JP2014041426A - Time series data processing system and time series data processing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To facilitate learning in the estimation of time series data, and to properly perform data estimation with small processing amounts in estimation.SOLUTION: A time series data processing system 1 comprises: an estimation function generation section 13 for using the set of the value of first data for learning, the value of second data for learning and correct answer data relating to the fluctuation of the value of the second data for learning as data for statistical learning, and for generating an estimation function for estimating the fluctuation of a value of a second classification by performing statistical learning in each measurement timing; an input value fluctuation estimation section 17 for estimating the input value fluctuation of the value of the second classification from an input value of a first classification and the input value of the second classification and the estimation function; and a derivation section 18 for deriving the value of the second classification from the input value fluctuation and the input value of the second classification.

Description

  The present invention relates to a time-series data processing system and a time-series data processing method.

  As a method of estimating the value of another time series data different from the one time series data from the measured one time series data value, a hidden condition that defines the state associated with each value of the one time series data is defined. There is known a method of preparing a plurality of state models based on the Markov model and estimating values of other time series data from the state model.

  Specifically, as a method of estimating time series data using a state model, acceleration (one time series data) is measured, and the number of steps per unit time is linked beforehand from the number of steps per unit time calculated from the acceleration. A method for selecting one of five attached action scenes (state models) and calculating a momentum (activity amount) (other time series data) based on an arithmetic expression and acceleration determined by the selected action scene Is known (see, for example, Patent Document 1). In such an estimation method, an experimentally determined amount of momentum is applied from state estimation based on transitions between a plurality of state models such as stop / walking / walking / running based on acceleration sensor values.

JP 2008-246181 A

  However, the conventional method estimates the momentum (other time series data) based on the arithmetic expression and acceleration (one time series data) determined for each state model. If an exercise that does not fit is performed, an appropriate state model cannot be selected from the acceleration (one time series data), and as a result, the momentum (other time series data) should be estimated appropriately. I could not.

  In addition, when a large number of state models are prepared in order to reduce the chances of exercise not conforming to the state model, acceleration (one time series data) and momentum (other time series data) for each state model There is a problem that it takes a lot of time to acquire learning data and to learn learning data. Then, as the number of state models increases, the amount of processing at the time of estimation increases, so there is a problem in terms of the amount of processing at the time of estimation.

  The present invention has been made in view of the above circumstances, and facilitates learning in time series data estimation, and at the time of estimation, can perform data estimation appropriately with a small amount of processing, and a time series data processing system and time series An object is to provide a data processing method.

  The time-series data processing system according to the present invention is data obtained by measuring values of a first type and a second type, which are different types, and the measurement timing of the values includes the first type and the second type. The learning data input means for inputting the first learning data and the second learning data, which are time series data associated with each other, and each value of the second learning data are The variation toward the timing value at which the value is measured next to the measured timing is defined as correct answer data regarding the variation in each value of the second learning data, and the first learning data The value of the first type is obtained by performing statistical learning using a set of the value, the value of the second learning data, and the correct answer data regarding the fluctuation of the value of the second learning data as statistical learning data. And the value of the second type, the second type An estimation function generating means for generating an estimation function for estimating a change in value; an input means for receiving an input value of the first type and an input value of the second type; an input value of the first type; The input value fluctuation for estimating the fluctuation of the input value, which is the fluctuation toward the timing next to the timing related to the input value of the second classification of the value of the second classification, from the input value of the classification of Deriving means for deriving a value of the second type at a timing next to the timing related to the input value of the second type from the estimation means, the input value variation, and the input value of the second type; Output means for outputting the derived value of the second type.

