Action state estimation device, action state estimation method, action state learning device, and action state learning method

Abstract

An action state estimation device (10) includes a sampling unit (11), an action state model storage unit (13B), and an estimation operation unit (14B). The sampling unit (11) samples a displacement measurement signal in a predetermined time period to generate displacement measurement data. The action state model storage unit (13B) stores an action state model that models a relationship between the displacement measurement data and a load state of a desired muscle. The estimation operation unit (14B) estimates the load state using the action state model with the displacement measurement data as an input vector.

Description

Technical Field

[0001] This invention relates to a technique for inferring action states including muscle load states based on tremor detection results, and a technique for generating action state models for this inference technique (action state learning technique). Background Technology

[0002] Patent Document 1 describes an action state estimation device, etc. The action state estimation device described in Patent Document 1 converts the measurement signal from a displacement detection sensor into frequency components. The action state estimation device described in Patent Document 1 estimates the action state based on the components of a predetermined frequency band.

[0003] Patent Document 1: Japanese Patent Application Publication No. 2011-182824

[0004] However, in the existing apparatus and method shown in Patent Document 1, it is necessary to convert the sensor's measurement signal into frequency components. Therefore, the processing load for generating the inference signal increases. Summary of the Invention

[0005] Therefore, the object of the present invention is to provide an action state prediction technique that achieves the desired prediction accuracy and suppresses the processing load.

[0006] The motion state estimation device of the present invention includes a first sampling unit, a motion state model storage unit, and an estimation calculation unit. The first sampling unit samples displacement measurement signals within a predetermined time to generate displacement measurement data. The motion state model storage unit stores a motion state model obtained by establishing a correlation between the displacement measurement data and the desired muscle load state. The estimation calculation unit uses the displacement measurement data as an input vector and the motion state model to estimate the load state.

[0007] In this structure, displacement measurement data is used directly to infer the load condition.

[0008] According to the present invention, the desired prediction accuracy can be achieved, and the processing load can be suppressed. Attached Figure Description

[0009] Figure 1 This is a functional block diagram of the action state prediction device according to the first embodiment of the present invention.

[0010] Figure 2 This is a table that sets an example of the importance of the first embodiment.

[0011] Figure 3 This is a flowchart illustrating the main processing steps of the action state prediction method according to the first embodiment of the present invention.

[0012] Figure 4This is a functional block diagram of the action state learning device according to the first embodiment of the present invention.

[0013] Figure 5 This is a flowchart illustrating the main processing steps of the action state learning method according to the first embodiment of the present invention.

[0014] Figure 6 (A) represents the time variation of signal strength. Figure 6 (B) represents the signal strength distribution. Figure 6 (C) represents the intensity block data.

[0015] Figure 7 This is a table that sets an example of the importance of the second implementation method.

[0016] Figure 8 (A) is a graph representing an example of the time-varying signal strength. Figure 8 (B) is a graph representing the intensity block data (average). Figure 8 (C) is a graph representing the intensity block data (cumulative value).

[0017] Figure 9 This is a functional block diagram of the action state prediction device according to the third embodiment of the present invention.

[0018] Figure 10 (A) Figure 10 (B) is a table representing a setting example of the importance of the third embodiment.

[0019] Figure 11 This is a flowchart illustrating the main processing steps of the action state prediction method according to the third embodiment of the present invention.

[0020] Figure 12 This is a functional block diagram of the action state learning device according to the third embodiment of the present invention.

[0021] Figure 13 This is a flowchart illustrating the main processing steps of the action state learning method according to the third embodiment of the present invention.

[0022] Figure 14 (A) Figure 14 (B) is a table representing a setting example of the importance of the fourth embodiment.

[0023] Figure 15 This is a flowchart illustrating the main processing steps of the action state learning method according to the fifth embodiment of the present invention.

[0024] Figure 16 This is a diagram representing the concept of synchronization.

[0025] Figure 17This is a functional block diagram of the action state prediction device according to the sixth embodiment of the present invention.

[0026] Figure 18 (A) is a diagram illustrating a waveform used to explain the importance of a time range, and is an example of a time range setting. Figure 18 (B) is a table representing an example of a setting based on the importance of a time range.

[0027] Figure 19 This is a flowchart illustrating the main processing steps of the action state prediction method according to the sixth embodiment of the present invention.

[0028] Figure 20 This is a functional block diagram of the action state learning device according to the sixth embodiment of the present invention.

[0029] Figure 21 This is a flowchart illustrating the main processing steps of the action state learning method according to the sixth embodiment of the present invention.

[0030] Figure 22 This is a functional block diagram of the action state prediction device according to the seventh embodiment of the present invention.

[0031] Figure 23 This is a flowchart illustrating the main processing steps of the action state prediction method according to the seventh embodiment of the present invention.

[0032] Figure 24 This is a functional block diagram of the action state learning device according to the seventh embodiment of the present invention.

[0033] Figure 25 This is a flowchart illustrating the main processing steps of the action state learning method according to the seventh embodiment of the present invention. Detailed Implementation

[0034] (First Implementation)

[0035] The action state prediction technique and action state model generation technique of the first embodiment of the present invention will be described with reference to the accompanying drawings.

[0036] (Structure and processing of the action status prediction device)

[0037] Figure 1 This is a functional block diagram of the action state prediction device according to the first embodiment of the present invention. Figure 1 As shown, the action state prediction device 10 includes a sampling unit 11, a statistical calculation unit 12, a prediction calculation unit 14, and an action state model storage unit 13. Each functional unit constituting the action state prediction device 10 can be implemented by electronic circuits, ICs, a storage medium storing a program that executes the functions of each functional unit, and an arithmetic processing device (CPU, etc.) that executes the program.

[0038] A displacement measurement signal is input to the sampling unit 11 from the displacement detection sensor 101. The sampling unit 11 samples the displacement measurement signal at a predetermined sampling frequency (e.g., 100Hz), thereby generating displacement measurement data. That is, the sampling unit 11 generates displacement measurement data without converting the displacement measurement signal into frequency components. The sampling unit 11 outputs the displacement measurement data to the statistical calculation unit 12.

[0039] Furthermore, the displacement detection sensor 101 is implemented using a piezoelectric sensor, an accelerometer, or the like. The displacement detection sensor 101 does not need to be positioned at the location of the muscle under the inferred load state; it only needs to be positioned at a location capable of measuring the tremors produced by the muscles of the inferred object. Additionally, the displacement detection sensor 101 can be a single sensor positioned at one location, or multiple sensors positioned at multiple locations. The displacement detection sensor 101 generates and outputs a displacement measurement signal. The displacement measurement signal is a signal that converts the displacement of the skin surface caused by tremors and deformation into a voltage signal.

[0040] The tremor described herein refers, for example, to involuntary movements indicating regular muscle activity. That is, the tremor of this invention is a subtle and rapid postural tremor that can be observed by a normal person, referred to as a physiological tremor, for example, with a frequency of 8 Hz to 12 Hz. Furthermore, movements observed in patients such as those with Parkinson's disease are pathological tremors, for example, with a frequency of 4 Hz to 7 Hz, and are not the subject of this invention. Using tremor provides various advantages over electromyography (EMG) in the following ways: For example, tremor can be detected (measured) even without direct attachment to the surface of the tested object (skin, etc.). Muscle contraction can be detected through tremor detection. Changes accompanying muscle fatigue can be detected through tremor detection.

[0041] The statistics calculation unit 12 calculates statistics based on displacement measurement data. The statistics calculation unit 12 calculates statistics based on multiple displacement measurement data within a specified period (e.g., 1 second).

[0042] As types of statistics, examples include the average, maximum, minimum, median, 1% value, 5% value, 25% value, 75% value, 95% value, 99% value, variance, skewness, kurtosis, and cumulative value. Furthermore, the types of statistics are not limited to these; any content obtainable from time-series displacement measurement data can be used, and other types are also possible. The statistics calculation unit 12 calculates various statistics from these. Additionally, the x% value refers to the value among multiple displacement measurement data within the period, counted in ascending order from the minimum value and located at x%.

[0043] The statistical calculation unit 12 outputs the calculated statistical quantities to the prediction calculation unit 14.

[0044] An action state model is stored in the action state model storage unit 13. The action state model sets the relationship between various statistics of displacement measurement data and the load state of the muscles of the predicted object. The action state model is, for example, pre-generated by the action state learning device 20 described later and stored in the action state model storage unit 13.

[0045] The prediction calculation unit 14 uses the action state model stored in the action state model storage unit 13, and takes multiple statistics as input vectors to predict the load state of the muscles of the prediction object. At this time, the prediction calculation unit 14 sets the importance of the statistics used for prediction based on the muscles of the prediction object. This importance is, for example, set in the action state model.

