicated by the feature vector V _Xi , and the processing between the nodes is weighted by the weight vector V _W. For example, in FIG. 8, in the case of conventional machine learning in which only the weight vector is updated, if the time interval for extracting the feature and the feature calculated in the time interval are predetermined, they cannot be changed. In the case of conventional machine learning, for example, when the time intervals are 1, 2 and 4 in FIG. 8 and the features calculated in each of the time intervals 1, 2 and 4 are A, C and D, these time intervals are used. And features are always fixed during learning. On the other hand, in the technique of the present embodiment, learning can be performed while updating the time interval and the feature suitable for estimating the athletic ability value. FIG. 8 shows an example in which the feature B calculated from the time interval 2 and the time interval 2 is updated to the time interval 3 and the feature C.

As an effect of the technique of the present embodiment, in the conventional machine learning, when the data of a huge number of dimensions is input as it is, the number of parameters to be learned becomes too large for the number of data, and the estimation accuracy and the identification accuracy are lowered. Feature extraction based on heuristics is essential. On the other hand, in the estimation knowledge construction method described in the present embodiment, since the search for the feature vector changing in the direction of the feature gradient vector and the adjustment of the weighting vector can be performed at the same time, a huge amount of long-time sensor data or the like can be obtained. Effective learning can be performed without prior knowledge for data with a large number of dimensions. The learning portion of the weight vector may be a multi-layer neural network as shown in FIG.

In the example of FIG. 9, the neural network is one of machine learning. A neural network is an information processing model modeled on the cranial nerve system. The neural network is composed of a plurality of node layers including an input layer and an output layer. The node layer contains one or more nodes. The model information of the neural network indicates the number of node layers constituting the neural network, the number of nodes included in each node layer, and the entire neural network or the type of each node layer. The neural network is composed of, for example, an input layer, an intermediate layer and an output layer. In the neural network, the information input to the node of the input layer is output processed from the input layer to the first intermediate layer, processed in the first intermediate layer, output processed from the first intermediate layer to the second intermediate layer, and the first The processing in the second intermediate layer, the output processing from the second intermediate layer to the output layer, and the processing in the output layer are sequentially performed, and the output result matching the input information is output. Each node of one layer is connected to each node of the next layer, and the connection between the nodes is weighted. The information of the node of one layer is given the weight of the connection between the nodes and output to the node of the next layer. These weighting elements form a weight vector _VW . In the learning of the neural network, when the feature amount calculated in the time interval and the time interval, that is, the feature vector V _Xi corresponding to the time interval is input to the input layer, the appropriate motor ability value y _i is obtained in the output layer. The weighting between the nodes is adjusted so that it is output.

[2-2. Processing operation of motor function estimation by the estimation device]
With reference to FIGS. 1 and 10, the processing operation of motor function estimation by the estimation device 300 will be described. Note that FIG. 10 is a flowchart showing an example of the operation flow of the motor function estimation in the motor function estimation system 1000 according to the embodiment.

(Step S100)
The sensor value acquisition unit 310 acquires the behavior sensor value over a predetermined time measured by the sensor 400 attached to the user. The behavior sensor value includes the measured values such as acceleration, heart rate, body temperature, and angular velocity in association with the measured time of the measured values.

(Step S110)
The feature acquisition unit 320 acquires a combination of time intervals and features to be extracted from the behavior sensor value from the first storage unit 200 in order to extract the feature amount used for motor function estimation from the behavior sensor value. The combination of the acquired time interval and the feature is stored in the first storage unit 200 in step S90 of the operation of generating the estimated information. The feature acquisition unit 320 acquires a combination of features and time intervals corresponding to the parameters of the measured values of the behavior sensor values.

(Step S120)
The feature amount calculation unit 330 extracts the time interval corresponding to the time interval acquired in step S110 from the behavior sensor value acquired in step S100, and further acquires it in step S110 from the measurement result of the extracted time interval. The feature amount of the created feature is calculated. Then, the feature amount calculation unit 330 calculates a feature vector having the calculated feature amount as an element.

