Feature Quantity Selection Device, Feature Quantity Selection Method, and Program

Abstract

In order to select feature quantities that are highly robust against jumps and outliers, this feature quantity selection device is provided with: an acquisition unit that acquires a plurality of data sets; a construction unit that changes the distribution of the data included in the data sets and thereby constructs a plurality of re-extracted data sets; an analysis unit that analyzes the plurality of re-extracted data sets using a Lasso regression method; a statistics unit that aggregates the values of the elements included in the plurality of re-extracted data sets according to the analysis results of the plurality of re-extracted data sets, and sets logical values to the elements included in the plurality of re-extracted data sets according to the aggregation results of the values of the elements; a selection unit that, according to a preset specification rule, selects a combination of feature quantities corresponding to the logical values set to the elements; and an output unit that outputs selection information relating to the selected combination of feature quantities.

Description

Technical Field

[0001] The present disclosure relates to a feature quantity selection device and the like that selects feature quantities used for estimation.

Background Art

[0002] With the spread of IoT (Internet of Things) technology, various information can be collected from various IoT devices. For example, in fields such as medicine, healthcare, and security, attempts have been made to utilize the information collected by IoT devices. If machine learning is applied to the information collected by IoT devices, such information can be utilized for applications such as estimating physical conditions. Since IoT devices are often placed in locations where power supply is difficult, high-level power saving is required. Among the power consumption of IoT devices, the proportion of power consumption spent on communication is large. For example, if the amount of information of feature quantities used for estimating physical conditions and the like can be reduced, the amount of information transmitted from IoT devices can be reduced, and the power consumption of IoT devices can be reduced.

[0003] Patent Document 1 discloses a technique for reducing data with a weak causal relationship with abnormal prediction of a device from sensor data collected in a factory or the like. Patent Document 1 discloses a technique for reducing data with a weak causal relationship by utilizing a plurality of sparse estimation methods. Patent Document 1 mentions LASSO (Least Absolute Shrinkage and Selection Operator) regression (also called lasso regression) as an example of a sparse estimation method. In the method of Patent Document 1, a plurality of data to which a sparse estimation method can be applied are input, and for a model that applies a plurality of sparse estimation methods to perform a predetermined output, the degree of application of each of the plurality of sparse estimation methods is learned. According to the method of Patent Document 1, an appropriate sparse estimation method can be selected according to the problem, and the degree of sparsity can be adjusted for the selected sparse estimation method.

Prior Art Documents

Patent Documents

[0004] [Patent Document 1] Japanese Patent Application Laid-Open No. 2021-149590 [Summary of the Invention] [Problems to be Solved by the Invention]

[0005] If the method of Patent Document 1 is used, it is possible to reduce unimportant feature quantities among the feature quantities used for estimating the physical state and the like. For example, by using Lasso regression, unnecessary feature quantities can be reduced by selecting important feature quantities among the feature quantities used for estimation. However, when the number of subjects is small, if feature quantities are selected using Lasso regression, it is easily affected by outliers or extreme values. If outliers or extreme values are arbitrarily excluded from a small amount of data, the effectiveness of the model used for estimation decreases. Therefore, it is required to select feature quantities with high robustness against outliers or extreme values.

[0006] An object of the present disclosure is to provide a feature quantity selection device and the like that can select feature quantities with high robustness against outliers or extreme values. [Means for Solving the Problems]

[0007] A feature quantity selection device according to an aspect of the present disclosure includes an acquisition unit that acquires a plurality of data sets, a construction unit that changes the distribution of data included in the data sets to construct a plurality of re-extracted data sets, an analysis unit that analyzes the plurality of re-extracted data sets using the method of Lasso regression, a statistical unit that aggregates the values of elements included in the plurality of re-extracted data sets according to the analysis results of the plurality of re-extracted data sets, and sets logical values for the elements included in the plurality of re-extracted data sets according to the aggregation results of the values of the elements, and a selection unit that selects a combination of feature quantities according to the logical values set for the elements according to a preset specific rule, and an output unit that outputs selection information regarding the selected combination of feature quantities.

[0008] A feature quantity according to an aspect of the present disclosure SelectedIn the method, a plurality of data sets are acquired, the distribution of the data included in the data sets is changed to construct a plurality of re-extracted data sets, the plurality of re-extracted data sets are analyzed using the Lasso regression method, the values of the elements included in the plurality of re-extracted data sets are aggregated according to the analysis results of the plurality of re-extracted data sets, logical values are set for the elements included in the plurality of re-extracted data sets according to the aggregation results of the values of the elements, combination feature amounts corresponding to the values of the logical values set for the elements are selected according to a preset specific rule, and selection information regarding the combination of the selected feature amounts is output.

[0009] A program according to one aspect of the present disclosure causes a computer to execute a process of acquiring a plurality of data sets, a process of changing the distribution of the data included in the data sets to construct a plurality of re-extracted data sets, a process of analyzing the plurality of re-extracted data sets using the Lasso regression method, a process of aggregating the values of the elements included in the plurality of re-extracted data sets according to the analysis results of the plurality of re-extracted data sets, a process of setting logical values for the elements included in the plurality of re-extracted data sets according to the aggregation results of the values of the elements, a process of selecting combination feature amounts corresponding to the values of the logical values set for the elements according to a preset specific rule, and a process of outputting selection information regarding the combination of the selected feature amounts.

Advantages of the Invention

[0010] According to the present disclosure, it becomes possible to provide a feature amount selection device or the like that can select a feature amount with high robustness against outliers and extreme values.

Brief Description of the Drawings

[0011]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

Figure 16

Figure 17

Figure 18

Figure 19

Figure 20

Figure 21

Figure 22

Figure 23

Figure 24

Figure 25

Figure 26

Figure 27

Figure 28

Figure 29

Figure 30

Figure 31

Figure 32

Figure 33

Figure 34

Figure 35

Figure 36

Mode for Carrying Out the Invention

[0012] Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings. However, although the embodiments described below have technically preferable limitations for carrying out the present invention, the scope of the invention is not limited thereto. In all the drawings used in the following description of the embodiments, the same reference numerals are given to the same parts unless otherwise specified. Also, in the following embodiments, repeated explanations regarding the same configuration and operation may be omitted.

[0013] (First Embodiment) First, a feature quantity selection device according to the first embodiment will be described with reference to the drawings. The feature quantity selection device of this embodiment selects feature quantities used for estimating physical states and the like using the method of LASSO (Least Absolute Shrinkage and Selection Operator) regression (hereinafter referred to as Lasso regression). Lasso regression is also called L1 regularization.

[0014] In the following, the selection of feature quantities used for estimating physical states will be described. For example, the feature quantities used for estimating physical states are extracted based on sensor data related to the movement of the feet according to the user's walking. For example, the sensor data related to the movement of the feet is measured by a measuring device installed in the footwear. For example, the measuring device includes an acceleration sensor and an angular velocity sensor. Note that the sensor data is not limited to the sensor data related to the movement of the feet, and may include features related to the gait. For example, the sensor data may be sensor data including features related to the gait measured using motion capture, smart apparel, or the like. Further, the following method can be applied not only to the selection of feature quantities related to the gait but also to the selection of feature quantities from arbitrary sensor data.

[0015] (Configuration) FIG. 1 is a block diagram showing an example of the configuration of a feature quantity selection device 10 according to this embodiment. The feature quantity selection device 10 includes an acquisition unit 11, a construction unit 12, an analysis unit 13, a statistics unit 15, a selection unit 17, and an output unit 19.

[0016] The acquisition unit 11 acquires a data set measured for a plurality of subjects and used for estimating physical states. The data set is data that combines an explanatory variable and an objective variable corresponding to the explanatory variable. For example, the data set is data that associates a measured value or feature quantity related to a subject with the physical state of the subject. For example, the explanatory variables used for estimating physical states are feature quantities extracted from sensor data related to the movement of the feet or the gait.

[0017] The construction unit 12 constructs a new dataset (also called a re-extracted dataset) by changing the distribution of the dataset regarding a plurality of subjects. For example, the construction unit 12 constructs a re-extracted dataset using the Leave-One-Subject-Out (also called LOSO) method. When using the LOSO method, one is removed from a plurality of datasets, and a reconstructed dataset is constructed using the remaining datasets. When using the LOSO method, the reconstructed datasets are generated by the number of subjects. For example, when there are 50 subjects, 50 re-extracted datasets can be constructed by using the LOSO method.

[0018] The construction unit 12 may construct a re-extracted dataset using the bootstrap method. In the bootstrap method, the properties of the population are estimated based on the values randomly extracted from the sample population by the sampling with replacement method. In the bootstrap method, the generation of new datasets using the values randomly extracted from the sample population is repeated, and statistical values are calculated. For example, if the generation of new datasets is repeated 1000 times, a re-extracted dataset of 1000 data can be constructed.

[0019] The analysis unit 13 performs lasso regression on the re-extracted dataset constructed by the construction unit 12. For example, the analysis unit 13 uses the loss function represented by the following formula 1.

Equation

[0020] The first term on the right side of Equation 1 is the term related to the sum of squared errors. The second term on the right side of Equation 1 is the regularization term. The regularization term is a function defined to return a larger value as the model parameter β j becomes larger. The regularization term corresponds to a penalty on the magnitude of the model parameter β j .

[0021] The regularization parameter λ is a meta-parameter set during the learning of the model. The regularization parameter λ adjusts the strength of regularization (penalty). When the value of the regularization parameter λ is large, the penalty of the regularization term becomes stronger, and overfitting is more strongly suppressed. If the value of the regularization parameter λ is too large, keeping the model parameters small is prioritized and the expressiveness of the model decreases. As a result, if the value of the regularization parameter λ is too large, a large bias remains.

[0022] Depending on the setting of the regularization parameter λ, in Lasso regression, the problem of the following Equation 2 is solved.

Equation

[0023] The following Equation 3 is a constraint condition for each element of the coefficient vector.

Equation

[0024] For the regularization parameter λ, there is a corresponding coefficient vector β. As the regularization parameter λ increases, the non-zero elements of the coefficient vector β decrease. That is, as the regularization parameter λ increases, the zero elements of the coefficient vector β increase, and the unnecessary feature quantities increase. On the contrary, as the regularization parameter λ decreases, the non-zero elements of the coefficient vector β increase, and the necessary feature quantities increase. If an appropriate regularization parameter λ is set, unnecessary zero elements can be reduced while leaving the non-zero elements necessary for estimation.

[0025] The analysis unit 13 executes Lasso regression on the re-extracted dataset constructed by the construction unit 12. The analysis unit 13 executes Lasso regression by changing the regularization parameter λ for each subject's re-extracted dataset. The analysis unit 13 generates a matrix (also called the first matrix) composed of the number of columns of the regularization parameter λ and the number of rows of the feature quantity. For example, when there are P regularization parameters λ, each regularization parameter λ is assigned a number from 1 to P (referred to as the λ number) (P is a natural number). The first matrix has the number of rows of the feature quantities used for estimation such as the physical state. When the number of feature quantities is p, each feature quantity is assigned a number from 1 to p (referred to as the feature quantity number) (p is a natural number).

[0026] Figure 2 is a conceptual diagram showing an example of the first matrix. In the first matrix of Figure 2, the hatched cells indicate non-zero elements. In the first matrix of Figure 2, the blank cells without hatching indicate zero elements.

[0027] Figure 3 is an example of the first matrix generated for 50 subjects using the reconstructed dataset constructed by LOSO. When the reconstructed dataset is constructed by LOSO, the first matrix for each subject is generated. In the case of the example in Figure 3, 50 patterns of the first matrix are generated according to the number of subjects (50 people). In the first matrix of Figure 3, the hatched cells indicate non-zero elements. In the first matrix of Figure 3, the blank cells without hatching indicate zero elements.