[0046] Figure 2 This is a table showing examples of settings indicating the importance of the first embodiment. Muscle M1, Muscle M2, Muscle M3, and Muscle M4 represent the types of muscles whose load state can be inferred from the measured displacement. For example, when the displacement detection sensor 101 is disposed at the tendon concentration area of ​​the ankle, more specifically at the front and back of the ankle, Muscle M1, Muscle M2, Muscle M3, and Muscle M4 can be set to the soleus, gastrocnemius, tibialis anterior, quadriceps, hamstring, etc. In addition, the statistics A1 to A15 are respectively set to any of the above-mentioned statistics (mean, maximum, minimum, median, 1% value, 5% value, 25% value, 75% value, 95% value, 99% value, variance, skewness, kurtosis, cumulative value, etc.).

[0047] For example, in Figure 2 In the case of muscle M1, the importance of the inference is set according to the following order: statistics A1, A2, A8, A12, and A7, in the order of 1st, 2nd, 3rd, 4th, and 5th place. Similarly, for muscle M2, the importance of the inference is set according to the following order: statistics A3, A5, A2, A14, and A15, in the order of 1st, 2nd, 3rd, 4th, and 5th place. The same applies to muscles M3 and M4. Figure 2 As shown, importance is set for the statistics.

[0048] If the prediction calculation unit 14 sets the muscle to be predicted, it uses the importance set according to the muscle to predict the muscle's load status (e.g., electromyographic value) based on multiple statistical quantities. Furthermore, the muscle's load status is not limited to electromyographic value; any parameter that can be expressed as a value can also be used.

[0049] More specifically, for example, the prediction calculation unit 14 calculates the prediction results of the load status of each statistic predicted based on the statistics with importance ranking from 1st to 5th and the action status model. Furthermore, the prediction calculation unit 14 weights each prediction result according to its importance, for example, by performing an additive average, thereby calculating the final prediction result of the load status. Moreover, the number of statistics used by the prediction calculation unit 14 in the prediction is not limited to this. For example, the prediction calculation can also be performed based on statistics with importance ranking from 1st to 10th.

[0050] Furthermore, if there is only one type of muscle to be predicted, the prediction calculation unit 14 uses the statistics and importance corresponding to that muscle to predict the load state. On the other hand, if there are multiple types of muscles to be predicted, the prediction calculation unit 14 sets statistics and importance for each muscle and predicts the load state for each muscle.

[0051] By using this structure, the action state estimation device 10 can estimate the muscle load state without performing processing that converts the measurement data into frequency components. Thus, the action state estimation device 10 can achieve the desired estimation accuracy and suppress processing load.

[0052] Furthermore, by using this structure, the action state estimation device 10 individually sets the types and importance of the statistics used for estimation for each muscle in the estimation object. As a result, the action state estimation device 10 can estimate the muscle load state with higher accuracy.

[0053] Furthermore, by using this structure, even if the position of the displacement detection sensor 101 is not the position of the muscle being predicted, the action state prediction device 10 can still predict the load state of that muscle. Thus, the action state prediction device 10 can predict the load state even for muscles not exposed on the body surface or muscles whose myoelectric potentials cannot be directly measured (for example, muscles far from the placement position of the displacement detection sensor 101). By placing the displacement detection sensor 101 on the ankle, the action state prediction device 10 can predict the load state of the quadriceps, hamstrings (biceps femoris, semimembranosus, semitendinosus, and rotatores major), tibialis anterior, gastrocnemius, soleus, gluteus maximus, etc. In addition, the action state prediction device 10 can predict the coordinated state of multiple muscles.

[0054] Furthermore, the effects of errors in the placement of the displacement detection sensor 101 can be suppressed. Therefore, the placement of the displacement detection sensor 101 becomes easier, and the operation for estimating the load condition becomes easier.

[0055] Furthermore, in this structure, the displacement measurement signal of the displacement detection sensor 101 can be used to infer the load state of multiple muscles, thus enabling the miniaturization of the motion state inference system that includes the sensor and the motion state inference device.

[0056] Furthermore, in this structure, the statistics and importance used are set individually for each muscle being predicted. Therefore, the action state prediction device 10 can use the displacement measurement signal of the displacement detection sensor 101 to predict the load state of multiple muscles with high accuracy.

[0057] Furthermore, in this structure, the load status can be inferred as a value rather than a classification. This allows the load status to be inferred from the value, enabling the action status inference device 10 to provide more accurate load status indications and management, and to provide more appropriate notifications to the subject, etc.

[0058] (Method for predicting action status)

[0059] Figure 3 This is a flowchart illustrating the main processes of the action state prediction method according to the first embodiment of the present invention. Furthermore, the specific details of each process are described in the above structural explanation, and therefore will be given a brief description below.

[0060] The motion state estimation device 10 receives a displacement measurement signal (S11). The motion state estimation device 10 samples the displacement measurement signal and generates displacement measurement data (S12). The motion state estimation device 10 calculates statistics based on the displacement measurement data (S13).

[0061] Action state prediction device 10 acquires action state model (S14). Action state prediction device 10 uses statistics as input vector and action state model to predict load state (S15).

[0062] In addition, the above-mentioned processing can also be pre-stored in a storage medium or the like and executed by a processing device such as a CPU.

[0063] (Structure and processing of the action state learning device)

[0064] The above-mentioned action state model is generated, for example, as follows. Figure 4 This is a functional block diagram of the action state learning device according to the first embodiment of the present invention.

[0065] like Figure 4As shown, the action state learning device 20 includes a sampling unit 11, a statistics calculation unit 12, an action state model storage unit 13, a sampling unit 21, a modeling baseline value calculation unit 22, and a learning calculation unit 24. The sampling unit 11, the statistics calculation unit 12, and the action state model storage unit 13 are described above and will not be further explained.

[0066] A muscle activity measurement signal (e.g., myoelectric potential signal) is input from the muscle activity detection sensor 102 to the sampling unit 21. The sampling unit 21 generates muscle activity measurement data by sampling the muscle activity measurement signal at a predetermined sampling frequency (e.g., 100 Hz). The sampling unit 21 outputs the muscle activity measurement data to the modeling reference value calculation unit 22.

[0067] Furthermore, the muscle activity detection sensor 102 is a sensor capable of measuring muscle activity, such as an electromyography (EMG) sensor. The muscle activity detection sensor 102 is positioned at the location of the muscle in which the load state is inferred. More specifically, the muscle activity detection sensor 102 is positioned at the location of the muscle that forms the basis for the muscle activity that generates the tremors measured by the muscle activity detection sensor 102. The muscle activity detection sensor 102 detects muscle activity, generates, and outputs a muscle activity measurement signal. The muscle activity detection sensor 102 can be a single sensor configured for one type of muscle, or multiple sensors configured for multiple types of muscles, one for each muscle.

[0068] The modeling baseline calculation unit 22 calculates the modeling baseline value based on muscle activity measurement data. For example, the modeling baseline calculation unit 22 calculates the absolute average value of muscle activity measurement data within a specified period as the modeling baseline value. The absolute average value refers to the average of the absolute values ​​of the measurement data.

[0069] Furthermore, the modeling baseline is not limited to the absolute mean; it can also use values ​​that can be regressed, such as the mean, maximum, minimum, median, 1% value, 5% value, 25% value, 75% value, 95% value, 99% value, variance, skewness, kurtosis, etc. Moreover, the modeling baseline can also be a value that can classify muscle activity measurement data into categories such as large, medium, and small loads.

[0070] The modeling baseline calculation unit 22 outputs the modeling baseline value to the learning calculation unit 24.

[0071] The learning computation unit 24 uses statistics and modeled baseline values ​​to learn and generate an action state model. More specifically, for example, the learning computation unit 24 uses statistics as explanatory variables and modeled baseline values ​​as target variables, and uses a gradient boosting method utilizing decision tree algorithms to learn. The learning computation unit 24 repeatedly performs this learning, and if a predetermined inference accuracy is obtained, the result is used to generate an action state model. Furthermore, the learning method is not limited to gradient boosting; boosting methods such as AdaBoost can also be used. In addition, other learning methods can include SVM, GMM, HMM, neural networks, learning Bayesian networks, etc. Moreover, multiple learners can be used in the learning computation unit 24, employing a weighted average method that uses majority voting after weighting the results of multiple learners.

[0072] By using this structure and processing, the action state learning device 20 is able to properly set the action state model.

[0073] (Method for generating action state models)

[0074] Figure 5 This is a flowchart illustrating the main processes of the action state learning method according to the first embodiment of the present invention. Furthermore, the specific details of each process are described in the above structural explanation, and therefore will be given a brief description below.