(Step S130)
The estimation knowledge acquisition unit 340 acquires a weight vector that associates the feature vector with the athletic ability value from the second storage unit 210. The acquired weight vector is stored in the second storage unit 210 in step S90 of the operation of generating the estimation information. When a plurality of weight vectors are stored in the second storage unit 210, the estimation knowledge acquisition unit 340 determines the weight vector to be acquired based on the correspondence with the combination of the feature and the time interval acquired in step S110. You may. As a result, the estimated knowledge acquisition unit 340 acquires the estimated knowledge for calculating the athletic ability value constituting the motor function.

(Step S140)
The motor function estimation unit 350 estimates the motor performance value from the feature vector calculated in step S120 and the weight vector acquired in step S130. For the estimation, the estimation knowledge composed of the weight vectors is used, and further, the processing operation in the estimation information generation device 100, that is, the same estimation method as the method used in machine learning is used. As the estimation method, the estimation method exemplified in FIG. 8 may be used, or the neural network exemplified in FIG. 9 may be used. By inputting the feature vector to the estimated knowledge to which the weight vector is applied, the estimated value of the athletic ability is output. Further, the motor function estimation unit 350 outputs the estimated motor performance value to the output unit 360.

(Step S150)
The output unit 360 outputs the athletic ability value estimated in step S140. The output destination may be the display unit 500, and the display unit 500 may directly display the athletic ability value. Further, the output destination may be a storage medium such as a memory, and the storage medium may store the athletic ability value.

By the processing of the above steps S110 to S150, the time interval and the feature useful for estimation are used, and thereby the estimation of the motor ability value with high accuracy from the behavior sensor value for a predetermined time even for a long time is performed. It will be possible to realize.

[Example]
An embodiment of the embodiment will be described. Specifically, the estimation result of the motor function by the estimation method in the motor function estimation system 1000 according to the embodiment is examined, and further, such an estimation result and the estimation result of the motor function by another estimation method are examined. Compared.

For example, referring to FIG. 11, the result of estimating the knee extension muscle strength of the subject using the estimation method of the present embodiment is shown. Note that FIG. 11 is a diagram showing the correlation between the measured value and the estimated value of the knee extension muscle strength of the subject. The number of subjects in the experiment was 102. In this experiment, neural network regression having one intermediate layer of 32 nodes was used for the estimation knowledge. The relationship between the measured value and the estimated value of knee extension muscle strength was evaluated by 6-fold cross validation (6-fold cross-validation). As a result of the evaluation, R = 0.85 was obtained as the correlation coefficient between the actually measured value and the estimated value by the estimation method of the present embodiment. From such a correlation coefficient, it can be seen that the motor function can be estimated with high accuracy.

For example, in the experimental results when the technique described in Patent Document 1 is used, the correlation coefficient R between the index used for estimation and the measured value is about R = 0.65 at the maximum. It can be said that the estimation method of the present embodiment has realized the estimation of the motor function with high accuracy.

For example, FIG. 12 shows a result of comparing the estimation result of the athletic ability value by the estimation method according to the embodiment with the estimation result of the athletic ability value by another method. FIG. 12 shows the correlation coefficient between the estimated value and the actually measured value and the average error rate of the error between the estimated value and the actually measured value in association with the estimation method. As shown in FIG. 12, a 6-minute walk test and a 10 m walk test were used as other estimation methods to be compared with the embodiments. The 6-minute walking test and the 10-m walking test are clinically commonly performed tests, and the 6-minute walking distance and the 10-m walking speed are used as simple indexes of the athletic ability value, respectively. be. The 10m walking test is a method for simply evaluating an athletic ability value by measuring a walking speed of 10m.