[0028] The statistics unit 15 assigns logical values (0, 1) to each cell of the first matrix of the generated multiple patterns. The process of assigning logical values (0, 1) to each cell of the first matrix of the multiple patterns generated by the Lasso regression is also called the first statistical process. In the first statistical process, the statistics unit 15 sets non-zero elements to TRUE (1) and zero elements to FALSE (0) for the multiple first matrices. The statistics unit 15 aggregates the logical values (0, 1) for each cell for all the first matrices. The statistics unit 15 aggregates the logical values (0, 1) for each cell by adding the logical values (1) of the non-zero elements for each cell for all the first matrices.

[0029] FIG. 4 is an example in which the aggregated values of the logical values for the first matrix of 50 subjects are entered in each cell of the matrix (also called the second matrix) corresponding to all the first matrices. In each cell of the second matrix, the number of non-zero elements (number of TRUEs) of the first matrix is entered for the multiple patterns.

[0030] The statistics unit 15 assigns logical values (0 or 1) to each cell of the second matrix according to the aggregated value of each cell included in all the first matrices. When the aggregated value for each cell of the second matrix is equal to or greater than a predetermined threshold, the statistics unit 15 sets that cell to TRUE (1). On the other hand, when the aggregated value is less than the predetermined threshold, the statistics unit 15 sets that cell to FALSE (0). The process of aggregating the logical values (0, 1) for each cell for all the first matrices and assigning logical values (0 or 1) corresponding to the aggregated values to each cell of the second matrix is also called the second statistical process.

[0031] FIG. 5 is a conceptual diagram showing an example in which logical values (0 or 1) are assigned to the aggregated values in FIG. 4. In the example of FIG. 5, cells with an aggregated value of 49 or more in FIG. 4 are set to TRUE (1). In the example of FIG. 5, cells set to TRUE (1) by the second statistical process are hatched. In the example of FIG. 5, cells set to FALSE (0) by the second statistical process are blank. For example, the result of the second statistical process such as in FIG. 5 may be displayed on a screen that can be confirmed by the user. In that case, the user can select a desired combination of feature amounts by selecting a λ number according to the result of the second statistical process displayed on the screen.

[0032] The statistics unit 15 may assign a logical value according to the average value of the aggregated values to each cell of the second matrix. For example, for each cell of the second matrix, the statistics unit 15 sets a cell with an average value of the aggregated value equal to or greater than a predetermined threshold to TRUE (1). On the other hand, the statistics unit 15 sets a cell with an average value of the aggregated value less than the predetermined threshold to FALSE (0). Such processing is also included in the second statistical process.

[0033] The selection unit 17 selects a λ number according to a preset specific rule. The specific rule is a rule for determining the λ number to be selected. For example, the specific rule is a rule of selecting a λ number corresponding to a preset reference value for the number of cells set to TRUE (1). The reference value may be set according to constraints on the calculation amount and communication amount. For example, the reference value is set to a value that does not exceed the load assignable to the calculation amount and communication amount. For example, the reference value is set to a value that does not exceed a ratio (for example, 50 to 80 percent, etc.) with respect to the load assignable to the calculation amount and communication amount. For example, when there are a plurality of λ numbers corresponding to the reference value for the number of cells set to TRUE (1), at least any one of the λ numbers may be selected. The selection unit 17 selects a combination of feature amounts in which the cells of the selected λ number are set to TRUE (1) based on the specific rule. For example, the selection unit 17 may select a combination of feature amounts according to a reference value set by the user.

[0034] The output unit 19 outputs information regarding the feature amount selected by the selection unit 17 (also referred to as selection information). The selection information is information regarding the combination of feature amounts used for estimating the physical state and the like. For example, in the case of sensor data regarding the movement of the foot, the selection information includes information indicating from which walking phase to extract the feature amount in the time-series data of acceleration and angular velocity for one walking cycle. The walking phase indicates the walking cycle (percentage) when one walking cycle is normalized to 0 to 100 percent. Feature amounts may be extracted over a plurality of consecutive walking phases. A block of a plurality of consecutive walking phases from which feature amounts are extracted is also referred to as a walking phase cluster.

[0035] The selection information output from the output unit 19 is used as a condition for extracting feature amounts from sensor data measured by a measuring device or the like. For example, the selection unit 17 may store the selection information in a storage unit (not shown). For example, the feature amounts extracted according to the selection information are used for learning an estimation model for estimating the physical state and the like. For example, the feature amounts to be extracted are extracted from sensor data measured by a measuring device or the like worn by a user who is the estimation target of the physical state.

[0036] FIGS. 6 to 7 are conceptual diagrams for explaining the difference in estimated values by an estimation model generated by learning using the feature amounts selected using the method of general lasso regression (comparative example) and the feature amounts selected using the method of the present embodiment. FIGS. 6 to 7 are examples in which the score of the TUG (Time Up and Go) test is estimated as the movement ability of the subject. The score of the TUG test is the time (also referred to as the TUG required time) from standing up from the chair, walking to a target 3 meters ahead, changing direction, and sitting back on the chair again.

[0037] FIG. 6 shows an estimation example using an estimation model generated using nine feature quantities selected by a general Lasso regression (comparative example) method. Regarding the true value (measured value) and the estimated value of the TUG required time, the intraclass correlation coefficient ICC (Intraclass Correlation Coefficients) was 0.602. Also, regarding the true value (measured value) and the estimated value of the TUG required time, the mean absolute error MAE (Mean Absolute Error) was 0.71.

[0038] FIG. 7 shows an estimation example using an estimation model generated using nine feature quantities selected by the method of this embodiment. In the example of FIG. 7, a re-extracted dataset constructed by LOSO was used. Regarding the true value (measured value) and the estimated value of the TUG required time, the intraclass correlation coefficient ICC was 0.682. Also, regarding the true value (measured value) and the estimated value of the TUG required time, the mean absolute error MAE was 0.63. As described above, both the ICC and the MAE were larger when using the method of this embodiment. That is, by using the method of this embodiment, the robustness against outliers and extreme values was improved.

[0039] FIG. 8 is a graph for explaining the influence of outliers and extreme values that may be included in sensor data measured for a plurality of subjects. The data within the range surrounded by the dashed circle corresponds to outliers and extreme values. L1 is the regression line when performing linear regression on a plurality of sensor data including outliers and extreme values. L2 is the regression line when arbitrarily excluding outliers and extreme values and performing linear regression on a plurality of sensor data. The regression line L1 is affected by outliers and extreme values and does not fit most of the sensor data. On the other hand, the regression line L2 is not affected by outliers and extreme values and fits most of the sensor data. Comparing the regression line L1 and the regression line L2, it seems that a more accurate estimation model can be constructed by using the regression line L2 obtained by arbitrarily excluding outliers and extreme values. However, arbitrarily excluding outliers and extreme values reduces the effectiveness of the estimation model. Therefore, it is required to select feature quantities with high robustness against outliers and extreme values without reducing the effectiveness of the estimation model.

[0040] In the method of this embodiment, after changing the distribution of the dataset using methods such as LOSO or the bootstrap method, lasso regression is performed. Further, in the method of this embodiment, not only are methods such as LOSO or the bootstrap method simply combined with lasso regression, but the above-described first statistical process and second statistical process are also executed. As a result, according to the method of this embodiment, an average solution with reduced influence of outliers and extreme values can be obtained. Further, according to the method of this embodiment, it is possible to select a feature amount with high robustness against outliers and extreme values without degrading the effectiveness of the estimation model.

[0041] (Operation) Next, the operation of the feature amount selection device 10 of this embodiment will be described with reference to the drawings. FIGS. 9 to 10 are flowcharts for explaining an example of the operation of the feature amount selection device 10. In the description using the flowcharts of FIGS. 9 to 10, the feature amount selection device 10 will be described as the main operating entity.

[0042] In FIG. 9, first, the feature amount selection device 10 acquires N datasets (step S111). The dataset number corresponds to the number of explanatory variables (feature amounts) included in the dataset (feature amount number).

[0043] Next, the feature amount selection device 10 sets the feature amount number n to 1 (step S112). n is the number of the dataset (feature amount).

[0044] Next, the feature amount selection device 10 excludes the data of the n-th subject (step S113).

[0045] Next, the feature amount selection device 10 performs lasso regression on the N - 1 datasets from which the data of the n-th subject has been excluded (step S114).

[0046] Next, the feature quantity selection device 10 executes first statistical processing (step S115). As the first statistical processing, the feature quantity selection device 10 assigns logical values to each cell of the first matrix (matrix B n ) generated by Lasso regression. For example, the feature quantity selection device 10 sets the non-zero elements of matrix B n to TRUE (1), and sets the zero elements of matrix B n to FALSE (0). For example, the feature quantity selection device 10 may set the cells whose element values in matrix B n are greater than or equal to the threshold value T0 to TRUE (1), and set the cells whose element values in matrix B n are less than the threshold value T0 to FALSE (0).

[0047] Next, the feature quantity selection device 10 increments the feature quantity number n by 1 (+1) (step S116).

[0048] Here, when the feature quantity number n is smaller than the number N of data sets (Yes in step S117), the process returns to step S113. On the other hand, when the feature quantity number n is greater than or equal to the number N of data sets (No in step S117), the process proceeds to step S121 in FIG. 10.

[0049] When No in step S117 of FIG. 9, the feature quantity selection device 10 executes second statistical processing (step S121). As the second statistical processing, for all the first matrices, the feature quantity selection device 10 aggregates the logical values (0 or 1) for each cell. The feature quantity selection device 10 sets logical values (0 or 1) for each cell of the aggregated second matrix according to the relationship between the aggregated value of the logical values for each cell and a predetermined threshold value. For example, the feature quantity selection device 10 sets the cells whose aggregated value is greater than or equal to the predetermined threshold value to TRUE (1). On the other hand, the feature quantity selection device 10 sets the cells whose aggregated value is less than the predetermined threshold value to FALSE (0).

[0050] Next, the feature quantity selection device 10 selects a λ number based on a specific rule according to the result of the second statistical processing (step S122).

[0051] Next, the feature quantity selection device 10 selects a combination of feature quantities corresponding to the selected λ number (step S123).

[0052] Next, the feature quantity selection device 10 outputs information (selection information) regarding the selected feature quantity (step S124). The selection information output from the feature quantity selection device 10 is used as a condition for extracting feature quantities from sensor data measured by a measuring device or the like.

[0053] As described above, the feature quantity selection device of the present embodiment includes an acquisition unit, a construction unit, an analysis unit, a statistics unit, a selection unit, and an output unit. The acquisition unit acquires a plurality of data sets. The construction unit constructs a plurality of re-extracted data sets by changing the distribution of the data included in the data set. The analysis unit analyzes the plurality of re-extracted data sets using the method of lasso regression. The statistics unit totals the values of the elements included in the plurality of re-extracted data sets according to the analysis results of the plurality of re-extracted data sets. The statistics unit sets a logical value for the elements included in the plurality of re-extracted data sets according to the total result of the values of the elements. The selection unit selects a combination of feature quantities according to the value of the logical value set for the element according to a preset specific rule. The output unit outputs selection information regarding the selected combination of feature quantities.

[0054] In this embodiment, the analysis unit executes Lasso regression for each of a plurality of preset regularization parameters with respect to a plurality of re-extracted data sets. The analysis unit generates a first matrix of a plurality of patterns composed of a column corresponding to the regularization parameter used in the Lasso regression and a row corresponding to the feature amount. The statistics unit executes a first statistical process of setting the first logical value of a cell with a non-zero element to 1 and setting the first logical value of a cell with a zero element to 0 with respect to the first matrix of the plurality of patterns. The statistics unit aggregates the first logical values for each cell constituting the first matrix of the plurality of patterns. The statistics unit executes a second statistical process of generating a second matrix in which 1 is set as the second logical value for a cell whose aggregated value of the first logical values satisfies a predetermined condition and 0 is set as the second logical value for a cell whose aggregated value of the first logical values does not satisfy the predetermined condition. The selection unit selects columns of the second matrix according to a preset specific rule and selects a combination of feature amounts corresponding to the selected columns.