[0075] The action state learning device 20 receives a displacement measurement signal (S21). The action state learning device 20 samples the displacement measurement signal and generates displacement measurement data (S22). The action state learning device 20 calculates statistics based on the displacement measurement data (S23).

[0076] The action state learning device 20 receives a muscle activity measurement signal (S31). The action state learning device 20 samples the muscle activity measurement signal and generates muscle activity measurement data (S32). The action state learning device 20 calculates a modeled baseline value based on the muscle activity measurement data (S33).

[0077] The action state learning device 20 performs learning using statistics and modeled baseline values ​​to generate an action state model (S41).

[0078] In addition, the above-mentioned processing can also be pre-stored in a storage medium or the like and executed by a processing device such as a CPU.

[0079] Furthermore, based on the structure of the action state estimation device 10 and the action state learning device 20 of this embodiment, it can be seen that by using the structure of this embodiment, although a relatively large electromyography (EMG) measurement unit such as an EMG meter is required during learning, a high-precision action state model can be generated. In actual use (when the action state estimation device 10 is used), it is not necessary to use a relatively large EMG measurement unit such as an EMG meter. That is, in actual use, the load state can be estimated (measured) with a simple structure, reducing the burden on the subject.

[0080] (Second Implementation)

[0081] The action state prediction technique and action state model generation technique of the second embodiment of the present invention will be described with reference to the accompanying drawings. Compared with the action state prediction technique shown in the first embodiment, the action state prediction technique of the second embodiment differs in the calculation method of the statistical quantities. Figure 6 (A) represents the time variation of signal strength. Figure 6 (B) represents the signal strength distribution. Figure 6 (C) represents the intensity block data.

[0082] The statistical calculation unit 12 calculates the signal strength distribution based on displacement measurement data within a specified period. The signal strength distribution is obtained by arranging the displacement measurement data within the specified period in descending order of signal strength. For example, if the statistical calculation unit 12 calculates the signal strength distribution as follows: Figure 6 As shown in (A), the signal strength from time t1 to time t100 is obtained according to the specified sampling period (sampling frequency), then... Figure 6 As shown in (B), the signal strength is arranged from large to small, from sequential R1 to sequential R100.

[0083] The statistical calculation unit 12 generates intensity block data based on the signal strength distribution and outputs it as a statistical measure. More specifically, the statistical calculation unit 12 sets up intensity blocks (signal strength blocks) according to a predetermined number, based on the signal strength distribution in descending order of signal strength. The statistical calculation unit 12 generates intensity block data by calculating a cumulative value for each intensity block. For example, in Figure 6 In case (C), the statistical calculation unit 12 divides the measurement data into 10 blocks. As an example, the statistical calculation unit 12 sets the signal strength from sequence R1 to sequence R10 as intensity block B1, and calculates the cumulative value of the signal strength from sequence R1 to sequence R10. The statistical calculation unit 12 performs this processing from intensity block B1 to intensity block B10 and outputs the statistical result.

[0084] The prediction calculation unit 14 uses statistics obtained based on the signal strength distribution to predict the action state. At this time, the prediction calculation unit 14 uses importance to predict the action state.

[0085] Figure 7 This is a table showing examples of how important the second implementation method is. For example... Figure 7 As shown, in the second embodiment, for the intensity blocks of the signal intensity distribution, importance is set for each muscle.

[0086] For example, in Figure 7 In the case of muscle M1, the importance of the predicted intensity is set according to the order of intensity blocks B10, B9, B6, B3, and B5, in the order of 1st, 2nd, 3rd, 4th, and 5th. Similarly, for muscle M2, the importance of the predicted intensity is set according to the order of intensity blocks B10, B8, B9, B7, and B1, in the order of 1st, 2nd, 3rd, 4th, and 5th. The same applies to muscles M3 and M4. Figure 7 As shown, set the importance for the statistics (intensity blocks).

[0087] If the estimation calculation unit 14 sets the muscle to be estimated, it uses the importance set according to the muscle to estimate the muscle load state based on multiple statistics (values ​​of intensity blocks).

[0088] With this structure, similar to the action state estimation device 10 of the first embodiment, the action state estimation device of the second embodiment can estimate the muscle load state without performing processing to convert the measurement data into frequency components. Therefore, the action state estimation device of the second embodiment can achieve the necessary estimation accuracy and suppress processing load.

[0089] Furthermore, in the above explanation (refer to...) Figure 6 (A) Figure 6 (B) Figure 6 In (C), multiple measurement data are arranged in order of intensity, multiple intensity blocks are set, and a cumulative value is calculated for each intensity block, which is then used as a statistic. However, in this case, the cumulative value can also be replaced by the average value.

[0090] Furthermore, referring to... Figure 8 (A) Figure 8 (B) Figure 8 In the method described in (C), multiple blocks (time blocks) are set for multiple measurement data within a time range, and the average value and cumulative value are calculated for each time block in the multiple time blocks and used as statistics.

[0091] When using the average value, the statistics calculation unit 12 divides multiple measurement data within the time period for statistics calculation into time series blocks (time blocks), and calculates the average value for each time block. When using the cumulative value, the statistics calculation unit 12 divides multiple measurement data within the time period for statistics calculation into time series blocks (time blocks), and calculates the cumulative value for each time block. The statistics calculation unit 12 uses the average value and the cumulative value as the value (statistic) of the time block.

[0092] Figure 8 (A) is a graph representing an example of the time-varying signal strength. Figure 8 (B) is a graph representing time block data (average). Figure 8 (C) is a graph representing time block data (cumulative values).

[0093] For example, in obtaining Figure 8 When measuring the data in (A), for the signal strength from time t1 to time t100, time blocks B1t-B10t are set in the time series. Time block B1t corresponds to times t1-t10, and time block B2t corresponds to times t11-t20. Similarly, time blocks B3t-B9t are set, and time block B10t corresponds to times t91-t100.

[0094] When using average values, the statistics calculation unit 12 calculates the average signal strength from time t1 to time t10 for time block B1t, as the statistics for time block B1t. Similarly, the statistics calculation unit 12 calculates average values ​​for time blocks B2t-B9t, as their respective statistical values. Furthermore, the statistics calculation unit 12 calculates the average signal strength from time t91 to time t100 for time block B10t, as the statistics for time block B10t.

[0095] When using cumulative values, the statistics calculation unit 12 calculates the cumulative signal strength from time t1 to time t10 for time block B1t, and uses this as the statistics for time block B1t. Similarly, the statistics calculation unit 12 calculates cumulative values ​​for time blocks B2t-B9t, and uses these as their respective statistics. Furthermore, the statistics calculation unit 12 calculates the cumulative signal strength from time t91 to time t100 for time block B10t, and uses this as the statistics for time block B10t.

[0096] (Third Implementation)

[0097] The action state prediction technique and action state model generation technique of the third embodiment of the present invention will be described with reference to the accompanying drawings.

[0098] (Structure and processing of the action status prediction device)

[0099] Figure 9 This is a functional block diagram of the action state prediction device according to the third embodiment of the present invention. Figure 9 As shown, compared to the action state estimation device 10 of the first embodiment, the action state estimation device 10A differs in that it includes a sampling unit 31 and a statistical calculation unit 32, as well as in that it has an action state model storage unit 13A and an estimation calculation unit 14A. The other structures of the action state estimation device 10A are the same as those of the action state estimation device 10, and the description of the same parts is omitted.

[0100] The action state prediction device 10A includes a sampling unit 11, a statistical calculation unit 12, a prediction calculation unit 14A, an action state model storage unit 13A, a sampling unit 31, and a statistical calculation unit 32. Each functional unit constituting the action state prediction device 10A can be implemented by electronic circuits, ICs, a storage medium storing a program that executes the functions of each functional unit, and an arithmetic processing device (CPU, etc.) that executes the program.

[0101] The statistics calculation unit 12 calculates the statistics shown in the first embodiment as displacement statistics and outputs them to the prediction calculation unit 14A.

[0102] A motion measurement signal is input from the motion detection sensor 300 to the sampling unit 31. The sampling unit 31 generates motion measurement data by sampling the motion measurement signal at a predetermined sampling frequency (e.g., 100Hz). That is, the sampling unit 31 generates motion measurement data without converting the motion measurement signal into frequency components. The sampling unit 31 outputs the motion measurement data to the statistics calculation unit 32.