  In this time-series data processing system, in the learning stage, by learning the fluctuations of the values of the first type and the second type and the values of the second learning data, An estimation function for estimating the fluctuation of the value of the two types is generated. Since the state model is not taken into consideration, acquisition of learning data and learning of learning data can be easily performed. Further, in the estimation stage, the fluctuation of the second type at the next timing can be estimated from the input value of the first type, the input value of the second type, and the estimation function. Since the state model is not used, it is possible to appropriately estimate the second time series data regardless of conformity / nonconformity with the state model. Further, the processing amount can be reduced by not using the state model. In addition, the value of the second type is not estimated directly, but the fluctuation is estimated, but the value of the second type at each time point is used to estimate the value of the second type at the next time point. By using it, it is possible to appropriately estimate not only the fluctuation but also the specific value. Then, by using the value of the second type when estimating the value of the second type at the next timing, appropriate data estimation with high accuracy is realized.

  As described above, according to the present invention, learning can be facilitated, and at the time of estimation, data can be estimated appropriately with a small amount of processing.

  In the time-series data processing system according to the present invention, the estimation function generating means multiplies a value of the first learning data and a value of the second learning data by a predetermined coefficient at each measurement timing. It is preferable that a set of the combined value and the correct answer data regarding the change in the value of the second learning data according to the predetermined coefficient is further set as statistical learning data.

  By preparing a plurality of sets of statistical learning data from data related to one measurement timing, it is possible to efficiently increase the amount of data for statistical learning and improve estimation accuracy.

  In the time-series data processing system according to the present invention, it is preferable that the input unit accepts the second type value output by the output unit as the second type input value at the next timing.

  By using the value of the second type output by the output means as the input value of the next timing, it is possible to appropriately estimate the value of the second type that is time-series data.

  In the time-series data processing system according to the present invention, the fluctuation is preferably either ascending or descending.

  By making the fluctuation into two-class classification of ascending / descending, appropriate data estimation can be performed with a smaller amount of processing, and learning of correct data can be further simplified.

  The time-series data processing system according to the present invention may have only a learning function, and the time-series data processing system having only a learning function has a first type and a second type that are different types. First learning data and second learning data, which are data obtained by measuring the value of the type, and the measurement timing of the value corresponds to the time series data corresponding to the first type and the second type And a learning data input means for inputting each of the second learning data, and a variation of each value of the second learning data toward a timing value at which the value is measured next to the timing at which each value is measured, Correct data regarding fluctuations in each value of the learning data, and correct data regarding fluctuations in the values of the first learning data, the second learning data, and the second learning data at each measurement timing , A set An estimation function generating means for generating an estimation function for estimating a variation in the value of the second type from the value of the first type and the value of the second type by performing statistical learning as data for statistical learning; .

  Since this time series data processing system does not consider the state model, acquisition of learning data and learning of learning data can be easily performed.

  In addition, the time series data processing system according to the present invention may be provided with only the estimation function, and the time series data processing system provided with only the estimation function is different from each other in the first type and the second type. From the input means for receiving the input value of the type, the input value of the first type, the input value of the second type, and the predetermined estimation function, the input value of the second type of the value of the second type From the input value fluctuation estimating means for estimating the input value fluctuation that is a fluctuation toward the next timing of the timing related to the above, the input value fluctuation, and the second type input value to the second type input value Deriving means for deriving a value of the second type at a timing next to such timing, and output means for outputting the value of the second type derived by the deriving means.

  Since this time-series data processing system does not use a state model, the second time-series data can be estimated appropriately regardless of whether the state model is compatible or not. Further, the processing amount can be reduced by not using the state model. In addition, the value of the second type is not estimated directly, but the fluctuation is estimated, but the value of the second type at each time point is used to estimate the value of the second type at the next time point. By using it, it is possible to appropriately estimate not only the fluctuation but also the specific value. Then, by using the value of the second type when estimating the value of the second type at the next timing, appropriate data estimation with high accuracy is realized.