[0055] According to this embodiment, by changing the data distribution and performing Lasso regression, it is possible to obtain feature amounts of average values closer to the true value as compared with the case of being directly affected by outliers or extreme values. Therefore, according to this embodiment, it is possible to select feature amounts with high robustness against outliers or extreme values.

[0056] In one aspect of this embodiment, the construction unit constructs a plurality of re-extracted data sets using the Leave-One-Subject-Out method. According to this aspect, by pseudo-changing the data distribution using the Leave-One-Subject-Out method, it is possible to approximate the data distribution to the distribution of the population that should originally exist.

[0057] In one aspect of this embodiment, the construction unit constructs a plurality of re-extracted data sets using the bootstrap method. According to this aspect, by pseudo-changing the data distribution using the bootstrap method, it is possible to approximate the data distribution to the distribution of the population estimated from the sample population.

[0058] In one aspect of the present embodiment, in the second statistical process, the statistics unit calculates the total value of the first logical values for each cell constituting the first matrix of a plurality of patterns. The statistics unit generates a second matrix in which the second logical value of a cell whose total value of the first logical values is equal to or greater than a predetermined threshold is set to 1, and the second logical value of a cell whose total value of the first logical values is less than the predetermined threshold is set to 0. According to this aspect, a combination of feature amounts can be selected based on the logical values of the second matrix set according to the total value of the first logical values.

[0059] In one aspect of the present embodiment, in the second statistical process, the statistics unit calculates the average value of the first logical values for each cell constituting the first matrix of a plurality of patterns. The statistics unit generates a second matrix in which the second logical value of a cell whose average value of the first logical values is equal to or greater than a predetermined threshold is set to 1, and the second logical value of a cell whose average value of the first logical values is less than the predetermined threshold is set to 0. According to this aspect, a combination of feature amounts can be selected based on the logical values of the second matrix set according to the average value of the first logical values.

[0060] (Second Embodiment) Next, a feature amount selection device according to the second embodiment will be described with reference to the drawings. The feature amount selection device of the present embodiment constructs an estimation model using the feature amounts selected by the method of the first embodiment. The feature amount selection device of the present embodiment selects a feature amount according to the estimation result of the constructed estimation model.

[0061] (Configuration) FIG. 11 is a block diagram showing an example of the configuration of a feature amount selection device 20 according to the present embodiment. The feature amount selection device 20 includes an acquisition unit 21, a construction unit 22, an analysis unit 23, a statistics unit 25, an estimation model construction unit 26, a selection unit 27, and an output unit 29.

[0062] The acquisition unit 21 has the same configuration as the acquisition unit 11 of the first embodiment. The acquisition unit 21 acquires a data set measured for a plurality of subjects and used for estimating a physical state.

[0063] The construction unit 22 has the same configuration as the construction unit 12 in the first embodiment. The construction unit 22 constructs a new dataset (also called a re-extracted dataset) by changing the distribution of the dataset regarding a plurality of subjects. For example, the construction unit 22 constructs a re-extracted dataset using the Leave-One-Subject-Out (also called LOSO) method. For example, the construction unit 12 may construct a re-extracted dataset using the bootstrap method.

[0064] The analysis unit 23 has the same configuration as the analysis unit 13 in the first embodiment. The analysis unit 23 performs lasso regression on the re-extracted dataset constructed by the construction unit 22. The analysis unit 23 generates a matrix (also called the first matrix) composed of a column of the number of regularization parameters λ and a row of the number of feature quantities. As a result, a first matrix having a column of the number of changed regularization parameters λ is generated.

[0065] The statistics unit 25 has the same configuration as the statistics unit 15 in the first embodiment. The statistics unit 25 executes a first statistical process of assigning logical values (0, 1) to each cell of a plurality of patterns of the first matrix generated for each subject. In the first statistical process, the statistics unit 25 sets non-zero elements to TRUE (1) and zero elements to FALSE (0) for a plurality of first matrices.

[0066] The statistics unit 25 executes a second statistical process of aggregating the logical values (0, 1) for each cell for all the first matrices and adding the logical values (1) of the non-zero elements for each cell. The statistics unit 25 assigns a logical value (0 or 1) to each cell of the second matrix according to the aggregated value of each cell of the first matrix. For each cell of the second matrix, when the aggregated value is equal to or greater than a predetermined threshold, the statistics unit 25 sets that cell to TRUE (1). On the other hand, when the aggregated value is less than the predetermined threshold, the statistics unit 25 sets that cell to FALSE (0).

[0067] The estimation model construction unit 26 constructs an estimation model by learning using the feature quantities selected by the selection unit 27. The estimation model construction unit 26 evaluates the constructed estimation model. For example, the estimation model construction unit 26 calculates evaluation indexes such as mean squared error, mean absolute error, mean relative error, coefficient of determination, and correlation coefficient. The estimation model construction unit 26 outputs the calculated evaluation indexes to the selection unit 27. For example, the evaluation result of the estimation model may be displayed on a screen that can be confirmed by the user. In that case, the user can select the most likely combination of feature quantities according to the evaluation result displayed on the screen.

[0068] The selection unit 27 selects the combination of feature quantities with the highest evaluation index calculated by the estimation model construction unit 26. For example, the selection unit 27 may select the most likely feature quantities according to an instruction input by the user.

[0069] The output unit 29 has the same configuration as the output unit 19 in the first embodiment. The output unit 29 outputs information (also referred to as selection information) regarding the feature quantities selected by the selection unit 27. The selection information output from the output unit 29 is used as a condition for extracting feature quantities from sensor data measured by a measuring device or the like. For example, the selection information may be stored in a storage unit (not shown) by the selection unit 27. For example, the feature quantities extracted according to the selection information are used for learning an estimation model for estimating a physical state or the like. For example, the feature quantities to be extracted are extracted from sensor data measured by a measuring device or the like worn by a user whose physical state is to be estimated.

[0070] (Operation) Next, the operation of the feature quantity selection device 20 of the present embodiment will be described with reference to the drawings. FIGS. 12 to 13 are flowcharts for explaining an example of the operation of the feature quantity selection device 20. In the description using the flowcharts of FIGS. 12 to 13, the feature quantity selection device 20 is described as the operating entity.

[0071] In FIG. 12, first, the feature quantity selection device 20 acquires N data sets (step S211). The number of the data set corresponds to the number of explanatory variables (feature quantities) included in the data set (feature quantity number).

[0072] Next, the feature quantity selection device 20 sets the feature quantity number n to 1 (step S212). n is the number of the data set (feature quantity).

[0073] Next, the feature quantity selection device 20 excludes the data of the n-th subject (step S213).

[0074] Next, the feature quantity selection device 20 performs Lasso regression on the N - 1 data sets from which the data of the n-th subject has been excluded (step S214).

[0075] Next, the feature quantity selection device 20 performs first statistical processing (step S215). As the first statistical processing, the feature quantity selection device 20 assigns logical values to each cell of the first matrix (matrix B n ) generated by Lasso regression. For example, the feature quantity selection device 20 sets the non-zero elements of matrix B n to TRUE (1), and sets the zero elements of matrix B n to FALSE (0). For example, the feature quantity selection device 20 may set the cells whose element values in matrix B n are equal to or greater than the threshold value T0 to TRUE (1), and set the cells whose element values in matrix B n are less than the threshold value T0 to FALSE (0).

[0076] Next, the feature quantity selection device 20 increments the feature quantity number n by 1 (+1) (step S216).

[0077] Here, if the feature quantity number n is smaller than the number N of data sets (Yes in step S217), the process returns to step S213. On the other hand, if the feature quantity number n is equal to or greater than the number N of data sets (No in step S217), the process proceeds to step S221 in FIG. 13.

[0078] If the answer is "No" in step S217 of FIG. 12, the feature quantity selection device 20 executes a second statistical process (step S221). As the second statistical process, the feature quantity selection device 20 totals the logical values (0, 1) for each cell for all the first matrices. The feature quantity selection device 20 assigns the sum of the logical values for each cell of the totaled first matrix to each cell of the second matrix. The feature quantity selection device 20 sets a logical value (0 or 1) for each cell of the totaled second matrix according to the value of each cell of the second matrix. For example, the feature quantity selection device 20 sets cells with a total value equal to or greater than a predetermined threshold to TRUE (1). On the other hand, the feature quantity selection device 20 sets cells with a total value less than the predetermined threshold to FALSE (0).

[0079] Next, the feature quantity selection device 20 executes a model evaluation process (step S222). The model evaluation process in step S222 will be described later (FIG. 14).

[0080] Next, the feature quantity selection device 20 searches for a λ number according to the evaluation index obtained by the model evaluation process (step S223).

[0081] Next, the feature quantity selection device 20 selects a combination of feature quantities corresponding to the searched λ number (step S224).

[0082] Next, the feature quantity selection device 20 outputs information (selection information) regarding the selected feature quantities (step S225). The selection information output from the feature quantity selection device 20 is used as a condition for extracting feature quantities from sensor data measured by a measuring device or the like.

[0083] 〔Model Evaluation Process〕 Next, the model evaluation process in step S222 of FIG. 13 will be described with reference to the drawings. FIG. 14 is a flowchart for explaining the model evaluation process. In the explanation using the flowchart of FIG. 14, the feature quantity selection device 20 will be described as the operating entity.

[0084] In FIG. 14, first, the feature quantity selection device 20 sets the λ number m to 1 (step S231). Here, m is the number of the regularization parameter λ.

[0085] Next, the feature quantity selection device 20 selects a combination of feature quantities corresponding to the λ number m (step S232).

[0086] Next, the feature quantity selection device 20 constructs an estimation model using the selected feature quantities (step S233).

[0087] Next, the feature quantity selection device 20 evaluates the constructed estimation model (step S234).

[0088] Next, the feature quantity selection device 20 outputs an evaluation index of the estimation model (step S235).

[0089] Next, the feature quantity selection device 20 increments the λ number m by 1 (+1) (step S236).

[0090] Here, if the λ number m is smaller than the number P of the regularization parameters λ (Yes in step S237), the process returns to step S232. On the other hand, if the λ number m is greater than or equal to the number P of the regularization parameters λ (Yes in step S237), the process proceeds to step S223 in FIG. 13.

[0091] As described above, the feature quantity selection device of the present embodiment includes an acquisition unit, a construction unit, an analysis unit, a statistics unit, a selection unit, an estimation model construction unit, and an output unit. The acquisition unit acquires a plurality of data sets. The construction unit constructs a plurality of re-extracted data sets by changing the distribution of the data included in the data set. The analysis unit analyzes the plurality of re-extracted data sets using the method of Lasso regression. The statistics unit aggregates the values of the elements included in the plurality of re-extracted data sets according to the analysis results of the plurality of re-extracted data sets. The statistics unit sets a logical value for the elements included in the plurality of re-extracted data sets according to the aggregation result of the values of the elements. The selection unit selects a combination of feature quantities according to the value of the logical value set for the elements according to a preset specific rule. The estimation model construction unit constructs an estimation model by learning using the selected feature quantities and evaluates the constructed estimation model. The selection unit selects a combination of feature quantities according to the evaluation result of the estimation model. The output unit outputs selection information regarding the selected combination of feature quantities.

[0092] In the present embodiment, the feature quantities are selected according to the evaluation result of the estimation model constructed using the feature quantities selected by the selection unit. Therefore, according to the present embodiment, by using the evaluation result of the estimation model, highly reliable feature quantities can be selected.