[0103] Furthermore, the motion detection sensor 300 is implemented using an accelerometer, angular velocity sensor, etc. The motion detection sensor 300 does not need to be positioned at the location of the muscle under the predicted load; it only needs to be positioned at a location capable of measuring the subject's movements produced by the muscles of the predicted subject. Additionally, the motion detection sensor 300 can be a single sensor positioned at one location, or multiple sensors positioned at multiple locations. The motion detection sensor 300 detects the subject's movements, generates, and outputs a measurement signal of the movement.

[0104] The statistics calculation unit 32 calculates motion statistics based on motion measurement data arranged in a time series. The statistics calculation unit 32 calculates motion statistics based on multiple motion measurement data within a specified period (e.g., 1 second).

[0105] Types of motion statistics include, for example, the average, maximum, minimum, median, 1% value, 5% value, 25% value, 75% value, 95% value, 99% value, variance, skewness, kurtosis, and cumulative value. Furthermore, the types of motion statistics are not limited to these; any type that can be obtained from time-series measurement data can be used. The statistics calculation unit 32 calculates various motion statistics from these. Additionally, the x% value refers to the value at which the maximum value among multiple motion measurement data within a period is set to 100%, equivalent to x% of its highest value.

[0106] The statistical calculation unit 32 outputs the calculated various action statistics to the prediction calculation unit 14A.

[0107] An action state model is stored in the action state model storage unit 13A. The action state model sets the relationship between various displacement statistics and various motion statistics and the muscle load state of the predicted object. The action state model is, for example, pre-generated by the action state learning device 20A described later, and stored in the action state model storage unit 13A.

[0108] The prediction calculation unit 14A uses the action state model stored in the action state model storage unit 13A, and takes displacement statistics and motion statistics as input vectors to predict the load state of the muscles of the prediction object. At this time, the prediction calculation unit 14A sets the importance of the displacement statistics and motion statistics used for prediction based on the muscles of the prediction object. These importance values ​​are, for example, set in the action state model.

[0109] Figure 10 (A) Figure 10 (B) is a table representing a setting example of the importance of the third embodiment. Figure 10 (A) indicates the importance of the displacement statistic. Figure 10 (B) indicates the importance of the action statistic. Additionally, in Figure 10 (A) Figure 10 In (B), muscle M1, muscle M2, muscle M3, and muscle M4 represent the types of muscles whose load state can be inferred from the measured displacement. For example, when the displacement detection sensor 101 is disposed at the tendon concentration area of ​​the ankle, more specifically at the front and back of the ankle, muscle M1, muscle M2, muscle M3, and muscle M4 can be set to the soleus, gastrocnemius, tibialis anterior, quadriceps, hamstrings, etc. Displacement statistics Ar1 to Ar15 are each set to any one of the above displacement statistics, and motion statistics Aa1 to Aa15 are each set to any one of the above motion statistics.

[0110] For example, in Figure 10In case (A), for muscle M1, the importance of the inference is set according to the displacement statistics Ar1, Ar2, Ar8, Ar12, and Ar7, in the order of 1st, 2nd, 3rd, 4th, and 5th position. Similarly, for muscle M2, the importance of the inference is set according to the displacement statistics Ar3, Ar5, Ar2, Ar14, and Ar15, in the order of 1st, 2nd, 3rd, 4th, and 5th position. The same applies to muscles M3 and M4. Figure 10 As shown in (A), importance is set for displacement statistics.

[0111] And, for example, in Figure 10 In case (B), for muscle M1, the importance of the inference is set according to the order of motion statistics Aa1, Aa2, Aa8, Aa12, and Aa7, in the order of 1st, 2nd, 3rd, 4th, and 5th. Similarly, for muscle M2, the importance of the inference is set according to the order of motion statistics Aa3, Aa5, Aa2, Aa14, and Aa15, in the order of 1st, 2nd, 3rd, 4th, and 5th. The same applies to muscles M3 and M4. Figure 10 As shown in (B), importance is set for motion statistics.

[0112] Furthermore, the importance of displacement statistics and motion statistics for a single muscle can be set using a shared importance system or set individually. For example, in Figure 10 (A) Figure 10 When the suffix (number) of the displacement statistic shown in (B) and the suffix (number) of the motion statistic represent the same type of statistic, Figure 10 (A) Figure 10 The importance of the displacement statistics and motion statistics shown in (B) can be set by using a shared importance setting.

[0113] On the other hand, if Figure 10 The suffix (number) of the displacement statistics shown in (A) is set according to the type of displacement statistics. Figure 10 The suffix (number) of the motion statistics shown in (B) is set according to the type of motion statistics, and the importance of displacement statistics and the importance of motion statistics are set separately.

[0114] By using a shared importance scale to set the importance of displacement and motion statistics, the setting of importance can be simplified, as can the prediction of action states. On the other hand, by setting the importance of displacement and motion statistics separately, the prediction conditions for action states can be set in a wider variety, and the prediction of action states can be made with higher accuracy.

[0115] If the estimation calculation unit 14A sets a muscle as the object of estimation, it uses an importance set according to the muscle to estimate the muscle's load state (e.g., myoelectric point value) based on multiple displacement statistics and motion statistics. Furthermore, the muscle's load state is not limited to myoelectric point value; it can be any parameter that can be expressed as a value.

[0116] More specifically, for example, the prediction calculation unit 14A calculates the predicted load state based on groups of displacement statistics and motion statistics with the same importance and the action state model. The prediction calculation unit 14A calculates the predicted load state for each group with the same importance. Furthermore, the prediction calculation unit 14A weights each predicted result according to its importance, for example, by performing an additive average, thereby calculating the final predicted load state.

[0117] Furthermore, the prediction calculation unit 14A can also independently calculate the prediction results of the load state based on displacement statistics and the motion state model, as well as the prediction results of the load state based on motion statistics and the motion state model, and calculate the final prediction result of the load state based on these prediction results. The number of posture statistics and motion statistics used by the prediction calculation unit 14 in the prediction is not limited to this, and can be appropriately set. For example, the prediction calculation can also be performed based on statistics with importance from the 1st to the 10th place.

[0118] Furthermore, if there is only one type of muscle to be predicted, the prediction calculation unit 14A uses the displacement statistics and motion statistics corresponding to that muscle, as well as their importance, to predict the load state. On the other hand, if there are multiple types of muscles to be predicted, the prediction calculation unit 14A sets the displacement statistics and motion statistics, as well as their importance, for each muscle each time, and predicts the load state for each muscle.

[0119] By using this structure, the action state estimation device 10A can infer muscle load status not only using displacement including tremors, but also using motion detection results such as acceleration and angular velocity. The sampled motion detection results such as acceleration and angular velocity are strongly influenced by the subject's movements and have a high correlation with the subject's muscle activity. Moreover, the action state estimation device 10A infers action state based on microscopic tremor measurement data and macroscopic motion measurement data, thus enabling high-precision estimation of muscle load status.

[0120] Furthermore, in this structure, displacement measurement data is also the sampled measurement data. Therefore, the action state estimation device 10A can suppress the loss of information related to tremors accompanied by frequency changes, and can estimate the muscle load state with higher accuracy.

[0121] Furthermore, by using this structure, the motion state estimation device 10A can individually set the types and importance of displacement and motion statistics used for estimation for each muscle being estimated. As a result, the motion state estimation device 10A can estimate the muscle load state with higher accuracy.

[0122] Furthermore, by using this structure, even if the positions of the displacement detection sensor 101 and the motion detection sensor 300 are not the positions of the muscles being predicted, the action state prediction device 10A can still predict the load state of those muscles. Thus, the action state prediction device 10A can predict the load state even for muscles not exposed on the body surface or muscles where myoelectric potentials cannot be directly measured (for example, muscles far from the placement positions of the displacement detection sensor 101 or the motion detection sensor 300). By placing the displacement detection sensor 101 and the motion detection sensor 300 on the ankle, the action state prediction device 10A can predict the load state of the quadriceps, hamstrings (biceps femoris, semimembranosus, semitendinosus, and rotatores major), tibialis anterior, gastrocnemius, soleus, gluteus maximus, etc. Additionally, the action state prediction device 10A can predict the coordinated state of multiple muscles.

[0123] Furthermore, the effects caused by errors in the placement of the displacement detection sensor 101 can be suppressed. Therefore, the placement of the displacement detection sensor 101 and the motion detection sensor 300 becomes easier, and the operation for estimating load conditions becomes easier.

[0124] Furthermore, in this structure, the displacement measurement signal of the displacement detection sensor 101 and the motion measurement signal of the motion detection sensor 300 can be used interchangeably to infer the load state of multiple muscles, thus enabling the miniaturization of the motion state estimation system that includes sensors and motion state estimation devices.