Since the method of Patent Document 1 uses acceleration in the traveling direction, it has a meaning close to the above-mentioned 6-minute walking distance. For the fixed feature amount (PCA of the entire walking) in Patent Document 1, the average walking waveform is calculated from the entire walking for 6 minutes without determining the point of interest as in the estimation method according to the embodiment, and the fixed feature amount is determined by the PCA. This is the calculated method.

Then, in FIG. 12, as a comparative method, an estimation method for estimating lower limb muscle strength by neural network regression with each of "6 minutes walking distance and 10 m walking speed" and "fixed feature amount (total walking PCA)" as feature amounts is used. There is. The results of the estimation method using "6 minutes walking distance and 10 m walking speed" as features are shown in the upper part of FIG. The result of the estimation method using the "fixed feature amount (PCA of the whole walking)" as the feature amount is shown in the middle part of FIG. The result of the estimation method of the present embodiment shown in the lower part of FIG. 12 is the result of estimating the motor function from the behavior sensor value such as acceleration measured from the walking test for 6 minutes.

From the results shown in FIG. 12, the method for estimating the motor ability value according to the embodiment has obtained an estimation accuracy higher than that of the comparison method, and it is possible to estimate the motor function with high accuracy from the low-load motion of walking. Recognize.

[others]
Although the motor function estimation system and the like according to one or more embodiments have been described above based on the embodiments, the present disclosure is not limited to the embodiments. As long as it does not deviate from the gist of the present disclosure, various modifications that can be conceived by those skilled in the art are applied to the present embodiment, and a form constructed by combining components in different embodiments is also within the scope of one or more embodiments. May be included within.

For example, the motor function estimation system 1000 according to the present disclosure estimates the motor performance value of the lower limbs of the user by using an acceleration sensor, a heart rate sensor, a temperature sensor, and an angular velocity sensor, but the present invention is not limited to this. The sensor may be attached to any movable part of the user such as an arm, whereby the motor function estimation system 1000 may estimate the motor ability value for the movable part.

For example, in the motor function estimation system 1000 according to the present disclosure, the estimation information generation device 100 constructs the estimation knowledge by using the behavior sensor value and the motor ability value of the user, but the data used for constructing the estimation knowledge is used. The target may be a specific user or a plurality of users. The plurality of users may be an unspecified number of users. The estimated knowledge constructed using the data of one specific user is the estimated knowledge dedicated to the user, and can output the athletic ability value reflecting the characteristics of the user. The estimation knowledge constructed using the data of an unspecified number of users is general-purpose estimation knowledge, and can output an athletic ability value that is not affected by the characteristics peculiar to a specific user.

In the present disclosure, all or part of a unit, device, member or part, or all or part of a functional block in a block diagram as shown in FIG. 1, is a semiconductor device, a semiconductor integrated circuit (IC, Integration Circuit). Alternatively, it may be executed by one or more electronic circuits including an LSI (Large Scale Integration). The LSI or IC may be integrated on one chip, or may be configured by combining a plurality of chips. For example, functional blocks other than storage elements such as storage units may be integrated on one chip. Here, it is called LSI or IC, but the name changes depending on the degree of integration, and it may be called system LSI, VLSI (Very Large Scale Integration), or ULSI (Ultra Large Scale Integration). A Field Programmable Gate Array (FPGA) programmed after the LSI is manufactured, or a Reconfigurable Logic Device (RLD) that can reconfigure the junction relationship inside the LSI or set up the circuit partition inside the LSI can also be used for the same purpose.

Further, all or part of the function or operation of a unit, device, member or part can be performed by software processing. In this case, the software is recorded on a non-temporary recording medium such as one or more ROMs, optical disks, hard disk drives, and is identified by the software when it is run by a processor. Functions are performed by processors and peripherals. The system or device may include one or more non-temporary recording media on which the software is recorded, a processor, and the required hardware device, such as an interface.