  The time-series data processing method according to the present invention is data obtained by measuring values of a first type and a second type, which are different types, and the measurement timing of the values includes the first type and the second type. The learning data input step for inputting the first learning data and the second learning data, which are time series data associated with each other, and the respective values of the second learning data are The variation toward the timing value at which the value is measured next to the measured timing is defined as correct answer data regarding the variation in each value of the second learning data, and the first learning data The value of the first type is obtained by performing statistical learning using a set of the value, the value of the second learning data, and the correct answer data regarding the fluctuation of the value of the second learning data as statistical learning data. And the value of the second type, the second type An estimation function generating step for generating an estimation function for estimating a variation in value; an input step for receiving an input value of the first type and an input value of the second type; an input value of the first type; The input value fluctuation for estimating the fluctuation of the input value, which is the fluctuation toward the timing next to the timing related to the input value of the second classification of the value of the second classification, from the input value of the classification of In the deriving step for deriving the value of the second type at the timing next to the timing related to the input value of the second type from the estimation step, the input value variation, and the input value of the second type, An output step of outputting the derived value of the second type. This time-series data processing method can achieve the same effects as the above-described time-series data processing system.

  The time-series data processing method according to the present invention is data obtained by measuring values of a first type and a second type, which are different types, and the measurement timing of the values includes the first type and the second type. The learning data input step for inputting the first learning data and the second learning data, which are time series data associated with each other, and the respective values of the second learning data are The variation toward the timing value at which the value is measured next to the measured timing is defined as correct answer data regarding the variation in each value of the second learning data, and the first learning data The value of the first type is obtained by performing statistical learning using a set of the value, the value of the second learning data, and the correct answer data regarding the fluctuation of the value of the second learning data as statistical learning data. And the value of the second type, the second type Including an estimation function generating step of generating an estimation function for estimating the variation of the value, a. This time-series data processing method can achieve the same effects as the above-described time-series data processing system.

  The time-series data processing method according to the present invention includes an input step for receiving input values of a first type and a second type, which are different types, an input value of the first type, and an input value of the second type. And an input value fluctuation estimation step for estimating an input value fluctuation, which is a fluctuation of the second type value toward the timing next to the timing related to the second type input value, from a predetermined estimation function; A derivation step for deriving a value of the second type at a timing next to the timing of the input value of the second type from the input value variation and the input value of the second type; And an output step for outputting a value of two types. This time-series data processing method can achieve the same effects as the above-described time-series data processing system.

  According to the present invention, it is possible to provide a time-series data processing system and a time-series data processing method capable of facilitating learning in estimating time-series data and appropriately performing data estimation with a small amount of processing during estimation. it can.

It is a figure which shows the function structure of the time series data processing system which concerns on embodiment of this invention. It is a figure which shows the acceleration and heart rate which were measured by the data measurement part for learning. It is a figure which shows the structure of the data measurement part for learning which concerns on embodiment of this invention. It is a figure which shows the hardware constitutions of the time series data processing server which concerns on embodiment of this invention. It is a figure which shows the learning flow of the time series data processing system which concerns on embodiment of this invention. It is a figure which shows the estimation flow of the time series data processing system which concerns on embodiment of this invention. It is the figure which showed the relationship between the estimated value of the heart rate using the estimation function at the time of performing statistical learning without using additional learning data, and the measured value of the heart rate. It is the figure which showed the relationship between the estimated value of the heart rate using the estimation function at the time of performing statistical learning using the data for additional learning, and the measured value of the heart rate. It is a modification of the structure of the data measurement part for learning. It is a modification of the structure of the data measurement part for learning. It is a modification of the structure of the data measurement part for learning.

  Hereinafter, preferred embodiments of a time-series data processing system and a time-series data processing method according to the present invention will be described in detail with reference to the drawings. In the description of the drawings, the same elements are denoted by the same reference numerals, and redundant description is omitted.

  The time-series data processing system according to the present invention estimates the value of the other time-series data by measuring one time-series data of two correlated time-series data. For example, if you want to measure the amount of exercise (for example, heart rate) for the purpose of health management, the person being measured wears a heart rate monitor (exercise meter) on the chest for measuring heart rate, etc. Is required, and the measurement is troublesome and burdensome. Also, heart rate monitors are generally expensive. The time-series data processing system according to the present invention does not measure time-series data (heart rate in the above example), which is a burden and a burden on such measurement, but is easily correlated with the time-series data. By measuring time series data that can be measured at low cost, the value of the time series data to be measured is estimated. In the case of measuring the heart rate described above, the acceleration that is correlated with the heart rate and easy to measure is measured (measured by an acceleration sensor (sensor device)), and the value of the heart rate is estimated from the measured acceleration. To do. And such estimation is implement | achieved by learning beforehand the data which measured the 2 time series data with correlation as learning data.