[0125] Furthermore, in this structure, the displacement statistics, motion statistics, and importance are individually set for each muscle being estimated. Therefore, the motion state estimation device 10A can use the measurement signals from the displacement detection sensor 101 and the motion detection sensor 300 together to estimate the load state of multiple muscles with high accuracy.

[0126] Furthermore, in this structure, the load status can be inferred as a value rather than a classification. This allows the load status to be inferred from the value, enabling the action status inference device 10A to provide more accurate load status indications and management, and to deliver more appropriate notifications to the subject.

[0127] (Method for predicting action status)

[0128] Figure 11 This is a flowchart illustrating the main processes of the action state prediction method according to the third embodiment of the present invention. Furthermore, the specific details of each process are described above in the structural explanation, and therefore will be given a brief overview below.

[0129] The motion state estimation device 10A receives a displacement measurement signal (S11A). The motion state estimation device 10A samples the displacement measurement signal and generates displacement measurement data (S12A). The motion state estimation device 10A calculates displacement statistics based on the displacement measurement data (S13A).

[0130] The motion state estimation device 10A receives a motion measurement signal (S21A). The motion state estimation device 10A samples the motion measurement signal and generates motion measurement data (S22A). The motion state estimation device 10A calculates motion statistics based on the motion measurement data (S23A).

[0131] Action state prediction device 10A acquires action state model (S14A). Action state prediction device 10A uses displacement statistics and motion statistics as input vectors and uses the action state model to predict load state (S15A).

[0132] In addition, the above-mentioned processing can also be pre-stored in a storage medium or the like and executed by a processing device such as a CPU.

[0133] (Structure and processing of the action state learning device)

[0134] The above-mentioned action state model is generated, for example, as follows. Figure 12 This is a functional block diagram of the action state learning device according to the third embodiment of the present invention.

[0135] like Figure 12 As shown, the action state learning device 20A includes a sampling unit 11, a statistics calculation unit 12, an action state model storage unit 13A, a sampling unit 21, a modeling baseline value calculation unit 22, a learning calculation unit 24A, a sampling unit 31, and a statistics calculation unit 32. The sampling unit 11, statistics calculation unit 12, action state model storage unit 13A, sampling unit 31, and statistics calculation unit 32 are described above and will not be further explained.

[0136] A muscle activity measurement signal (e.g., myoelectric potential signal) is input from the muscle activity detection sensor 102 to the sampling unit 21. The sampling unit 21 generates muscle activity measurement data by sampling the muscle activity measurement signal at a predetermined sampling frequency (e.g., 100 Hz). The sampling unit 21 outputs the muscle activity measurement data to the modeling reference value calculation unit 22.

[0137] Furthermore, the muscle activity detection sensor 102 is a sensor capable of measuring muscle activity, such as an electromyography (EMG) sensor. The muscle activity detection sensor 102 is positioned at the location of the muscle to infer its load state. More specifically, the muscle activity detection sensor 102 is positioned at the location of the muscle that forms the basis of muscle activity, producing tremors that are measured by the muscle activity detection sensor 102. The muscle activity detection sensor 102 detects muscle activity, generates, and outputs a measurement signal of the muscle activity. The muscle activity detection sensor 102 can also be a single sensor configured for one type of muscle, or multiple sensors configured for each type of muscle.

[0138] The modeling baseline calculation unit 22 calculates the modeling baseline value based on muscle activity measurement data. For example, the modeling baseline calculation unit 22 calculates the absolute average value of muscle activity measurement data within a specified period as the modeling baseline value. The absolute average value refers to the average of the absolute values ​​of the measurement data.

[0139] Furthermore, the modeling baseline is not limited to the absolute mean; it can also use values ​​that can be regressed, such as the mean, maximum, minimum, median, 1% value, 5% value, 25% value, 75% value, 95% value, 99% value, variance, skewness, kurtosis, etc. Moreover, the modeling baseline can also be a value that can classify muscle activity measurement data into categories such as large, medium, and small loads.

[0140] The modeling baseline calculation unit 22 outputs the modeling baseline value to the learning calculation unit 24A.

[0141] The learning and calculation unit 24A uses displacement statistics, motion statistics, and modeled baseline values ​​to learn and generate an action state model. More specifically, for example, the learning and calculation unit 24A uses displacement statistics and motion statistics as explanatory variables, and the modeled baseline values ​​as target variables, and uses a gradient boosting method utilizing a decision tree algorithm to learn. The learning and calculation unit 24A repeatedly performs this learning, and if a predetermined inference accuracy is obtained, the result is used to generate an action state model.

[0142] Furthermore, the learning method is not limited to gradient boosting; methods such as AdaBoost can also be used. In addition, other learning methods include SVM, GMM, HMM, neural networks, and learning-based Bayesian networks. Moreover, multiple learners can be used in the learning computation unit 24A, employing a general method that weights the results from multiple learners and then uses majority voting.

[0143] By using this structure and processing, the action state learning device 20A is able to properly set the action state model.

[0144] (Action-based learning method)

[0145] Figure 13 This is a flowchart illustrating the main processing steps of the action state learning method according to the third embodiment of the present invention.

[0146] The motion state learning device 20A receives displacement measurement signals and motion measurement signals as input (S41A). The motion state learning device 20A samples the displacement measurement signals to generate displacement measurement data and samples the motion measurement signals to generate motion measurement data (S42A). The motion state learning device 20A calculates displacement statistics based on the displacement measurement data and calculates motion statistics based on the motion measurement data (S43A).

[0147] The action state learning device 20A inputs a muscle activity measurement signal (S51A). The action state learning device 20A samples the muscle activity measurement signal and generates muscle activity measurement data (S52A). The action state learning device 20A calculates a modeled baseline value based on the muscle activity measurement data (S53A).

[0148] The action state learning device 20A performs learning using displacement statistics, motion statistics, and modeling baseline values ​​to generate an action state model (S61A).

[0149] In addition, the above-mentioned processing can also be pre-stored in a storage medium or the like and executed by a processing device such as a CPU.

[0150] (Fourth Implementation)

[0151] The action state prediction technique and action state model generation technique of the fourth embodiment of the present invention will be described. Compared with the action state prediction technique shown in the third embodiment, the action state prediction technique of the fourth embodiment differs in the use of intensity block data of displacement statistics and motion statistics. The method for generating intensity block data of motion statistics is the same as the method for generating intensity block data of displacement statistics in the third embodiment, and the description of specific examples is omitted.

[0152] The statistical calculation unit 12 generates intensity block data (displacement intensity block data) of the displacement measurement data based on the signal intensity distribution of the displacement measurement data, and outputs it as displacement statistics. More specifically, the statistical calculation unit 12 sets up intensity blocks (signal intensity blocks) according to a predetermined number of signal intensity blocks in descending order of signal intensity distribution. The statistical calculation unit 12 generates displacement intensity block data by calculating the cumulative value for each intensity block.

[0153] The statistics calculation unit 32 has the same structure as the statistics calculation unit 12, and performs the same processing on the motion measurement data as the statistics calculation unit 12. As a result, the statistics calculation unit 32 calculates and outputs motion statistics composed of intensity block data (motion intensity block data) of the motion measurement data.

[0154] The prediction calculation unit 14A uses displacement statistics and motion statistics obtained based on the signal intensity distribution to predict the action state. At this time, the prediction calculation unit 14A uses importance to predict the action state.

[0155] Figure 14 (A) Figure 14 (B) is a table representing a setting example of the importance of the fourth embodiment. Figure 14 (A) indicates the importance of the displacement statistic. Figure 14 (B) indicates the importance of the action statistic. For example... Figure 14 (A) Figure 14 As shown in (B), in the fourth embodiment, for the intensity blocks of the signal intensity distribution, importance is set for each muscle.

[0156] about Figure 14 The importance setting shown in (A), besides the fact that the displacement statistic is the value of the strength block, is related to... Figure 10 The importance setting shown in (A) is the same. Regarding Figure 14 The importance setting shown in (B), besides the fact that the motion statistics are the values ​​of the intensity blocks, is related to... Figure 10 The importance setting shown in (B) is the same. Therefore, detailed explanation is omitted.

[0157] If the estimation calculation unit 14A sets the muscle to be estimated, it uses the importance set according to the muscle to estimate the muscle load state based on multiple displacement statistics (intensity block values) and motion statistics (intensity block values).

[0158] With this structure, similar to the action state estimation device 10A of the third embodiment, the action state estimation device of the fourth embodiment can estimate the muscle load state with high accuracy.

[0159] (Fifth Implementation)

[0160] The action state estimation technique and action state model generation technique of the fifth embodiment of the present invention will be described. Compared with the action state estimation technique shown in the first embodiment, the action state estimation technique of the fifth embodiment differs in that it synchronizes the displacement measurement signal and the muscle activity measurement signal during learning. Other methods of the action state estimation technique of the fifth embodiment are the same as those of the action state estimation technique of the first embodiment, and the description of the same parts is omitted.