In the present disclosure, it is possible to easily and highly accurately estimate the motor ability by calculating the time interval and the feature effective for the estimation from the behavior sensor value of the predetermined time worn on the user's body. The techniques of the present disclosure are applicable to any application for achieving the above.

100 Estimated information generator 110 Data acquisition unit 120 Initial feature determination unit 130 Initial feature calculation unit 140 Weight adjustment unit 150 Feature gradient calculation unit 160 Feature candidate calculation unit 170 Feature search unit 180 Judgment unit 190 Estimated knowledge output unit 200 First storage unit 210 Second storage unit 300 Estimator 310 Sensor value acquisition unit 320 Feature acquisition unit 330 Feature amount calculation unit 340 Estimated knowledge acquisition unit 350 Motor function estimation unit 360 Output unit

Claims (9)

A sensor that measures at least one selected from the user's acceleration, heart rate, body temperature, and angular velocity at a given time.
Equipped with a first processing circuit
The first processing circuit is
(A1) The user's muscular strength, endurance, or balance ability when at least one sensor value selected from the user's acceleration, heart rate, body temperature, and angular velocity is acquired and the at least one sensor value is acquired. Obtain the first athletic ability value for
(A2) The characteristics of the sensor value in time change and the time interval within the predetermined time are determined.
(A3) Using the sensor value, a feature vector corresponding to the feature amount of the feature in the time interval is calculated.
(A4) A coefficient indicating the weight of an element of the feature vector included in an estimation formula for estimating the second motor ability value of the user based on machine learning by inputting the feature vector and the first motor ability value. Get the first weight vector as an element
(A5) By partially differentiating the estimation error in the estimation of the second motor ability value with the feature vector, a gradient vector showing the gradient of the estimation error with respect to the feature vector is calculated.
(A6) Within the predetermined time, a new time interval that is an arbitrary time interval different from the time interval and a new feature that is a feature of the sensor value in the new time interval are determined.
(A7) A feature candidate vector, which is a candidate of the feature vector corresponding to the feature amount of the new feature in the new time interval, is calculated.
(A8) The feature candidate vector satisfying a predetermined condition that the difference between the feature candidate vector and the feature vector and the degree of coincidence with the gradient vector is the highest is determined.
(A9) The weight of the element of the feature vector included in the estimation formula for estimating the second motor ability value of the user estimated from the feature candidate vector satisfying the predetermined condition and the first motor ability value. A second weight vector having the indicated coefficient as an element is acquired, and the first weight vector is modified to the second weight vector.
(A10) The feature and the time interval corresponding to the feature candidate vector satisfying the predetermined condition, and the second weight vector are stored in the storage unit.
Motor function estimation information generator. The weight vector constitutes a neural network and weights between the nodes of the neural network.
The motor function estimation information generation device according to claim 1. The motor function estimation information generator according to claim 1 and
Equipped with a second processing circuit
The second processing circuit is
Acquires at least one sensor value selected from the user's acceleration, heart rate, body temperature, and angular velocity,
An estimated feature vector is calculated based on the correspondence between the feature and the time interval stored in the storage unit, the feature in the temporal change of the sensor value, and the time interval from which the sensor value is acquired .
Using the second weight vector stored in the storage unit and the estimated feature vector as inputs, the motor ability values related to muscle strength, endurance, or balance ability are estimated based on machine learning .
Output the athletic ability value,
Motor function estimation system. The processor,
(B1) The user's muscular strength, endurance, or balance ability when at least one sensor value selected from the user's acceleration, heart rate, body temperature, and angular velocity is acquired and the at least one sensor value is acquired. Obtain the first athletic ability value,
(B2) The characteristics of the sensor value in time change and the time interval within a predetermined time are determined.
(B3) Using the sensor value, a feature vector corresponding to the feature amount of the feature in the time interval is calculated.
(B4) A coefficient indicating the weight of an element of the feature vector included in an estimation formula for estimating the second motor ability value of the user based on machine learning by inputting the feature vector and the first motor ability value. Get the first weight vector as an element