  First, the function of the time-series data processing system 1 according to the present embodiment will be described with reference to FIG. FIG. 1 is a diagram showing a functional configuration of a time-series data processing system 1 according to the embodiment of the present invention. As shown in FIG. 1, the time-series data processing system 1 according to the present embodiment has a configuration of a time-series data processing server 2, a learning data input unit 11, an estimation function generation unit 13, an input unit 14, an initial value. The input unit 16, the input value variation estimation unit 17, the derivation unit 18, the output unit 19, the learning data measurement unit 12, and the acceleration sensor 15 are configured.

  The time-series data processing server 2 is a server operated by a business operator, for example. In addition, the learning data measurement unit 12 is worn by a person to be measured for learning data in order to measure learning data to be described later. The acceleration sensor 15 is worn by a person to be measured in order to measure acceleration used for heart rate estimation described later. The person wearing the learning data measurement unit 12 and the person wearing the acceleration sensor 15 may be the same or different. The time-series data processing server 2, the learning data measurement unit 12, and the acceleration sensor 15 can transmit and receive information to and from each other via, for example, a mobile communication network.

  The time-series data processing system 1 is a system that estimates the value of another time-series data different from the one time-series data from the measured value of one time-series data. In the present embodiment, the measured time series data will be described as acceleration, and other time series data to be estimated different from the one time series data will be described as heart rate. That is, by measuring the acceleration, the value of the heart rate that is a measure of the amount of exercise is estimated. Such a system for estimating the heart rate from the acceleration is used for estimating the amount of exercise (heart rate) from the sensor value (acceleration) in health care.

  In addition, the time series data processing system 1 uses these time series data in advance in order to estimate the value of the heart rate, which is another time series data, from the measured acceleration value, which is one time series data. The measured data is used as learning data, and statistical learning is performed. As described above, the functional configuration of the time-series data processing system 1 is large and includes a learning function and an estimation function.

  Among the functional configurations of the time series data processing system 1 described above, the learning data input unit 11, the learning data measurement unit 12, and the estimation function generation unit 13 are functional configurations relating to the learning function of the time series data processing system 1. It is. Note that a system that does not have an estimation function and includes only a functional configuration related to a learning function can be a time-series data processing system.

  Among the functional configurations of the time-series data processing system 1 described above, the input unit 14, the acceleration sensor 15, the initial value input unit 16, the input value variation estimation unit 17, the derivation unit 18, and the output unit 19 These are functional configurations related to the estimation function of the time-series data processing system 1. A system that does not have a learning function but has only a functional configuration related to the estimation function can be a time-series data processing system.

  The learning data input unit 11 is data obtained by measuring values of a first type and a second type, which are different types, and the measurement timing of the values is associated with the first type and the second type. Learning data input means for inputting first learning data and second learning data, which are time series data. Here, in the present embodiment, the first type is acceleration, and the second type is heart rate. Therefore, the first learning data is time-series data relating to acceleration, and the second learning data is time-series data relating to heart rate. Specifically, the first learning data is data in which an acceleration value is associated with a time indicating a measurement timing of the value. Specifically, the second learning data is data in which the value of the heart rate is associated with the time indicating the measurement timing of the value. Alternatively, the acceleration value and the heart rate value corresponding to each other may be sequentially input to the learning data input unit 11 (each time the value is measured). In that case, the time indicating the measurement timing may not necessarily correspond.