[0161] Figure 15 This is a flowchart illustrating the main processes of the action state learning method according to the fifth embodiment of the present invention. For example... Figure 15 As shown, the fifth embodiment of the action state learning method differs from the first embodiment in that it adds synchronization processing. Other processing in the fifth embodiment is the same as in the first embodiment, and descriptions of identical parts are omitted.

[0162] If the action state learning device 20 calculates statistics and modeled baseline values, it synchronizes the statistics with the modeled baseline values. Figure 16 This is a diagram representing the concept of synchronization. For example... Figure 16 As shown, based on the response difference between the displacement detection sensor 101 and the muscle activity detection sensor 102, a time difference Δt is generated between the reference time t0t of the displacement measurement signal and the reference time tom of the muscle activity measurement signal.

[0163] Therefore, the learning and calculation unit 24 detects the reference time tot of the displacement measurement signal and the reference time tom of the muscle activity measurement signal, and detects the time difference Δt by calculating their difference. The learning and calculation unit 24 uses the time difference Δt to synchronize the statistic with the modeled reference value.

[0164] The learning and calculation unit 24 performs learning using synchronized statistics and modeled baseline values ​​to generate an action state model (S41).

[0165] By performing such processing, the action state estimation device 10 is able to estimate the muscle load state with higher accuracy.

[0166] Furthermore, when using displacement measurement signals, motion measurement signals, and muscle activity measurement signals for learning, it is sufficient to synchronize the displacement measurement signals and motion measurement signals with the muscle activity measurement signals during learning.

[0167] (Sixth Implementation Method)

[0168] The action state estimation technique and action state model generation technique of the sixth embodiment of the present invention will be described. Compared with the action state estimation technique and action state model generation technique shown in the first embodiment, the action state estimation technique and action state model generation technique of the sixth embodiment differ in that they do not use feature quantities. Other methods of the action state estimation technique and action state model generation technique of the sixth embodiment are the same as those of the action state estimation technique and action state model generation technique of the first embodiment, and the description of the same parts is omitted.

[0169] (Structure and processing of the action status prediction device)

[0170] Figure 17 This is a functional block diagram of the action state prediction device according to the sixth embodiment of the present invention. Figure 17 As shown, the action state prediction device 10B includes a sampling unit 11, an action state model storage unit 13B, and a prediction calculation unit 14B.

[0171] The sampling unit 11 outputs the displacement measurement data, that is, the displacement measurement data expressed by the time function (displacement measurement data at multiple times), to the estimation calculation unit 14B.

[0172] An action state model is stored in the action state model storage unit 13B. The action state model sets the relationship between displacement measurement data and the load state of the muscles of the predicted object. The action state model is, for example, pre-generated by the action state learning device 20B (described later) and stored in the action state model storage unit 13B.

[0173] The prediction calculation unit 14B uses the action state model stored in the action state model storage unit 13B, and takes displacement measurement data at multiple times as input vectors to predict the load state of the muscle of the prediction object.

[0174] In this way, the action state estimation device 10B can estimate the muscle load state of the object being estimated even without calculating characteristic quantities. As a result, the action state estimation device 10B can reduce the processing load.

[0175] At this time, the estimation calculation unit 14B can set the importance of the displacement measurement data used for estimation based on the muscles of the estimation object. This importance is set, for example, in the action state model.

[0176] For example, the estimation calculation unit 14B uses the measurement start time as a reference to assign higher importance to displacement measurement data within a specified time range after a specified time has elapsed than to displacement measurement data within other time ranges. Alternatively, for example, the estimation calculation unit 14B uses the measurement start time as a reference to group the displacement measurement data according to each specified time range and sets an importance for each group.

[0177] Figure 18 (A) is a diagram illustrating a waveform used to explain the importance of a time range, and is an example of a time range setting. Figure 18 (B) is a table representing an example of a setting based on the importance of a time range. Furthermore, in Figure 18 (A) Figure 18 In (B), it indicates that the prescribed time range used in the prediction of muscle load status is divided into 4 time ranges, but the number of divisions is not limited to this.

[0178] like Figure 18 As shown in (A), the estimation calculation unit 14B divides the predetermined time range Tt used in estimating the muscle load state into multiple time ranges B1t, B2t, B3t, and B4t. The estimation calculation unit 14B assigns multiple displacement measurement data within the predetermined time range Tt to the multiple time ranges B1t, B2t, B3t, and B4t according to their respective measurement times (acquisition times).

[0179] like Figure 18 As shown in (B), the prediction calculation unit 14B stores the importance of multiple time ranges B1t, B2t, B3t, and B4t for the load state of the muscle of the prediction object. For example, in Figure 18 In case (B), if it is muscle M1, the importance of time ranges B1t and B4t is set to high, and the importance of time ranges B2t and B3t is set to low. Conversely, if it is muscle M2, the importance of time ranges B2t and B3t is set to high, and the importance of time ranges B1t and B4t is set to low.

[0180] For example, the importance can be set based on the positional relationship between the muscle of the inferred object and the displacement detection sensor 101.

[0181] As a specific example, if the distance between the muscle of the object to be predicted and the displacement detection sensor 101 is short, that is, if the displacement detection sensor 101 is located near the muscle of the object to be predicted, the displacement measurement data of the first and last time ranges (the end time range) within the specified time range of the object to be measured will have a relatively large impact on the prediction result, while the displacement measurement data of the middle time range will have a relatively small impact on the prediction result.

[0182] Therefore, in the above example, when the distance between muscle M1 and displacement sensor 101 is short, for muscle M1, the time ranges B1t and B4t corresponding to the initial and final time ranges are given higher importance, while the time ranges B2t and B3t corresponding to the intermediate time ranges are given lower importance.

[0183] On the other hand, if the distance between the muscle of the object to be predicted and the displacement detection sensor 101 is long, that is, if the displacement detection sensor 101 is located far away from the muscle of the object to be predicted, the displacement measurement data of the first and last time ranges (end time range) of the specified time range of the object to be measured has a relatively small impact on the prediction result, while the displacement measurement data of the middle time range has a relatively large impact on the prediction result.

[0184] Therefore, in the above example, when the distance between muscle M2 and displacement sensor 101 is relatively long, for muscle M2, the time ranges B1t and B4t corresponding to the initial and final time ranges are given lower importance, while the time ranges B2t and B3t corresponding to the intermediate time ranges are given higher importance.

[0185] By setting importance for time ranges in this way, the action state estimation device 10B can estimate the muscle load state of the object being estimated with higher accuracy.

[0186] Furthermore, time ranges where importance is set to low can also be excluded from prediction. Thus, the action status prediction device 10B can reduce the prediction load.

[0187] (Method for predicting action status)

[0188] Figure 19 This is a flowchart illustrating the main processing steps of the action state prediction method according to the sixth embodiment of the present invention.

[0189] The motion state estimation device 10B receives a displacement measurement signal (S11). The motion state estimation device 10B samples the displacement measurement signal and generates displacement measurement data (S12).

[0190] The action state prediction device 10B acquires an action state model using displacement measurement data (S14B). The action state prediction device 10B uses the displacement measurement data as an input vector and the action state model to predict the load state (S15B).

[0191] In addition, the above-mentioned processing can also be pre-stored in a storage medium or the like and executed by a processing device such as a CPU.

[0192] (Structure and processing of the action state learning device)

[0193] The above-mentioned action state model is generated, for example, as shown below.

[0194] Figure 20 This is a functional block diagram of the action state learning device according to the sixth embodiment of the present invention.

[0195] like Figure 20 As shown, the action state learning device 20B includes a sampling unit 11, an action state model storage unit 13B, a sampling unit 21, a modeling reference value calculation unit 22, and a learning calculation unit 24B. The sampling unit 11, the sampling unit 21, and the modeling reference value calculation unit 22 are the same as those in the action state learning device 20 of the first embodiment, and their description is omitted.

[0196] The learning and calculation unit 24B uses displacement measurement data and modeled benchmark values ​​to learn and generate an action state model. More specifically, for example, the learning and calculation unit 24B uses displacement measurement data as explanatory variables and modeled benchmark values ​​as target variables, and uses a gradient boosting method that utilizes a decision tree algorithm for learning. Furthermore, the learning method is not limited to gradient boosting; various methods described above can also be used.

[0197] The learning and calculation unit 24B repeatedly performs this learning process. If a predetermined inference accuracy is obtained, the result is used to generate an action state model. The learning and calculation unit 24B stores the generated action state model in the action state model storage unit 13B.