(B5) By partially differentiating the estimation error in the estimation of the second motor ability value with the feature vector, a gradient vector showing the gradient of the estimation error with respect to the feature vector is calculated.
(B6) Within the predetermined time, a new time interval that is an arbitrary time interval different from the time interval and a new feature that is a feature of the sensor value in the new time interval are determined.
(B7) A feature candidate vector, which is a candidate of the feature vector corresponding to the feature amount of the new feature in the new time interval, is calculated.
(B8) The feature candidate vector satisfying a predetermined condition that the difference between the feature candidate vector and the feature vector and the degree of coincidence with the gradient vector is the highest is determined.
(B9) The weight of the element of the feature vector included in the estimation formula for estimating the second motor ability value of the user estimated from the feature candidate vector satisfying the predetermined condition and the first motor ability value. A second weight vector having the indicated coefficient as an element is acquired, and the first weight vector is modified to the second weight vector.
(B10) The feature and the time interval corresponding to the feature candidate vector satisfying the predetermined condition, and the second weight vector are stored in the storage unit.
Motor function estimation information generation method. The weight vector constitutes a neural network and weights between the nodes of the neural network.
The method for generating motor function estimation information according to claim 4. The processor,
(B1) The user's muscular strength, endurance, or balance ability when at least one sensor value selected from the user's acceleration, heart rate, body temperature, and angular velocity is acquired and the at least one sensor value is acquired. Obtain the first athletic ability value,
(B2) The characteristics of the sensor value in time change and the time interval within a predetermined time are determined.
(B3) Using the sensor value, a feature vector corresponding to the feature amount of the feature in the time interval is calculated.
(B4) A coefficient indicating the weight of the element of the feature vector included in the estimation formula for estimating the second motor ability value of the user based on machine learning by inputting the feature vector and the first motor ability value. Get the first weight vector as an element
(B5) By partially differentiating the estimation error in the estimation of the second motor ability value with the feature vector, a gradient vector showing the gradient of the estimation error with respect to the feature vector is calculated.
(B6) Within the predetermined time, a new time interval which is an arbitrary time interval different from the time interval and a new feature which is a feature of the sensor value in the new time interval are determined.
(B7) A feature candidate vector, which is a candidate of the feature vector corresponding to the feature amount of the new feature in the new time interval, is calculated.
(B8) The feature candidate vector satisfying a predetermined condition that the difference between the feature candidate vector and the feature vector and the degree of coincidence with the gradient vector is the highest is determined.
(B9) The weight of the element of the feature vector included in the estimation formula for estimating the second motor ability value of the user estimated from the feature candidate vector satisfying the predetermined condition and the first motor ability value. A second weight vector having the indicated coefficient as an element is acquired, and the first weight vector is modified to the second weight vector.
(B10) The feature and the time interval corresponding to the feature candidate vector satisfying the predetermined condition, and the second weight vector are stored in the storage unit.
(B11) Further, at least one sensor value selected from the user's acceleration, heart rate, body temperature, and angular velocity is acquired, and the user's acceleration, heart rate, body temperature, and angular velocity are acquired.
(B12) An estimated feature vector is calculated based on the correspondence between the feature and the time interval stored in the storage unit, the feature in the temporal change of the sensor value, and the time interval in which the sensor value is acquired. death,
(B13) Using the second weight vector stored in the storage unit and the estimated feature vector as inputs, the motor ability values related to muscle strength, endurance, or balance ability are estimated based on machine learning .
Motor function estimation method. A recording medium comprising a control program for causing a device including a processor to execute a process, wherein the recording medium is non-volatile and can be read by a computer.
(B1) The user's muscular strength, endurance or balance ability when at least one sensor value selected from the user's acceleration, heart rate, body temperature, and angular velocity is acquired and the at least one sensor value is acquired. Get the first athletic ability value,