[0093] (Third Embodiment) Next, the feature quantity selection device according to the third embodiment will be described with reference to the drawings. The feature quantity selection device of the present embodiment is different from the first embodiment in that the first statistical process is omitted and the second statistical process is executed for the average value of each cell regarding a plurality of first matrices.

[0094] (Configuration) FIG. 15 is a block diagram showing an example of the configuration of the feature quantity selection device 30 according to the present embodiment. The feature quantity selection device 30 includes an acquisition unit 31, a construction unit 32, an analysis unit 33, a statistics unit 35 , selected a determination unit 37, and an output unit 39.

[0095] The acquisition unit 31 has the same configuration as the acquisition unit 11 in the first embodiment. The acquisition unit 31 acquires a dataset used for estimating the physical state, which has been measured for a plurality of subjects.

[0096] The construction unit 32 has the same configuration as the construction unit 12 in the first embodiment. The construction unit 32 constructs a new dataset (also called a re-extracted dataset) by changing the distribution of the datasets for a plurality of subjects. For example, the construction unit 32 constructs a re-extracted dataset using the Leave-One-Subject-Out (LOSO) method. For example, the construction unit 32 may construct a re-extracted dataset using the bootstrap method.

[0097] The analysis unit 33 has the same configuration as the analysis unit 13 in the first embodiment. The analysis unit 33 performs lasso regression on the re-extracted dataset constructed by the construction unit 32. The analysis unit 33 generates a matrix (also called a first matrix) composed of the number of columns of the regularization parameter λ and the number of rows of the feature quantity. As a result, a first matrix having the number of columns of the changed regularization parameter λ is generated.

[0098] The statistics unit 35 calculates the average value of each cell for a plurality of patterns of the first matrix generated for each subject. The statistics unit 35 generates a second matrix in which the average value of each cell is assigned to each cell of the first matrix of a plurality of patterns. For example, the statistics unit 35 sets a cell whose average value of the aggregated value is equal to or greater than a predetermined threshold to TRUE (1) for each cell of the second matrix. On the other hand, the statistics unit 35 sets a cell whose average value of the aggregated value is less than the predetermined threshold to FALSE (0). This process is included in the second statistical process.

[0099] The selection unit 37 has the same configuration as the selection unit 17 in the first embodiment. The selection unit 17 selects a λ number based on a preset specific rule. The selection unit 17 selects a combination of feature quantities in which the cell of the selected λ number is set to TRUE (1) based on the specific rule.

[0100] The output unit 39 has the same configuration as the output unit 19 of the first embodiment. The output unit 39 outputs information regarding the feature amount selected by the selection unit 27 (also referred to as selection information). The selection information output from the output unit 39 is used as a condition for extracting the feature amount from the sensor data measured by a measuring device or the like. For example, the selection information may be stored in a storage unit (not shown) by the selection unit 37. For example, the feature amount extracted according to the selection information is used for learning an estimation model for estimating a physical state or the like. For example, the feature amount to be extracted is extracted from the sensor data measured by a measuring device or the like worn by the user whose physical state is to be estimated.

[0101] (Operation) Next, the operation of the feature amount selection device 30 of the present embodiment will be described with reference to the drawings. FIGS. 16 to 17 are flowcharts for explaining an example of the operation of the feature amount selection device 30. In the description using the flowcharts of FIGS. 16 to 17, the feature amount selection device 30 will be described as the operating entity.

[0102] In FIG. 16, first, the feature amount selection device 30 acquires N data sets (step S311). The number of the data set corresponds to the number of the explanatory variables (feature amounts) included in the data set (feature amount number).

[0103] Next, the feature amount selection device 30 sets the feature amount number n to 1 (step S312). n is the number of the data set (feature amount).

[0104] Next, the feature amount selection device 30 excludes the data of the n-th subject (step S313).

[0105] Next, the feature amount selection device 30 performs lasso regression on the N - 1 data sets from which the data of the n-th subject has been excluded (step S314).

[0106] Next, the feature amount selection device 30 increments the feature amount number n by 1 (+1) (step S315).

[0107] Here, when the feature quantity number n is smaller than the number N of datasets (Yes in step S316), the process returns to step S313. On the other hand, when the feature quantity number n is equal to or greater than the number N of datasets (No in step S316), the process proceeds to step S321 in FIG. 17.

[0108] When No in step S316 of FIG. 16, the feature quantity selection device 30 executes the second statistical process (step S321). As the first stage of the second statistical process, the feature quantity selection device 30 generates a second matrix in which the average value of each cell is assigned to each cell of the first matrix of a plurality of patterns generated for each subject. As the second stage of the second statistical process, the feature quantity selection device 30 sets a logical value (0, 1) for each cell with respect to the generated second matrix as the second statistical process. For example, the feature quantity selection device 30 sets a cell whose average value of the aggregated value is equal to or greater than a predetermined threshold to TRUE (1) for each cell of the second matrix. On the other hand, the feature quantity selection device 30 sets a cell whose average value of the aggregated value is less than the predetermined threshold to FALSE (0).

[0109] Next, the feature quantity selection device 30 selects a λ number based on a specific rule according to the result of the second statistical process (step S322).

[0110] Next, the feature quantity selection device 30 selects a combination of feature quantities corresponding to the selected λ number (step S323).

[0111] Next, the feature quantity selection device 30 outputs information (selection information) regarding the selected feature quantity (step S324). The selection information output from the feature quantity selection device 30 is used as a condition for extracting feature quantities from sensor data measured by a measuring device or the like.

[0112] As described above, the feature quantity selection device according to the present embodiment includes an acquisition unit, a construction unit, an analysis unit, a statistics unit, a selection unit, and an output unit. The acquisition unit acquires a plurality of data sets. The construction unit constructs a plurality of re-extracted data sets by changing the distribution of the data included in the data sets. The analysis unit analyzes the plurality of re-extracted data sets using the method of lasso regression. The statistics unit aggregates the values of the elements included in the plurality of re-extracted data sets according to the analysis results of the plurality of re-extracted data sets. The statistics unit sets a logical value for the elements included in the plurality of re-extracted data sets according to the aggregation result of the values of the elements. The selection unit selects a combination of feature quantities according to the value of the logical value set for the elements according to a preset specific rule. The output unit outputs selection information regarding the selected combination of feature quantities.

[0113] In the present embodiment, the analysis unit executes lasso regression for each of the plurality of re-extracted data sets for each of the plurality of preset regularization parameters. The analysis unit generates a first matrix of a plurality of patterns composed of columns corresponding to the regularization parameters used in the lasso regression and rows corresponding to the feature quantities. The statistics unit aggregates the values of the elements for each cell constituting the first matrix of the plurality of patterns. The statistics unit executes a second statistical process of generating a second matrix in which 1 is set as the second logical value for cells whose average value of the element values is equal to or greater than a predetermined threshold, and 0 is set as the second logical value for cells whose average value of the element values is less than the predetermined threshold. The selection unit selects a column of the second matrix according to a preset specific rule and selects a combination of feature quantities corresponding to the selected column.

[0114] According to the present embodiment, by changing the distribution of the data and performing lasso regression, it is possible to obtain feature quantities having average values closer to the true value as compared with the case of directly receiving the influence of outliers or extreme values. Therefore, according to the present embodiment, it is possible to select feature quantities with high robustness against outliers or extreme values.

[0115] (Fourth Embodiment) Next, a feature quantity selection device according to the fourth embodiment will be described with reference to the drawings. The feature quantity selection device of this embodiment has a simplified configuration compared to the feature quantity selection devices of the first to third embodiments.

[0116] FIG. 18 is a block diagram showing an example of the configuration of a feature quantity selection device 40 according to this embodiment. The feature quantity selection device 40 includes an acquisition unit 41, a construction unit 42, an analysis unit 43, a statistics unit 45, a selection unit 47, and an output unit 49.

[0117] The acquisition unit 41 acquires a plurality of data sets. The construction unit 42 constructs a plurality of re-extracted data sets by changing the distribution of the data included in the data set. The analysis unit 43 analyzes the plurality of re-extracted data sets using the method of lasso regression. The statistics unit 45 aggregates the values of the elements included in the plurality of re-extracted data sets according to the analysis results of the plurality of re-extracted data sets. The statistics unit 45 sets a logical value for the elements included in the plurality of re-extracted data sets according to the aggregation result of the values of the elements. The selection unit 47 selects a combination of feature quantities according to the value of the logical value set for the elements according to a preset specific rule. The output unit 49 outputs selection information regarding the selected combination of feature quantities.

[0118] According to this embodiment, by changing the distribution of the data and performing lasso regression, it is possible to obtain a feature quantity of an average value closer to the true value as compared with the case of being directly affected by outliers or extreme values. Therefore, according to this embodiment, it is possible to select a feature quantity with high robustness against outliers or extreme values.

[0119] (Fifth Embodiment) Next, a learning system according to the fifth embodiment will be described with reference to the drawings. The learning system of this embodiment executes learning using the feature quantities selected by the feature quantity selection device of the first to fourth embodiments.

[0120] FIG. 19 is a block diagram showing an example of the configuration of the learning system 5 according to the present embodiment. The learning system 5 includes a gait measurement device 50 and a learning device 55. The gait measurement device 50 and the learning device 55 may be connected by wire or wirelessly. The gait measurement device 50 and the learning device 55 may be configured as a single device. Further, the learning system 5 may be configured only by the learning device 55 excluding the gait measurement device 50 from the configuration of the learning system 5. Although only one gait measurement device 50 is shown in FIG. 19, two gait measurement devices 50 may be arranged, one on each of the left and right feet. Further, the learning device 55 may be configured to execute learning using the feature amount data that has been generated by the gait measurement device 50 in advance and stored in a database without being connected to the gait measurement device 50.

[0121] The gait measurement device 50 is installed on at least one of the left and right feet. The gait measurement device 50 has the same configuration as the gait measurement device 50 of the first embodiment. The gait measurement device 50 includes an acceleration sensor and an angular velocity sensor. The gait measurement device 50 converts the measured physical quantity into digital data (also referred to as sensor data). The gait measurement device 50 generates gait waveform data for one normalized walking cycle from the time-series data of the sensor data. The gait measurement device 50 generates feature amount data used for estimating the physical state. The gait measurement device 50 transmits the generated feature amount data to the learning device 55. Note that the gait measurement device 50 may be configured to transmit the feature amount data to a database (not shown) accessed by the learning device 55. The feature amount data accumulated in the database is used for learning by the learning device 55.

[0122] The learning device 55 receives feature amount data from the gait measurement device 50. The feature amount data received by the learning device 55 includes the feature amounts selected by the feature amount selection device of the first to fourth embodiments. When using the feature amount data stored in a database (not shown), the learning device 55 receives the feature amount data from the database. The learning device 55 performs learning using the received feature amount data. For example, the learning device 55 learns teacher data with the feature amount data extracted from the walking waveform data of a plurality of subjects as explanatory variables and the value regarding the physical state corresponding to the feature amount data as objective variables. There is no particular limitation on the learning algorithm executed by the learning device 55. The learning device 55 generates an estimation model learned using teacher data regarding a plurality of subjects. The learning device 55 stores the generated estimation model. The estimation model learned by the learning device 55 may be stored in a storage device external to the learning device 55.

[0123] 〔Learning device〕 Next, the details of the learning device 55 will be described with reference to the drawings. FIG. 20 is a block diagram showing an example of the detailed configuration of the learning device 55. The learning device 55 includes a reception unit 551, a learning unit 553, and a storage unit 555.