  An image of time series data relating to the measured acceleration and heart rate will be described with reference to FIG. FIG. 2 is a diagram showing acceleration and heart rate measured by the learning data measurement unit 12. In FIG. 2, the left axis shows the heart rate value, the right axis shows the acceleration value, and the horizontal axis shows the time. The measured values of heart rate and acceleration are measured at the corresponding timing (time). As shown in FIG. 2, the rise and fall of the heart rate and the acceleration norm conversion value are almost the same, and it can be confirmed that there is a correlation between the heart rate and the acceleration.

The acceleration norm is the square root of the sum of the squares of vector elements each having the acceleration value of each axis when the acceleration is measured by a triaxial acceleration sensor or the like. The acceleration norm conversion value is a value obtained by performing a predetermined conversion based on the acceleration norm value. The acceleration norm conversion value n _t at time t is 99.98% of the acceleration norm conversion value n _t-1 at time t−1, which is the timing at which acceleration was measured before time t, and the acceleration norm measurement at time t. It is calculated by adding the value n ′ _t at a ratio of 0.02%. An equation for calculating the acceleration norm conversion value n _t at time t is shown below. Incidentally, w is a weight multiplied to the measured value _n't acceleration norm in order to give acceleration norm conversion value n _t, as described above, in FIG. 2, is set to w = 0.0002.

  The time-series data related to acceleration and the time-series data related to heart rate input by the learning data input unit 11 correspond to the measurement timings of the values. Here, “corresponding” does not necessarily mean that the acceleration and the heart rate are measured at the same timing, but at the time of estimation, from the input value of acceleration, the heart rate corresponding to the input value of acceleration The measurement timings of acceleration and heart rate need to be approximated to such an extent that fluctuations in the above can be estimated and the heart rate can be derived. The time-series data related to acceleration and the time-series data related to heart rate input by the learning data input unit 11 are measured by the learning data measuring unit 12 described later and transmitted to the time-series data processing server 2. is there.

  The learning data measurement unit 12 measures time-series data related to acceleration and time-series data related to heart rate, and transmits them to the time-series data processing server 2. Data transmission from the learning data measurement unit 12 to the time-series data processing server 2 is performed by wireless communication such as mobile communication, for example.

  The configuration of the learning data measurement unit 12 is shown in FIG. FIG. 3 is a diagram showing a configuration of the learning data measurement unit 12 according to the embodiment of the present invention. The learning data measurement unit 12 includes a heart rate monitor 21, a portable terminal 22, and a wristwatch type terminal 23, and the acceleration and heart rate are measured by wearing each of them by a person to be measured. The

The heart rate monitor 21 is worn by a person and measures the heart rate. Examples of the heart rate monitor 21 include a band type attached to the chest. The heart rate monitor 21 transmits the heart rate to the portable terminal 22 by wireless communication or the like. Here, for example, ANT ⁺ can be considered as the wireless communication.

  The portable terminal 22 has a function of receiving the heart rate of the person wearing the heart rate monitor 21 measured by the heart rate monitor 21 and transmitting the heart rate to the watch type terminal 23 by communicating with the watch type terminal 23. Is. As the mobile terminal 22, for example, a mobile phone can be considered. Communication between the portable terminal 22 and the wristwatch-type terminal 23 is performed by wireless communication or the like. As the wireless communication, short-range wireless communication, for example, Bluetooth (registered trademark) or the like can be considered. When communication is performed using Bluetooth (registered trademark), the portable terminal 22 functions as a server and the wristwatch-type terminal 23 functions as a client.

  The wristwatch-type terminal 23 measures acceleration by the acceleration sensor 24 and stores the acceleration measured by the acceleration sensor 24 and the heart rate received from the portable terminal 22 in the storage 25. As the acceleration sensor 24, a triaxial acceleration sensor or the like is used. The acceleration and heart rate stored in the storage 25 are transmitted to the learning data input unit 11 by wireless communication such as mobile communication.