(B2) The characteristics of the sensor value in time change and the time interval within a predetermined time are determined.
(B3) Using the sensor value, a feature vector corresponding to the feature amount of the feature in the time interval is calculated.
(B4) A coefficient indicating the weight of the element of the feature vector included in the estimation formula for estimating the second motor ability value of the user based on machine learning by inputting the feature vector and the first motor ability value. Get the first weight vector as an element ,
(B5) By partially differentiating the estimation error in the estimation of the second motor ability value with the feature vector, a gradient vector showing the gradient of the estimation error with respect to the feature vector is calculated.
(B6) Within the predetermined time, a new time interval that is an arbitrary time interval different from the time interval and a new feature that is a feature of the sensor value in the new time interval are determined.
(B7) The feature candidate vector, which is a candidate of the feature vector corresponding to the feature amount of the new feature in the new time interval, is calculated.
(B8) The feature candidate vector satisfying a predetermined condition that the difference between the feature candidate vector and the feature vector and the degree of coincidence with the gradient vector is the highest is determined.
(B9) The weight of the element of the feature vector included in the estimation formula for estimating the second motor ability value of the user estimated from the feature candidate vector satisfying the predetermined condition and the first motor ability value. A second weight vector having the indicated coefficient as an element is acquired, and the first weight vector is corrected to the second weight vector.
(B10) The feature and the time interval corresponding to the feature candidate vector satisfying the predetermined condition, and the second weight vector are stored in the storage unit.
Recording medium including that. The weight vector constitutes a neural network and weights between the nodes of the neural network.
The recording medium according to claim 7. A recording medium comprising a control program for causing a device including a processor to execute a process, wherein the recording medium is non-volatile and can be read by a computer.
(B1) The user's muscular strength, endurance or balance ability when at least one sensor value selected from the user's acceleration, heart rate, body temperature, and angular velocity is acquired and the at least one sensor value is acquired. Get the first athletic ability value,
(B2) The characteristics of the sensor value in time change and the time interval within a predetermined time are determined.
(B3) Using the sensor value, a feature vector corresponding to the feature amount of the feature in the time interval is calculated.
(B4) A coefficient indicating the weight of the element of the feature vector included in the estimation formula for estimating the second motor ability value of the user based on machine learning by inputting the feature vector and the first motor ability value. Get the first weight vector as an element ,
(B5) By partially differentiating the estimation error in the estimation of the second motor ability value with the feature vector, a gradient vector showing the gradient of the estimation error with respect to the feature vector is calculated.
(B6) Within the predetermined time, a new time interval that is an arbitrary time interval different from the time interval and a new feature that is a feature of the sensor value in the new time interval are determined.
(B7) The feature candidate vector, which is a candidate of the feature vector corresponding to the feature amount of the new feature in the new time interval, is calculated.
(B8) The feature candidate vector satisfying a predetermined condition that the difference between the feature candidate vector and the feature vector and the degree of coincidence with the gradient vector is the highest is determined.
(B9) The weight of the element of the feature vector included in the estimation formula for estimating the second motor ability value of the user estimated from the feature candidate vector satisfying the predetermined condition and the first motor ability value. A second weight vector having the indicated coefficient as an element is acquired, and the first weight vector is corrected to the second weight vector.
(B10) The feature and the time interval corresponding to the feature candidate vector satisfying the predetermined condition, and the second weight vector are stored in the storage unit.
(B11) Further, at least one sensor value selected from the user's acceleration, heart rate, body temperature, and angular velocity is acquired.
(B12) An estimated feature vector is calculated based on the correspondence between the feature and the time interval stored in the storage unit, the feature in the temporal change of the sensor value, and the time interval in which the sensor value is acquired. Let me
(B13) Using the second weight vector stored in the storage unit and the estimated feature vector as inputs, the motor ability values related to muscle strength, endurance, or balance ability are estimated based on machine learning .
Recording medium including that.

Images