[0124] The reception unit 551 receives feature amount data from the gait measurement device 50. The reception unit 551 outputs the received feature amount data to the learning unit 553. The reception unit 551 may receive the feature amount data from the gait measurement device 50 via a wired connection such as a cable, or may receive the feature amount data from the gait measurement device 50 via wireless communication. For example, the reception unit 551 is configured to receive the feature amount data from the gait measurement device 50 via a wireless communication function (not shown) compliant with standards such as Bluetooth (registered trademark) and WiFi (registered trademark). Note that the communication function of the reception unit 551 may comply with standards other than Bluetooth (registered trademark) and WiFi (registered trademark).

[0125] The learning unit 553 acquires feature quantity data from the receiving unit 551. The learning unit 553 executes learning using the acquired feature quantity data. For example, the learning unit 553 uses, as an explanatory variable, the feature quantity data extracted regarding the walking style of a subject, and learns, as teacher data, a data set in which the physical state of the subject is the objective variable. For example, the learning unit 553 targets, for learning, physical states such as the grip strength, whole-body muscle strength, lower-limb muscle strength, movement ability, dynamic balance, and static balance of a subject. For example, the learning unit 553 generates an estimation model that estimates the physical state in response to the input of feature quantity data, which has been learned for a plurality of users. For example, the learning unit 553 generates an estimation model that performs estimation according to attributes, using explanatory variables including attribute data such as gender, age, height, and weight. The learning unit 553 stores the estimation model learned for a plurality of subjects in the storage unit 555.

[0126] For example, the learning unit 553 executes learning using the algorithm of linear regression. For example, the learning unit 553 executes learning using the algorithm of support vector machine (SVM). For example, the learning unit 553 executes learning using the algorithm of Gaussian process regression (GPR). For example, the learning unit 553 executes learning using the algorithm of random forest (RF). For example, the learning unit 553 may execute unsupervised learning that classifies the subject who is the source of the feature quantity data according to the feature quantity data. The algorithm of the learning executed by the learning unit 553 is not particularly limited.

[0127] The learning unit 553 may execute learning using, as an explanatory variable, the walking waveform data (sensor data) for one walking cycle. For example, the learning unit 553 executes supervised learning using, as explanatory variables, the walking waveform data of the acceleration in three axial directions, the angular velocity around three axes, and the angle (posture angle) around three axes, and using, as the objective variable, the correct value of the physical state to be estimated.

[0128] FIG. 21 is a conceptual diagram for explaining learning for generating an estimation model. FIG. 21 is a conceptual diagram showing an example in which a learning unit 553 is caused to learn using, as teacher data, a data set of feature amounts F1 to Fn that are explanatory variables and a score related to a physical state that is an objective variable. For example, the learning unit 553 learns data related to a plurality of subjects and generates an estimation model that outputs an output (estimated value) related to the physical state of a subject in response to an input of a feature amount extracted from sensor data.

[0129] The storage unit 555 stores an estimation model learned with respect to a plurality of subjects. For example, the storage unit 555 stores an estimation model for estimating a physical state learned with respect to a plurality of subjects. For example, the estimation model stored in the storage unit 555 is used for estimating the physical state by the physical state estimation system of the sixth embodiment described later.

[0130] As described above, the learning system of the present embodiment includes a gait measurement device and a learning device. The gait measurement device acquires time-series data of sensor data related to the movement of a foot. The gait measurement device extracts gait waveform data for one walking cycle from the time-series data of the sensor data and normalizes the extracted gait waveform data. The gait measurement device extracts a feature amount related to the physical state of the estimation target from the normalized gait waveform data. The gait measurement device extracts the feature amounts selected by the feature amount selection device of the first to fourth embodiments. The gait measurement device generates feature amount data including the extracted feature amounts. The gait measurement device outputs the generated feature amount data to the learning device.

[0131] The learning device includes a reception unit, a learning unit, and a storage unit. The reception unit acquires the feature amount data generated by the gait measurement device. The learning unit executes learning using the feature amount data. The learning unit generates an estimation model that outputs a physical state in response to an input of a feature amount extracted from time-series data of sensor data measured as the user walks. The estimation model generated by the learning unit is stored in the storage unit.

[0132] The learning system of this embodiment generates an estimation model using the feature amount data measured by a gait measurement device. The learning system of this embodiment executes learning using the feature amounts selected by the feature amount selection devices of the first to fourth embodiments. Therefore, according to this aspect, it is possible to generate an estimation model that can appropriately estimate the physical state in daily life using highly robust feature amounts.

[0133] (Sixth Embodiment) Next, a physical state estimation system according to the sixth embodiment will be described with reference to the drawings. The physical state estimation system of this embodiment measures sensor data related to the movement of the feet according to the user's walking. The physical state estimation system of this embodiment estimates the physical state of the user using the measured sensor data. For example, the physical state estimation system of this embodiment estimates, as the physical state, muscle strength indexes such as grip strength and knee extension strength, dynamic balance, lower limb muscle strength, mobility, static balance, and the like. Note that the sensor data may be sensor data including features related to gait measured using motion capture, smart apparel, or the like.

[0134] (Configuration) FIG. 22 is a block diagram showing an example of the configuration of a physical state estimation system 6 according to this embodiment. The physical state estimation system 6 includes a gait measurement device 60 and an estimation device 63. In this embodiment, an example in which the gait measurement device 60 and the estimation device 63 are configured as separate hardware will be described. For example, the gait measurement device 60 is installed in the footwear or the like of a subject (user) whose physical state is to be estimated. For example, the function of the estimation device 63 is installed in a mobile terminal carried by the subject (user). Hereinafter, the configurations of the gait measurement device 60 and the estimation device 63 will be described individually.

[0135] 〔Gait Measurement Device〕 FIG. 23 is a block diagram showing an example of the configuration of the gait measurement device 60. The gait measurement device 60 includes a sensor 61 and a feature amount data generation unit 62. In the present embodiment, an example in which the sensor 61 and the feature amount data generation unit 62 are integrated is given. The sensor 61 and the feature amount data generation unit 62 may be provided as separate devices.

[0136] As shown in FIG. 23, the sensor 61 includes an acceleration sensor 611 and an angular velocity sensor 612. FIG. 23 shows an example in which the acceleration sensor 611 and the angular velocity sensor 612 are included in the sensor 61. The sensor 61 may include sensors other than the acceleration sensor 611 and the angular velocity sensor 612. Explanation of sensors other than the acceleration sensor 611 and the angular velocity sensor 612 that may be included in the sensor 61 is omitted.

[0137] The acceleration sensor 611 is a sensor that measures the acceleration in three axial directions (also called spatial acceleration). The acceleration sensor 611 measures the acceleration (also called spatial acceleration) as a physical quantity related to the movement of the foot. The acceleration sensor 611 outputs the measured acceleration to the feature amount data generation unit 62. For example, as the acceleration sensor 611, a piezoelectric type, piezoresistive type, capacitive type, or other type of sensor can be used. The sensor used as the acceleration sensor 611 is not limited to the measurement method as long as it can measure the acceleration.

[0138] The angular velocity sensor 612 is a sensor that measures the angular velocity around three axes (also called spatial angular velocity). The angular velocity sensor 612 measures the angular velocity (also called spatial angular velocity) as a physical quantity related to the movement of the foot. The angular velocity sensor 612 outputs the measured angular velocity to the feature amount data generation unit 62. For example, as the angular velocity sensor 612, a vibration type, capacitive type, or other type of sensor can be used. The sensor used as the angular velocity sensor 612 is not limited to the measurement method as long as it can measure the angular velocity.

[0139] The sensor 61 is realized by, for example, an inertial measurement device that measures acceleration and angular velocity. As an example of an inertial measurement device, an IMU (Inertial Measurement Unit) can be mentioned. The IMU includes an acceleration sensor 611 that measures acceleration in three axial directions and an angular velocity sensor 612 that measures angular velocity around three axes. The sensor 61 may be realized by an inertial measurement device such as a VG (Vertical Gyro) or an AHRS (Attitude Heading). Further, the sensor 61 may be realized by a GPS / INS (Global Positioning System / Inertial Navigation System). The sensor 61 may be realized by a device other than an inertial measurement device as long as it can measure physical quantities related to the movement of the foot.

[0140] FIG. 24 is a conceptual diagram showing an example in which the gait measurement device 60 is disposed inside the right shoe 600. In the example of FIG. 24, the gait measurement device 60 is installed at a position that hits the back side of the arch of the foot. For example, the gait measurement device 60 is disposed on an insole inserted into the shoe 600. For example, the gait measurement device 60 may be disposed on the bottom surface of the shoe 600. For example, the gait measurement device 60 may be embedded in the main body of the shoe 600. The gait measurement device 60 may be detachable from the shoe 600 or may not be detachable from the shoe 600. The gait measurement device 60 may be installed at a position other than the back side of the arch of the foot as long as it can measure sensor data related to the movement of the foot. Further, the gait measurement device 60 may be installed on socks worn by the user or on ornaments such as anklets worn by the user. Further, the gait measurement device 60 may be directly attached to or embedded in the foot. FIG. 24 shows an example in which the gait measurement device 60 is installed in the right shoe 600. The gait measurement device 60 may be installed in the shoes 600 for both feet.

[0141] In the example of FIG. 24, a local coordinate system including the x-axis in the left-right direction, the y-axis in the front-back direction, and the z-axis in the up-down direction is set with the gait measurement device 60 (sensor 61) as a reference. The x-axis is positive to the left, the y-axis is positive to the rear, and the z-axis is positive upward. The directions of the axes set for the sensor 61 may be the same for the left and right feet, or may be different for the left and right feet. For example, when sensors 61 produced with the same specifications are arranged in the left and right shoes 600, the up-down directions (directions in the Z-axis direction) of the sensors 61 arranged in the left and right shoes 600 are the same. In that case, the three axes of the local coordinate system set for the sensor data derived from the left foot and the three axes of the local coordinate system set for the sensor data derived from the right foot are the same on the left and right.

[0142] FIG. 25 is a conceptual diagram for explaining the local coordinate system (x-axis, y-axis, z-axis) set for the gait measurement device 60 (sensor 61) installed on the back side of the arch of the foot and the world coordinate system (X-axis, Y-axis, Z-axis) set with respect to the ground. In the world coordinate system (X-axis, Y-axis, Z-axis), with the user standing upright facing the traveling direction, the lateral direction of the user is the X-axis direction (positive to the left), the direction of the user's back is the Y-axis direction (positive to the rear), and the direction of gravity is the Z-axis direction (positive vertically upward). Note that the example of FIG. 25 conceptually shows the relationship between the local coordinate system (x-axis, y-axis, z-axis) and the world coordinate system (X-axis, Y-axis, Z-axis), and does not accurately show the relationship between the local coordinate system and the world coordinate system that varies according to the user's walking.

[0143] FIG. 26 is a conceptual diagram for explaining the plane set for the human body (also called the human body plane). In the present embodiment, a sagittal plane that divides the body into left and right, a coronal plane that divides the body into front and back, and a horizontal plane that divides the body into upper and lower are defined. Note that as shown in FIG. 26, in the state of standing upright with the center line of the foot facing the traveling direction, the world coordinate system and the local coordinate system coincide. In the present embodiment, the rotation in the sagittal plane with the x-axis as the rotation axis is defined as roll, the rotation in the coronal plane with the y-axis as the rotation axis is defined as pitch, and the rotation in the horizontal plane with the z-axis as the rotation axis is defined as yaw. Also, the rotation angle in the sagittal plane with the x-axis as the rotation axis is defined as the roll angle, the rotation angle in the coronal plane with the y-axis as the rotation axis is defined as the pitch angle, and the rotation angle in the horizontal plane with the z-axis as the rotation axis is defined as the yaw angle.