[0198] By using this structure and processing, the action state learning device 20B can properly set the action state model even without using feature quantities.

[0199] At this point, by performing the synchronization processing shown in the fifth embodiment, the negative impact on the generation of the action state model caused by the time difference between displacement measurement data and muscle activity measurement data can be suppressed. Therefore, the action state learning device 20B can more appropriately set the action state model.

[0200] (Method for generating action state models)

[0201] Figure 21 This is a flowchart illustrating the main processing steps of the action state learning method according to the sixth embodiment of the present invention.

[0202] The motion state learning device 20B receives a displacement measurement signal (S21). The motion state learning device 20 performs sampling on the displacement measurement signal and generates displacement measurement data (S22).

[0203] The action state learning device 20B inputs a muscle activity measurement signal (S31). The action state learning device 20B samples the muscle activity measurement signal and generates muscle activity measurement data (S32). The action state learning device 20B calculates a modeled baseline value based on the muscle activity measurement data (S33).

[0204] The action state learning device 20B performs learning using displacement measurement data and modeled reference values ​​to generate an action state model (S41B).

[0205] In addition, the above-mentioned processing can also be pre-stored in a storage medium or the like and executed by a processing device such as a CPU.

[0206] (Seventh Implementation)

[0207] The action state prediction technique and action state model generation technique of the seventh embodiment of the present invention will be described with reference to the accompanying drawings.

[0208] (Structure and processing of the action status prediction device)

[0209] Figure 22 This is a functional block diagram of the action state prediction device according to the seventh embodiment of the present invention. Figure 22 As shown, compared to the action state estimation device 10A of the third embodiment, the action state estimation device 10C differs in that it lacks a statistical calculation unit and in that it has an action state model storage unit 13C and an estimation calculation unit 14C. The other structures of the action state estimation device 10C are the same as those of the action state estimation device 10A, and the description of the same parts is omitted.

[0210] The action state prediction device 10C includes a sampling unit 11, a prediction calculation unit 14C, an action state model storage unit 13C, and a sampling unit 31.

[0211] The sampling unit 31 outputs the calculated motion measurement data to the estimation calculation unit 14C. At this time, the sampling unit 31 obtains the motion measurement data using orthogonal triaxial components or composite values ​​obtained by synthesizing these orthogonal triaxial components. By using such motion measurement data, an input vector can be set that can more accurately estimate the muscle load state of the target object. Alternatively, both orthogonal triaxial components and composite values ​​can be used.

[0212] An action state model is stored in the action state model storage unit 13C. The action state model sets the relationship between displacement measurement data, motion measurement data, and the load state of the muscles of the predicted object. The action state model is, for example, pre-generated by the action state learning device 20C described later, and stored in the action state model storage unit 13C.

[0213] The prediction calculation unit 14C uses the action state model stored in the action state model storage unit 13C, and takes displacement measurement data and motion measurement data as input vectors to predict the load state of the muscles of the prediction object. At this time, the prediction calculation unit 14C can set the importance of the displacement measurement data and the motion measurement data used for prediction. This importance is set, for example, in the action state model.

[0214] For example, the estimation calculation unit 14C uses the measurement start time as a reference to assign higher importance to displacement measurement data within a specified time range after a predetermined time interval than to displacement measurement data within other time ranges. Alternatively, for example, the estimation calculation unit 14B uses the measurement start time as a reference to group the displacement measurement data according to each specified time range and sets an importance for each group. The importance of motion measurement data can also be set using the same method as for displacement measurement data.

[0215] In this way, the motion state estimation device 10C can estimate the load state of the muscle of the object being estimated even without calculating characteristic quantities for multiple measurement data (displacement measurement data and motion measurement data).

[0216] (Method for predicting action status)

[0217] Figure 23 This is a flowchart illustrating the main processing steps of the action state prediction method according to the seventh embodiment of the present invention.

[0218] The motion state estimation device 10C receives a displacement measurement signal (S11C). The motion state estimation device 10C samples the displacement measurement signal and generates displacement measurement data (S12C).

[0219] The motion state estimation device 10C receives a motion measurement signal (S21C). The motion state estimation device 10C samples the motion measurement signal and generates motion measurement data (S22C).

[0220] The motion state prediction device 10C acquires the motion state model (S14C). The motion state prediction device 10C uses displacement measurement data and motion measurement data as input vectors and uses the motion state model to predict the load state (S15C).

[0221] In addition, the above-mentioned processing can also be pre-stored in a storage medium or the like and executed by a processing device such as a CPU.

[0222] (Structure and processing of the action state learning device)

[0223] The above-mentioned action state model is generated, for example, as follows. Figure 24 This is a functional block diagram of the action state learning device according to the seventh embodiment of the present invention.

[0224] like Figure 24 As shown, the action state learning device 20C includes a sampling unit 11, an action state model storage unit 13C, a sampling unit 21, a modeling reference value calculation unit 22, a learning calculation unit 24C, and a sampling unit 31.

[0225] The learning and calculation unit 24C uses displacement measurement data, motion measurement data, and modeled baseline values ​​to learn and generate an action state model. More specifically, for example, the learning and calculation unit 24C uses displacement measurement data and motion measurement data as explanatory variables, and the modeled baseline values ​​as target variables, and uses a gradient boosting method utilizing a decision tree algorithm to learn. The learning and calculation unit 24C repeatedly performs this learning, and if a predetermined inference accuracy is obtained, the result is used to generate an action state model. Furthermore, the learning method is not limited to gradient boosting; various methods described above can also be used.

[0226] The learning and calculation unit 24C repeatedly performs this learning process. If the prescribed inference accuracy is obtained, the result is used to generate an action state model. The learning and calculation unit 24C stores the generated action state model in the action state model storage unit 13C.

[0227] By using this structure and processing, the action state learning device 20C is able to properly set the action state model.

[0228] (Action-based learning method)

[0229] Figure 25 This is a flowchart illustrating the main processing steps of the action state learning method according to the seventh embodiment of the present invention.

[0230] The motion state learning device 20C receives displacement measurement signals and motion measurement signals (S41C). The motion state learning device 20C samples the displacement measurement signals to generate displacement measurement data, and samples the motion measurement signals to generate motion measurement data (S42C).

[0231] The action state learning device 20C inputs a muscle activity measurement signal (S51C). The action state learning device 20C samples the muscle activity measurement signal and generates muscle activity measurement data (S52C). The action state learning device 20A calculates a modeled baseline value based on the muscle activity measurement data (S53C).

[0232] The action state learning device 20C performs learning using displacement measurement data, motion measurement data, and modeled reference values ​​to generate an action state model (S61C).

[0233] In addition, the above-mentioned processing can also be pre-stored in a storage medium or the like and executed by a processing device such as a CPU.

[0234] The structures and processes described in the above embodiments can be combined. Furthermore, by using these combinations, the action state estimation device and the action state estimation method can achieve higher estimation accuracy. For example, both the statistics obtained through sampling and the statistics obtained by calculating the signal strength distribution can be used (using both statistics).

[0235] Furthermore, the above description indicates that an electromyography (EMG) sensor is used as the muscle activity detection sensor 102. However, the muscle activity detection sensor 102 can also be other sensors capable of measuring muscle activity, such as MRI.

[0236] Furthermore, the above description indicates that an inertial sensor such as an accelerometer or angular velocity sensor is used as the motion detection sensor 300. However, the motion detection sensor 300 can also be, for example, a motion sensor or an image sensor.

[0237] Furthermore, the above description indicates the method of sampling the displacement measurement signal. However, when sampling the motion measurement signal, frequency sampling of the displacement measurement signal is also possible. By sampling the displacement measurement signal as described above, the motion state estimation device can estimate the muscle load state with higher accuracy.

[0238] With this in mind, a piezoelectric sensor is preferred for displacement detection sensor 101. That is, while other sensors such as accelerometers improve accuracy by extracting frequency components, a piezoelectric sensor can make high-precision predictions even without such frequency component extraction.

[0239] Furthermore, in the method that does not use the aforementioned characteristic quantities (sixth embodiment), since the displacement detection sensor 101 is a piezoelectric sensor, it functions more effectively.

[0240] Furthermore, the above structure and processing represent a method for inferring muscle load status as a behavioral state. However, it is also possible to infer other behavioral states of the subject that are related to the muscle load status.

[0241] Furthermore, the aforementioned structure and processing describe a method of inferring the action state by directly using displacement and motion measurement data obtained through sampling, or by using displacement and motion statistics obtained through calculating signal intensity distribution. However, in the action state estimation device and method, the difference (change) and rate of change of each statistic can also be used. Specifically, in the action state estimation device and method, the difference (change) and rate of change of adjacent statistics among multiple statistics are calculated, and this calculated value is used. Thus, changes in the action state can also be inferred in the action state estimation device and method.