[0144] As shown in FIG. 23, the feature amount data generation unit 62 (also referred to as a feature amount data generation device) includes an acquisition unit 621, a normalization unit 622, an extraction unit 623, a generation unit 625, and a feature amount data output unit 627. For example, the feature amount data generation unit 62 is realized by a microcomputer or a microcontroller that performs overall control and data processing of the gait measurement device 60. For example, the feature amount data generation unit 62 includes a CPU (Central Processing Unit), a RAM (Random Access Memory), a ROM (Read Only Memory), a flash memory, and the like. The feature amount data generation unit 62 controls the acceleration sensor 611 and the angular velocity sensor 612 to measure the angular velocity and acceleration. For example, the feature amount data generation unit 62 may be mounted on the side of a mobile terminal (not shown) carried by a subject (user).

[0145] The acquisition unit 621 acquires the acceleration in the three-axis directions from the acceleration sensor 611. Also, the acquisition unit 621 acquires the angular velocity around the three axes from the angular velocity sensor 612. For example, the acquisition unit 621 performs AD conversion (Analog-to-Digital Conversion) on physical quantities (analog data) such as the acquired angular velocity and acceleration. Note that the physical quantities (analog data) measured by the acceleration sensor 611 and the angular velocity sensor 612 may be converted into digital data in each of the acceleration sensor 611 and the angular velocity sensor 612. The acquisition unit 621 outputs the converted digital data (also referred to as sensor data) to the normalization unit 622. The acquisition unit 621 may be configured to store the sensor data in a storage unit (not shown). The sensor data includes at least acceleration data converted into digital data and angular velocity data converted into digital data. The acceleration data includes an acceleration vector in the three-axis directions. The angular velocity data includes an angular velocity vector around the three axes. The acquisition time of these data is associated with the acceleration data and the angular velocity data. Also, the acquisition unit 621 may apply corrections such as mounting error, temperature correction, and linearity correction to the acceleration data and the angular velocity data.

[0146] The normalization unit 622 acquires sensor data from the acquisition unit 621. The normalization unit 622 extracts time-series data for one walking cycle (also referred to as walking waveform data) from the time-series data of the acceleration in three axial directions and the angular velocity around three axes included in the sensor data. The normalization unit 622 normalizes the time of the extracted walking waveform data for one walking cycle to a walking cycle of 0 to 100% (also referred to as the first normalization). Timings such as 1% and 10% included in the walking cycle of 0 to 100% are also referred to as walking phases. Further, the normalization unit 622 normalizes (also referred to as the second normalization) the walking waveform data for one walking cycle that has been first-normalized so that the stance phase is 60% and the swing phase is 40%. The stance phase is a period during which at least a part of the sole of the foot is in contact with the ground. The swing phase is a period during which the sole of the foot is separated from the ground. By performing the second normalization on the walking waveform data, it is possible to suppress the deviation of the walking phase from which the feature amount is extracted from being disturbed by external disturbances.

[0147] FIG. 27 is a conceptual diagram for explaining one walking cycle with the right foot as a reference. One walking cycle with the left foot as a reference is the same as that of the right foot. The horizontal axis in FIG. 27 represents one walking cycle of the right foot starting from the point in time when the heel of the right foot touches the ground and ending at the point in time when the heel of the right foot touches the ground next. The horizontal axis in FIG. 27 is first-normalized with the walking cycle being 100%. Further, the horizontal axis in FIG. 27 is second-normalized so that the stance phase is 60% and the swing phase is 40%. One walking cycle of one foot is roughly divided into a stance phase in which at least a part of the sole of the foot is in contact with the ground and a swing phase in which the sole of the foot is separated from the ground. The stance phase is further subdivided into a load response period T1, a mid-stance period T2, a terminal stance period T3, and a pre-swing period T4. The swing phase is further subdivided into an initial swing period T5, a mid-swing period T6, and a terminal swing period T7. Note that FIG. 27 is an example and does not limit the periods constituting one walking cycle or the names of those periods.

[0148] As shown in FIG. 27, in walking, a plurality of events (also referred to as walking events) occur. E1 represents the event of the heel of the right foot touching the ground (heel contact) (HC: Heel Contact). E2 represents the event of the tip of the left foot leaving the ground while the sole of the right foot is in contact with the ground (opposite toe off) (OTO: Opposite Toe Off). E3 represents the event of the heel of the right foot rising while the sole of the right foot is in contact with the ground (heel rise) (HR: Heel Rise). E4 is the event of the heel of the left foot touching the ground (opposite heel strike) (OHS: Opposite Heel Strike). E5 represents the event of the tip of the right foot leaving the ground while the sole of the left foot is in contact with the ground (toe off) (TO: Toe Off). E6 represents the event of the left and right feet crossing while the sole of the left foot is in contact with the ground (foot adjacent) (FA: Foot Adjacent). E7 represents the event of the tibia of the right foot becoming substantially perpendicular to the ground while the sole of the left foot is in contact with the ground (tibia vertical) (TV: Tibia Vertical). E8 represents the event of the heel of the right foot touching the ground (heel contact) (HC: Heel Contact). E8 corresponds to the end point of the walking cycle starting from E1 and also corresponds to the starting point of the next walking cycle. Note that FIG. 27 is an example and does not limit the events occurring in walking or the names of those events.

[0149] FIG. 28 is a diagram for explaining an example of detecting heel contact HC and toe-off TO from time-series data (solid line) of the acceleration in the traveling direction (Y-direction acceleration). The timing of heel contact HC is the timing of the minimum peak immediately after the maximum peak appearing in the time-series data of the acceleration in the traveling direction (Y-direction acceleration). The maximum peak serving as a mark for the timing of heel contact HC corresponds to the maximum peak of the walking waveform data for one walking cycle. The interval between consecutive heel contacts HC is one walking cycle. The timing of toe-off TO is the rising timing of the maximum peak that appears after the period of the stance phase where no variation appears in the time-series data of the acceleration in the traveling direction (Y-direction acceleration). FIG. 28 also shows time-series data (broken line) of the roll angle (angular velocity around the X axis). The timing at the midpoint between the timing when the roll angle is minimum and the timing when the roll angle is maximum corresponds to the mid-stance phase. For example, parameters such as walking speed, step length, cadence, internal rotation / external rotation, plantar flexion / dorsiflexion (also called gait parameters) can be obtained with reference to the mid-stance phase.

[0150] FIG. 29 is a diagram for explaining an example of the walking waveform data normalized by the normalization unit 622. The normalization unit 622 detects heel contact HC and toe-off TO from the time-series data of the acceleration in the traveling direction (Y-direction acceleration). The normalization unit 622 extracts the interval between consecutive heel contacts HC as the walking waveform data for one walking cycle. The normalization unit 622 converts the horizontal axis (time axis) of the walking waveform data for one walking cycle to a walking cycle of 0 to 100% by the first normalization. In FIG. 29, the walking waveform data after the first normalization is shown by a broken line. In the walking waveform data (broken line) after the first normalization, the timing of toe-off TO is deviated from 60%.

[0151] In the example of FIG. 29, the normalization unit 622 normalizes the section from the heel contact HC with a walking phase of 0% to the toe-off TO following the heel contact HC to 0 to 60%. Further, the normalization unit 622 normalizes the section from the toe-off TO to the heel contact HC with a walking phase of 100% following the toe-off TO to 60 to 100%. As a result, the walking waveform data for one walking cycle is normalized into a section with a walking cycle of 0 to 60% (stance phase) and a section with a walking cycle of 60 to 100% (swing phase). In FIG. 8, the walking waveform data after the second normalization is shown by a solid line. In the walking waveform data (solid line) after the second normalization, the timing of the toe-off TO coincides with 60%.

[0152] FIGS. 28 to 29 show an example of extracting / normalizing the walking waveform data for one walking cycle based on the acceleration in the traveling direction (Y-direction acceleration). Regarding the acceleration / angular velocity other than the acceleration in the traveling direction (Y-direction acceleration), the normalization unit 622 extracts / normalizes the walking waveform data for one walking cycle in accordance with the walking cycle of the acceleration in the traveling direction (Y-direction acceleration). Further, the normalization unit 622 may generate time-series data of the angles around the three axes by integrating the time-series data of the angular velocities around the three axes. In that case, the normalization unit 622 also extracts / normalizes the walking waveform data for one walking cycle in accordance with the walking cycle of the acceleration in the traveling direction (Y-direction acceleration) regarding the angles around the three axes.

[0153] The normalization unit 622 may extract / normalize the walking waveform data for one walking cycle based on the acceleration / angular velocity other than the forward acceleration (Y-direction acceleration) (the drawings are omitted). For example, the normalization unit 622 may detect the heel contact HC and the toe-off TO from the time series data of the vertical acceleration (Z-direction acceleration). The timing of the heel contact HC is the timing of the sharp minimum peak appearing in the time series data of the vertical acceleration (Z-direction acceleration). At the timing of the sharp minimum peak, the value of the vertical acceleration (Z-direction acceleration) becomes almost 0. The minimum peak serving as a mark for the timing of the heel contact HC corresponds to the minimum peak of the walking waveform data for one walking cycle. The interval between consecutive heel contacts HC is one walking cycle. The timing of the toe-off TO is the timing of the inflection point during the gentle increase after the time series data of the vertical acceleration (Z-direction acceleration) passes through a section with little variation after the maximum peak immediately following the heel contact HC. Further, the normalization unit 622 may extract / normalize the walking waveform data for one walking cycle based on both the forward acceleration (Y-direction acceleration) and the vertical acceleration (Z-direction acceleration). Further, the normalization unit 622 may extract / normalize the walking waveform data for one walking cycle based on the acceleration, angular velocity, angle, etc. other than the forward acceleration (Y-direction acceleration) and the vertical acceleration (Z-direction acceleration).

[0154] The extraction unit 623 acquires the walking waveform data for one walking cycle normalized by the normalization unit 622. The extraction unit 623 extracts the feature amounts used for the estimation of the body state from the walking waveform data for one walking cycle. The extraction unit 623 extracts the feature amounts for each walking phase cluster from the walking phase clusters obtained by integrating the temporally consecutive walking phases based on the preset conditions. The walking phase cluster includes at least one walking phase. The walking phase cluster also includes a single walking phase. The walking waveform data and the walking phases from which the feature amounts used for the estimation of the body state are extracted will be described later.

[0155] FIG. 30 is a conceptual diagram for explaining the extraction of feature quantities for estimating the physical state from the walking waveform data for one walking cycle. For example, the extraction unit 623 extracts the temporally continuous walking phases I to I+m as the walking phase cluster C (I and m are natural numbers). The walking phase cluster C includes m walking phases (constituent elements). That is, the number of walking phases (constituent elements) constituting the walking phase cluster C (also referred to as the number of constituent elements) is m. In FIG. 30, an example in which the walking phase is an integer value is given, but the walking phase may be subdivided down to the decimal point. When the walking phase is subdivided down to the decimal point, the number of constituent elements of the walking phase cluster C becomes a number corresponding to the number of data points in the interval of the walking phase cluster. The extraction unit 623 extracts feature quantities from each of the walking phases I to I+m. When the walking phase cluster C is composed of a single walking phase J, the extraction unit 623 extracts a feature quantity from the single walking phase J (J is a natural number).

[0156] The generation unit 625 applies a feature quantity composition formula to the feature quantities (first feature quantities) extracted from each of the walking phases constituting the walking phase cluster to generate a feature quantity (second feature quantity) of the walking phase cluster. The feature quantity composition formula is a calculation formula set in advance for generating the feature quantity of the walking phase cluster. For example, the feature quantity composition formula is a calculation formula related to arithmetic operations. For example, the second feature quantity calculated using the feature quantity composition formula is the integrated average value, arithmetic average value, slope, variation, etc. of the first feature quantity in each walking phase included in the walking phase cluster. For example, the generation unit 625 applies a calculation formula for calculating the slope and variation of the first feature quantity extracted from each of the walking phases constituting the walking phase cluster as the feature quantity composition formula. For example, when the walking phase cluster is composed of a single walking phase, since the slope and variation cannot be calculated, a feature quantity composition formula for calculating the integrated average value, arithmetic average value, etc. may be used.

[0157] The feature amount data output unit 627 outputs the feature amount data for each walking phase cluster generated by the generation unit 625. The feature amount data output unit 627 outputs the feature amount data of the generated walking phase cluster to the estimation device 63 that uses the feature amount data.

[0158] 〔Estimation Device〕 FIG. 31 is a block diagram showing an example of the configuration of the estimation device 63. The estimation device 63 includes a data acquisition unit 631, a storage unit 632, an estimation unit 633, and an output unit 635.

[0159] The data acquisition unit 631 acquires the feature amount data from the gait measurement device 60. The data acquisition unit 631 outputs the received feature amount data to the estimation unit 633. The data acquisition unit 631 may receive the feature amount data from the gait measurement device 60 via a wired connection such as a cable, or may receive the feature amount data from the gait measurement device 60 via wireless communication. For example, the data acquisition unit 631 is configured to receive the feature amount data from the gait measurement device 60 via a wireless communication function (not shown) conforming to a standard such as Bluetooth (registered trademark) or WiFi (registered trademark). Note that the communication function of the data acquisition unit 631 may conform to a standard other than Bluetooth (registered trademark) or WiFi (registered trademark).

[0160] The storage unit 632 stores an estimation model for estimating the physical state using the feature amount data extracted from the walking waveform data. The storage unit 632 stores an estimation model obtained by learning the relationship between the feature amount data related to the physical states of a plurality of subjects and their physical states. For example, the storage unit 632 stores an estimation model for estimating the physical state learned for a plurality of subjects. For example, the storage unit 632 may store an estimation model according to the attributes.

[0161] The estimation model may be stored in the storage unit 632 at the time of factory shipment of the product or at the time of calibration before the user uses the physical state estimation system 6. For example, it may be configured to use the estimation model stored in a storage device such as an external server. In that case, it may be configured to use the estimation model via an interface (not shown) connected to the storage device.

[0162] The estimation unit 633 acquires feature quantity data from the data acquisition unit 631. The estimation unit 633 uses the acquired feature quantity data to execute the estimation of the physical state. The estimation unit 633 inputs the feature quantity data into the estimation model stored in the storage unit 632. The estimation unit 633 outputs an estimation result corresponding to the physical state output from the estimation model. When using an estimation model stored in an external storage device constructed in a cloud, a server, or the like, the estimation unit 633 is configured to use the estimation model via an interface (not shown) connected to the storage device.

[0163] The output unit 635 outputs the estimation result of the physical state by the estimation unit 633. For example, the output unit 635 causes the estimation result of the physical state to be displayed on the screen of the mobile terminal of the subject (user). For example, the output unit 635 outputs the estimation result to an external system or the like that uses the estimation result. There is no particular limitation on the use of the physical state output from the estimation device 63.

[0164] For example, the estimation device 63 is connected to an external system or the like constructed in a cloud or a server via a mobile terminal (not shown) carried by a subject (user). The mobile terminal (not shown) is a portable communication device. For example, the mobile terminal is a portable communication device having a communication function such as a smartphone, a smartwatch, or a mobile phone. For example, the estimation device 63 is connected to the mobile terminal via a wired connection such as a cable. For example, the estimation device 63 is connected to the mobile terminal via wireless communication. For example, the estimation device 63 is connected to the mobile terminal via a wireless communication function (not shown) compliant with standards such as Bluetooth (registered trademark) or WiFi (registered trademark). Note that the communication function of the estimation device 63 may comply with standards other than Bluetooth (registered trademark) or WiFi (registered trademark). The estimation result of the physical state may be used by an application installed in the mobile terminal. In that case, the mobile terminal executes processing using the estimation result by application software or the like installed in the mobile terminal.

[0165] FIG. 32 is a conceptual diagram showing an example in which feature amounts F1 to Fn extracted from sensor data measured along with a user's walking are input to an estimation model 651 pre-constructed for estimating a physical state by a learning system according to a fifth embodiment, and a score related to the physical state is output. The estimation model 651 outputs a score of the physical state in response to the input of the feature amounts F1 to Fn. For example, the estimation model 651 is generated by learning using teacher data having the feature amounts F1 to Fn used for estimating the physical state as explanatory variables and the physical state as an objective variable. There is no limitation on the estimation result of the estimation model 651 as long as an estimation result related to the physical state is output in response to the input of the feature amount data for estimating the physical state. For example, the estimation model 651 may be a model for estimating the physical state using attributes such as gender, age, height, and weight as explanatory variables in addition to the feature amounts F1 to Fn used for estimating the physical state.

[0166] For example, in the storage unit 632, an estimation model for estimating the physical state is stored using the multiple regression prediction method. For example, in the storage unit 632, parameters for estimating the score S of the physical state are stored using the following Equation 1. S = f1×F1 + f2×F2 + ··· + fn×Fn + f0 ··· (1) In the above Equation 1, F1, F2, ···, Fn are feature amounts for each gait phase cluster used for estimating the physical state. f1, f2, ···, fn are coefficients multiplied by F1, F2, ···, Fn. f0 is a constant term. For example, in the storage unit 632, coefficients such as f1, f2, ···, fn are stored.

[0167] (Operation) Next, the operation of the physical state estimation system 6 will be described with reference to the drawings. Here, the gait measurement device 60 and the estimation device 63 included in the physical state estimation system 6 will be individually described. Regarding the gait measurement device 60, the operation of the feature amount data generation unit 62 included in the gait measurement device 60 will be described.

[0168] 〔Gait measurement device〕 FIG. 33 is a flowchart for explaining the operation of the feature amount data generation unit 62 included in the gait measurement device 60. In the explanation along the flowchart of FIG. 33, the feature amount data generation unit 62 will be described as the main operating body.

[0169] In FIG. 33, first, the feature amount data generation unit 62 acquires time-series data of sensor data related to the gait (step S601).

[0170] Next, the feature amount data generation unit 62 extracts gait waveform data for one gait cycle from the time-series data of the sensor data (step S602). The feature amount data generation unit 62 detects heel strike and toe-off from the time-series data of the sensor data. The feature amount data generation unit 62 extracts the time-series data of the section between consecutive heel strikes as the gait waveform data for one gait cycle.

[0171] Next, the feature amount data generation unit 62 normalizes the extracted gait waveform data for one gait cycle (step S603). The feature amount data generation unit 62 normalizes the gait waveform data for one gait cycle to a gait cycle of 0 to 100% (first normalization). Further, the feature amount data generation unit 62 normalizes the ratio of the stance phase to the swing phase of the gait waveform data for one gait cycle that has been first-normalized to 60:40 (second normalization).

[0172] Next, the feature amount data generation unit 62 extracts features from the gait phase used for estimating the body state with respect to the normalized gait waveform (step S604). For example, the feature amount data generation unit 62 extracts the features to be input into the estimation model constructed for each gender.

[0173] Next, the feature amount data generation unit 62 generates feature amounts for each gait phase cluster using the extracted feature amounts (step S605).

[0174] Next, the feature amount data generation unit 62 integrates the feature amounts for each gait phase cluster to generate feature amount data for one gait cycle (step S606).

[0175] Next, the feature amount data generation unit 62 outputs the generated feature amount data to the estimation device 63 (step S607).

[0176] 〔Estimation Device〕 FIG. 33 is a flowchart for explaining the operation of the estimation device 63. In the explanation along the flowchart of FIG. 33, the estimation device 63 will be described as the operating entity.

[0177] In FIG. 33, first, the estimation device 63 acquires the feature amount data generated using the sensor data regarding the gait appearance (step S631).

[0178] Next, the estimation device 63 inputs the acquired feature amount data into an estimation model for estimating the body state (step S632).

[0179] Next, the estimation device 63 estimates the physical state of the user according to the output (estimated value) from the estimation model (step S633).

[0180] Next, the estimation device 63 outputs information regarding the estimated physical state (step S634). For example, the physical state is output to a terminal device (not shown) carried by the user. For example, the physical state is output to a system that executes processing using the physical state.

[0181] (Application Example) Next, an application example according to the present embodiment will be described with reference to the drawings. In the following application example, an example is shown in which the function of the estimation device 63 installed in the portable terminal carried by the user estimates the physical state using the feature amount data measured by the gait measurement device 60 disposed on the shoe.

[0182] FIG. 35 is a conceptual diagram showing an example in which the estimation result by the estimation device 63 is displayed on the screen of the portable terminal 660 carried by the user walking while wearing the shoe 600 on which the gait measurement device 60 is disposed. FIG. 35 is an example in which information corresponding to the estimation result of the physical state using the feature amount data corresponding to the sensor data measured during the user's walking is displayed on the screen of the portable terminal 660.

[0183] FIG. 35 shows an example in which information corresponding to an estimated value related to a physical state is displayed on the screen of the mobile terminal 660. In the example of FIG. 35, as an estimation result related to the physical state, a score quantified based on a preset standard is displayed on the display unit of the mobile terminal 660. Also, in the example of FIG. 35, information related to the estimation result of the physical state, "The overall muscle strength of the whole body is decreasing.", is displayed on the display unit of the mobile terminal 660 according to the estimated value related to the physical state. Further, in the example of FIG. 35, according to the estimated value of the physical state, recommendation information corresponding to the estimation result of the overall muscle strength of the whole body, "Training A is recommended. Please watch the following video.", is displayed on the display unit of the mobile terminal 660. The user who has confirmed the information displayed on the display unit of the mobile terminal 660 can practice training that leads to the improvement of the overall muscle strength of the whole body by referring to the video of Training A and exercising according to the displayed recommendation information.

[0184] As described above, the physical state estimation system of the present embodiment includes a gait measurement device and a physical state estimation device. The gait measurement device includes a sensor and a feature amount data generation unit. The sensor has an acceleration sensor and an angular velocity sensor. The sensor measures the spatial acceleration using the acceleration sensor. The sensor measures the spatial angular velocity using the angular velocity sensor. The sensor generates sensor data related to the movement of the foot using the measured spatial acceleration and spatial angular velocity. The sensor outputs the generated sensor data to the feature amount data generation unit. The feature amount data generation unit acquires time-series data of the sensor data related to the movement of the foot. The feature amount data generation unit extracts gait waveform data for one gait cycle from the time-series data of the sensor data. The feature amount data generation unit normalizes the extracted gait waveform data. The feature amount data generation unit extracts a feature amount related to the physical state to be estimated from a gait phase cluster composed of at least one gait phase that is temporally continuous from the normalized gait waveform data. The feature amount data generation unit extracts the feature amounts selected by the feature amount selection device of the first to fourth embodiments. The feature amount data generation unit generates feature amount data including the extracted feature amounts. The feature amount data generation unit outputs the generated feature amount data.

[0185] The body state estimation device includes a data acquisition unit, a storage unit, an estimation unit, and an output unit. The data acquisition unit acquires feature quantity data including feature quantities used for estimating the user's body state, which are extracted from the characteristics of the user's gait. The storage unit stores an estimation model that outputs a body state in response to the input of the feature quantity data. The estimation unit inputs the acquired feature quantity data into the estimation model to estimate the user's body state. The output unit outputs information regarding the estimated body state.

[0186] The body state estimation system according to the present embodiment estimates the user's body state using feature quantities extracted from the characteristics of the user's gait. Therefore, according to the body state estimation system of the present embodiment, the body state can be appropriately estimated in daily life using highly robust feature quantities.

[0187] (Hardware) Here, the hardware configuration for executing the processing according to each embodiment of the present disclosure will be described by taking the information processing apparatus 90 in FIG. 36 as an example. Note that the information processing apparatus 90 in FIG. 36 is a configuration example for executing the processing of each embodiment, and does not limit the scope of the present disclosure.