[0242] Furthermore, in the estimation of muscle load status using the aforementioned action state model, the following biological information-related items can be added as input vectors. For example, at least one of the following can be added: the subject's BMI, height, weight, body fat percentage, muscle mass, grip strength (left and right, first time, second time), minimum calf circumference, age group (20, 30, 40, 50, 60 years old), and gender (female, male). This allows for more accurate estimation of muscle load status.

[0243] Explanation of reference numerals in the attached figures

[0244] 10, 10A, 10B...Movement state prediction device; 11...Sampling unit; 12...Statistical calculation unit; 13, 13A, 13B...Movement state model storage unit; 14, 14A, 14B...Prediction calculation unit; 20, 20A, 20B...Movement state learning device; 21...Sampling unit; 22...Modeling baseline value calculation unit; 24, 24A, 24B...Learning calculation unit; 31...Sampling unit; 32...Statistical calculation unit; 101...Displacement detection sensor; 102...Muscle activity detection sensor; 300...Movement detection sensor.

Claims

1. An action state estimation device, comprising: The first sampling unit samples the displacement signal obtained by converting the displacement caused by the physiological tremors and deformations of the muscles on the surface of the subject into voltage within a specified time, thereby generating displacement measurement data. The action state model storage unit stores action state models derived by establishing a correlation between displacement measurement data and muscle load status; and The estimation calculation unit uses the displacement measurement data as an input vector and the action state model to estimate the load state. The action state model includes an importance level used by the prediction calculation unit. Regarding the aforementioned importance, For each muscle considered as a hypothetical target of the load state, the importance is pre-defined based on the statistics of the time block of the displacement measurement data for each hypothetical muscle or the statistics of the intensity block of the displacement measurement data for each hypothetical muscle. When calculating the final load state prediction based on multiple predictions derived from each of the aforementioned statistics, the importance is used to determine the weighting assigned to each of the multiple predictions. The prediction calculation unit uses multiple statistical measures to calculate prediction results for multiple load states, performs weighting on the prediction results for the multiple load states based on the importance, and calculates the final prediction result for the load state by performing an additive average calculation on the weighted prediction results for the multiple load states.

2. The action state prediction device according to claim 1, wherein, The action state estimation device includes a second sampling unit that samples the motion measurement signal generated and output by the motion detection sensor detecting the subject's actions within a specified time period, thereby generating motion measurement data. The action state model storage unit stores an action state model obtained by establishing a correlation between the displacement measurement data, the motion measurement data, and the muscle load state. The inference calculation unit uses the displacement measurement data and the motion measurement data as input vectors and uses the action state model to infer the load state.

3. The action state prediction device according to claim 2, wherein, In the action state model, for each predicted object of the load state, i.e., for each muscle, for each time range of the displacement measurement data and for each time range of the motion measurement data, an importance is assigned.

4. The action state prediction device according to claim 3, wherein, The importance of each time range of the displacement measurement data and the importance of each time range of the motion measurement data are set by a shared importance.

5. The action state prediction device according to claim 3, wherein, The importance of each time range of the displacement measurement data and the importance of each time range of the motion measurement data are set separately.

6. The action state prediction device according to claim 2, wherein, The motion measurement signal is a measurement signal of at least one of acceleration and angular velocity.

7. The action state prediction device according to claim 6, wherein, The motion measurement signal includes at least one of an orthogonal triaxial component and a composite component obtained by synthesizing the orthogonal triaxial component.

8. The action state prediction device according to any one of claims 1 to 7, wherein, The displacement measurement signal is a signal obtained by voltage transformation of the displacement amount, which includes the effects caused by physiological tremors.

9. The action state prediction device according to any one of claims 1 to 7, wherein, The input vector contains biological information.

10. A method for inferring action state, comprising: The first sampling process involves sampling the displacement of the subject's skin surface caused by physiological tremors and deformations of muscles within a specified time period, converting the displacement into a voltage signal (displacement measurement signal), and generating displacement measurement data; and The inference processing uses a motion state model derived by establishing a correlation between displacement measurement data and muscle load state. The displacement measurement data is used as an input vector to infer the load state. The action state model is configured with an importance level that is used by the inference calculation. Regarding the aforementioned importance, For each muscle considered as a hypothetical target of the load state, the importance is pre-defined based on the statistics of the time block of the displacement measurement data for each hypothetical muscle or the statistics of the intensity block of the displacement measurement data for each hypothetical muscle. When calculating the final load state prediction based on multiple predictions derived from each of the aforementioned statistics, the importance is used to determine the weighting assigned to each of the multiple predictions. The inference calculation process uses multiple statistical measures to calculate inference results for multiple load states, performs weighting on the inference results for the multiple load states based on the importance, and calculates the final inference result for the load state by performing an additive average calculation on the weighted inference results for the multiple load states.

11. The action state prediction method according to claim 10, wherein, The action state estimation method includes a second sampling process, in which motion measurement signals generated and output by motion detection sensors detecting the subject's actions are sampled in a time sequence to generate motion measurement data. In the inference calculation process, an action state model is used to establish a correlation between displacement measurement data, motion measurement data, and muscle load state. The displacement measurement data and the motion measurement data are used as input vectors to infer the load state.

12. An action state learning device, comprising: The first sampling unit samples the displacement signal obtained by converting the displacement caused by the physiological tremors and deformations of the muscles on the surface of the subject into voltage within a specified time, thereby generating displacement measurement data. The third sampling unit samples the muscle activity measurement signal, which is composed of muscle potentials obtained by detecting the muscle activity of the subject, within a specified time, and generates muscle activity measurement data. The modeling baseline calculation unit calculates the modeling baseline value of the action state model based on the muscle activity measurement data. as well as The learning and computing unit performs learning using the displacement measurement data and the modeled reference values ​​to generate the action state model. The action state model includes an importance level used by the learning computation unit. Regarding the aforementioned importance, For each muscle considered as a load state, the importance is pre-defined based on the statistics of the time block of displacement measurement data for each muscle of the predicted load state, or the statistics of the intensity block of displacement measurement data for each muscle of the predicted load state. When calculating the final load state prediction based on multiple predictions derived from each of the aforementioned statistics, the importance is used to determine the weighting assigned to each of the multiple predictions. The action state learning device is used to calculate the final load state prediction result by weighting the predicted results of multiple load states calculated using multiple statistical measures and by averaging the weighted predicted results of the multiple load states.

13. The action state learning device according to claim 12, wherein, The learning and computing unit synchronizes the displacement measurement data and the modeled reference value to generate an action state model.

14. The action state learning device according to claim 12 or 13, wherein, The action state learning device includes a second sampling unit that samples the action measurement signals obtained from detecting the subject's actions within a specified time period, thereby generating action measurement data. The learning and computing unit uses the displacement measurement data, the motion measurement data, and the modeling reference value to generate an action state model.

15. The action state learning device according to claim 14, wherein, The learning and computing unit synchronizes the displacement measurement data, the motion measurement data, and the modeling reference value to generate an action state model.

16. An action state learning method, comprising: The first sampling process involves sampling the displacement signal obtained by converting the displacement of the subject's skin surface caused by physiological tremors and deformations produced by muscles into voltage within a specified time, thereby generating displacement measurement data. The third sampling process involves sampling the muscle activity measurement signal, which is composed of muscle potentials obtained from detecting the muscle activity of the subject, within a specified time period, and generating muscle activity measurement data. Modeling baseline value calculation processing: Calculate the modeling baseline value of the action state model based on the muscle activity measurement data; as well as The learning computation is performed, using the displacement measurement data and the modeled benchmark values, to generate the action state model. The action state model is configured with importance values ​​that are processed by the learning computation. Regarding the aforementioned importance, For each muscle considered as a load state, the importance is pre-defined based on the statistics of the time block of displacement measurement data for each muscle of the predicted load state, or the statistics of the intensity block of displacement measurement data for each muscle of the predicted load state. When calculating the final load state prediction based on multiple predictions derived from each of the aforementioned statistics, the importance is used to determine the weighting assigned to each of the multiple predictions. The action state learning method is used when calculating the final load state prediction result by weighting the load state prediction results calculated using multiple statistical measures and then averaging the weighted load state prediction results.

17. The action state learning method according to claim 16, wherein, The action state learning method includes a second sampling process, in which the action measurement signals obtained from detecting the subject's actions are sampled in a time sequence to generate action measurement data. In the learning and processing, the displacement measurement data, the motion measurement data, and the modeling reference value are used to generate an action state model.

Images