[0188] As shown in FIG. 36, the information processing apparatus 90 includes a processor 91, a main storage device 92, an auxiliary storage device 93, an input / output interface 95, and a communication interface 96. In FIG. 36, the interface is abbreviated as I / F (Interface). The processor 91, the main storage device 92, the auxiliary storage device 93, the input / output interface 95, and the communication interface 96 are connected to each other via a bus 98 so as to be capable of data communication. Further, the processor 91, the main storage device 92, the auxiliary storage device 93, and the input / output interface 95 are connected to a network such as the Internet or an intranet via the communication interface 96.

[0189] Processor 91 expands the program stored in auxiliary storage device 93 or the like into main storage device 92. Processor 91 executes the program expanded in main storage device 92. In this embodiment, a configuration may be adopted in which the software program installed in information processing apparatus 90 is used. Processor 91 executes the processing according to each embodiment.

[0190] Main storage device 92 has an area where the program is expanded. In main storage device 92, the program stored in auxiliary storage device 93 or the like is expanded by processor 91. Main storage device 92 is realized by a volatile memory such as DRAM (Dynamic Random Access Memory), for example. Further, as main storage device 92, a non-volatile memory such as MRAM (Magnetoresistive Random Access Memory) may be configured / added.

[0191] Auxiliary storage device 93 stores various data such as programs. Auxiliary storage device 93 is realized by a local disk such as a hard disk or a flash memory. Note that it is also possible to adopt a configuration in which various data is stored in main storage device 92 and auxiliary storage device 93 is omitted.

[0192] Input / output interface 95 is an interface for connecting information processing apparatus 90 and peripheral devices based on standards and specifications. Communication interface 96 is an interface for connecting to an external system or device through a network such as the Internet or an intranet based on standards and specifications. Input / output interface 95 and communication interface 96 may be shared as an interface for connecting to external devices.

[0193] The information processing apparatus 90 may be connected with input devices such as a keyboard, a mouse, and a touch panel as necessary. These input devices are used for inputting information and settings. When using a touch panel as an input device, the display screen of the display device may also serve as the interface of the input device. The data communication between the processor 91 and the input device may be mediated by the input / output interface 95.

[0194] Further, the information processing apparatus 90 may be equipped with a display device for displaying information. When equipped with a display device, it is preferable that the information processing apparatus 90 be provided with a display control device (not shown) for controlling the display of the display device. The display device may be connected to the information processing apparatus 90 via the input / output interface 95.

[0195] Further, the information processing apparatus 90 may be equipped with a drive device. The drive device mediates the reading of data and programs from the recording medium and the writing of the processing results of the information processing apparatus 90 to the recording medium between the processor 91 and the recording medium (program recording medium). The drive device may be connected to the information processing apparatus 90 via the input / output interface 95.

[0196] The above is an example of the hardware configuration for enabling the processing according to each embodiment of the present invention. Note that the hardware configuration in FIG. 36 is an example of the hardware configuration for executing the processing according to each embodiment, and does not limit the scope of the present invention. Also, a program for causing a computer to execute the processing according to each embodiment is included in the scope of the present invention. Furthermore, a program recording medium on which the program according to each embodiment is recorded is included in the scope of the present invention. The recording medium can be realized by, for example, an optical recording medium such as a CD (Compact Disc) or a DVD (Digital Versatile Disc). The recording medium may be realized by a semiconductor recording medium such as a USB (Universal Serial Bus) memory or an SD (Secure Digital) card. Also, the recording medium may be realized by a magnetic recording medium such as a flexible disk or other recording media. When the program executed by the processor is recorded on the recording medium, the recording medium corresponds to a program recording medium.

[0197] The components of each embodiment may be arbitrarily combined. Also, the components of each embodiment may be realized by software or by a circuit.

[0198] The present invention has been described above with reference to the embodiments, but the present invention is not limited to the above embodiments. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

Explanation of Reference Numerals

[0199] 5 Learning system 6 Physical state estimation system 10, 20, 30, 40 Feature selection device 11, 21, 31, 41 Acquisition unit 12, 22, 32, 42 Construction unit 13, 23, 33, 43 Analysis unit 15, 25, 35, 45 Statistics unit 17, 27, 37, 47 Selection unit Output units 19, 29, 39, 49 Estimation model construction unit 26 Gait measurement devices 50, 60 Learning device 55 Sensor 61 Feature data generation unit 62 Estimation device 63 Receiving unit 551 Learning unit 553 Memory unit 555 Data acquisition unit 631 Memory unit 632 Estimation unit 633 Output unit 635

Claims

1. An acquisition means for acquiring a plurality of data sets; A construction means for changing the distribution of data included in the data set to construct a plurality of re-extracted data sets; An analysis means for analyzing a plurality of the re-extracted data sets using a Lasso regression method; A statistical means for aggregating the values of elements included in a plurality of the re-extracted data sets according to the analysis results of the plurality of the re-extracted data sets, and setting a logical value for the elements included in the plurality of the re-extracted data sets according to the aggregation result of the values of the elements; A selection means for selecting a combination of feature quantities corresponding to the value of the logical value set for the element according to a preset specific rule; An output means for outputting selection information regarding the selected combination of feature quantities, comprising: The analysis means: Executes the Lasso regression for each of a plurality of preset regularization parameters with respect to a plurality of the re-extracted data sets; Generates a first matrix of a plurality of patterns constituted by a column corresponding to the regularization parameter used in the Lasso regression and a row corresponding to the feature quantity; The statistical means: Executes a first statistical process of setting the first logical value of a cell of a non-zero element to 1 and setting the first logical value of a cell of a zero element to 0 with respect to the first matrix of a plurality of patterns; Executes a second statistical process of aggregating the first logical value for each cell constituting the first matrix of a plurality of patterns, setting 1 as the second logical value for a cell where the aggregation value of the first logical value satisfies a predetermined condition, and setting 0 as the second logical value for a cell where the aggregation value of the first logical value does not satisfy the predetermined condition, to generate a second matrix; The selection means: A feature quantity selection device that selects columns of the second matrix according to the preset specific rule and selects a combination of the feature quantities corresponding to the selected columns.

2. The construction means: The feature quantity selection device according to claim 1, which constructs a plurality of the re-extracted data sets using a Leave-One-Subject-Out method.

3. The construction means: The feature quantity selection device according to claim 1, which constructs a plurality of the re-extracted data sets using a bootstrap method.

4. The statistical means: In the second statistical process: Calculates the total value of the first logical value for each cell constituting the first matrix of a plurality of patterns. A feature quantity selection device according to any one of claims 1 to 3, which generates the second matrix in which the second logical value of a cell whose total value of the first logical values is equal to or greater than a predetermined threshold is set to 1, and the second logical value of a cell whose total value of the first logical values is less than the predetermined threshold is set to 0.

5. The statistical means In the second statistical process, calculates the average value of the first logical values for each cell constituting the first matrix of a plurality of patterns, A feature quantity selection device according to any one of claims 1 to 3, which generates the second matrix in which the second logical value of a cell whose average value of the first logical values is equal to or greater than a predetermined threshold is set to 1, and the second logical value of a cell whose average value of the first logical values is less than the predetermined threshold is set to 0.

6. acquisition means for acquiring a plurality of data sets; construction means for changing the distribution of data included in the data set to construct a plurality of re-extracted data sets; analysis means for analyzing a plurality of the re-extracted data sets using a Lasso regression method; statistical means for aggregating the values of elements included in the plurality of re-extracted data sets according to the analysis results of the plurality of re-extracted data sets, and setting logical values for the elements included in the plurality of re-extracted data sets according to the aggregation result of the values of the elements; selection means for selecting a combination of feature quantities according to the value of the logical value set for the element according to a preset specific rule; output means for outputting selection information regarding the selected combination of feature quantities, and The analysis means executes the Lasso regression for each of a plurality of preset regularization parameters for each of the plurality of re-extracted data sets, generates a first matrix of a plurality of patterns constituted by a column corresponding to the regularization parameter used in the Lasso regression and a row corresponding to the feature quantity, The statistical means executes a second statistical process of aggregating the values of the elements for each cell constituting the first matrix of a plurality of patterns, and generating a second matrix in which 1 is set as the second logical value for a cell whose average value of the element values is equal to or greater than a predetermined threshold, and 0 is set as the second logical value for a cell whose average value of the element values is less than the predetermined threshold, The selection means is a feature quantity selection device that selects a column of the second matrix according to the preset specific rule and selects a combination of feature quantities corresponding to the selected column.

7. estimating model construction means for constructing an estimating model by learning using the selected feature quantity and evaluating the constructed estimating model. The selection means The feature quantity selection device according to any one of claims 1 to 6, which selects a combination of the feature quantities according to an evaluation result of the estimation model.

8. A computer acquires a plurality of data sets, changes a distribution of data included in the data set to construct a plurality of re-extracted data sets, analyzes the plurality of re-extracted data sets using a Lasso regression method, aggregates values of elements included in the plurality of re-extracted data sets according to analysis results of the plurality of re-extracted data sets, sets logical values for the elements included in the plurality of re-extracted data sets according to an aggregation result of the values of the elements, selects a combination of feature quantities according to the logical values set for the elements according to a preset specific rule, outputs selection information regarding the selected combination of the feature quantities, In the analysis, for the plurality of re-extracted data sets, the Lasso regression is executed for each of a plurality of preset regularization parameters, generates a first matrix of a plurality of patterns constituted by a column corresponding to the regularization parameter used in the Lasso regression and a row corresponding to the feature quantity, In the setting, for the first matrix of the plurality of patterns, executes a first statistical process of setting a first logical value of a cell of a non-zero element to 1 and setting a first logical value of a cell of a zero element to 0, for each cell constituting the first matrix of the plurality of patterns, aggregates the first logical value, and generates a second matrix in which 1 is set as a second logical value for a cell where an aggregation value of the first logical value satisfies a predetermined condition, and 0 is set as the second logical value for a cell where the aggregation value of the first logical value does not satisfy the predetermined condition by executing a second statistical process, In the selection, A feature quantity selection method that selects a column of the second matrix according to the preset specific rule and selects a combination of the feature quantities corresponding to the selected column.

9. A computer is caused to perform a process of acquiring a plurality of data sets, perform a process of changing a distribution of data included in the data set to construct a plurality of re-extracted data sets, perform a process of analyzing the plurality of re-extracted data sets using a Lasso regression method, perform a process of aggregating values of elements included in the plurality of re-extracted data sets according to analysis results of the plurality of re-extracted data sets, perform a process of setting logical values for the elements included in the plurality of re-extracted data sets according to an aggregation result of the values of the elements, processing for selecting a feature quantity of a combination according to the logical value set for the element in accordance with a preset specific rule; processing for outputting selection information regarding the selected combination of feature quantities; in the processing for analyzing; processing for executing the Lasso regression for each of a plurality of preset regularization parameters with respect to the plurality of re-extracted data sets; processing for generating a first matrix of a plurality of patterns constituted by a column corresponding to the regularization parameter used in the Lasso regression and a row corresponding to the feature quantity; in the processing for setting; processing for executing a first statistical process of setting the first logical value of a cell of a non-zero element to 1 and setting the first logical value of a cell of a zero element to 0 with respect to the first matrix of a plurality of patterns; processing for executing a second statistical process of generating a second matrix in which, for each cell constituting the first matrix of a plurality of patterns, the first logical values are aggregated, 1 is set as a second logical value for a cell in which the aggregated value of the first logical values satisfies a predetermined condition, and 0 is set as the second logical value for a cell in which the aggregated value of the first logical values does not satisfy the predetermined condition; in the processing for selecting; a program for causing execution of processing for selecting columns of the second matrix according to the preset specific rule and selecting a combination of the feature quantities corresponding to the selected columns.

Images