  In this way, the heart rate measured by the heart rate monitor 21 and the acceleration measured by the acceleration sensor 24 of the wristwatch type terminal 23 are periodically measured at a plurality of timings at corresponding timings. The timing that corresponds is not necessarily the same, but at the time of estimation, the fluctuation of the heart rate corresponding to the input value of acceleration is estimated from the input value of acceleration, and the heart rate can be derived. The measurement timing of acceleration and heart rate must be approximated. Further, both the acceleration and the heart rate need to be measured at a plurality of times because the learning data needs to be time-series data that is numerical data arranged in time order.

  Specifically, in order to measure a plurality of timings at timings corresponding to acceleration and heart rate, it is conceivable that the sampling rate is the same for acceleration and heart rate. In this case, for example, it is conceivable that both the acceleration sensor 24 and the heart rate monitor 21 are provided with timers and the measurement timing is synchronized to measure at a predetermined time.

  The estimation function generation unit 13 measures each value of the time series data regarding the heart rate, which is the second learning data, input by the learning data input unit 11 and the timing after each value is measured. Compared to the measured timing value to determine how each value fluctuates toward the next measured timing value, and the variation is calculated for each time-series data related to heart rate. The correct data regarding the fluctuation of the value is used.

  Fluctuation means that the value changes. In addition, the fluctuation toward the value of the timing at which the value is measured next to the timing at which the value is measured means how the value changes toward the timing at which the value is measured next (for example, an increase , Descent, etc.) and the difference in value from the timing at which the value was measured next. In the present embodiment, the fluctuation means how each value changes toward the timing when the value is measured next, and is either rising or falling. Therefore, heart rate fluctuation is simply a two-class classification of ascending and descending.

  Thus, for example, when the heart rate at timing 1 which is a predetermined measurement timing is 70 and the heart rate at timing 2 which is the next measurement timing after timing 1 is 75, time-series data regarding the heart rate The correct answer data regarding the change in the value of the heart rate at the timing 1 is “increased”. Similarly, if the heart rate at timing 2 which is a predetermined measurement timing is 75 and the heart rate at timing 3 which is the next measurement timing after timing 2 is 70, the time series regarding the heart rate The correct answer data regarding the fluctuation of the value of the heart rate at the data timing 2 is “down”.

  Note that filtering (smoothing) may be performed on the time-series data related to the measured heart rate before the rise or fall is determined from the time-series data related to the heart rate as the correct answer data related to the fluctuation. For example, in order to remove a noise component of data, a filtering process is performed by appropriately weighting and adding the values acquired at a certain timing to the values acquired at a plurality of measurement timings before and after the value. It is conceivable that the value after the filtering process is set to a value acquired at a certain timing. Note that a conventional smoothing method for time series data can be applied to the filtering process.

  The estimation function generation unit 13 also sets time-series data values relating to acceleration, time-series data values relating to heart rate, and time-series relating to heart rate at each timing when the learning data is measured. A set of correct data related to fluctuations in data values is used as statistical learning data, and by performing statistical learning, the time-series data value related to acceleration and the time-series data value related to heart rate are used to calculate the heart rate. It is an estimation function generation means for generating an estimation function for estimating a variation in value. Note that statistical learning is performed by a random forest or the like that can be classified into two classes. The estimation function plays a role as an ascending / descending classifier in order to estimate fluctuations in the value of the heart rate at the time of estimation, that is, ascending / descending of the heart rate.

  That is, the estimation function generation unit 13 uses the acceleration value and the heart rate value measured at the corresponding timing as arguments corresponding to the correct data of the estimation function, and the heart rate as the correct data (output of the estimation function). The correct answer data (rising or descending) regarding the fluctuation of the value is taken as one set, and statistical learning is performed for all measured timings. As a result of the statistical learning, when acceleration and a heart rate are input, an estimation function capable of estimating a fluctuation (increase or decrease) in the value of the heart rate is generated.

The estimation function f is defined by the following equation, and f = 1 when the correct answer data is rising, and f = −1 when the correct data is falling. Note that x represents an acceleration value at a predetermined timing, and r represents a heart rate value (usually at the same timing) corresponding to the timing related to the acceleration value. Further, sin (Δr) is a variable indicating an increase or decrease in heart rate.

  In addition, the estimation function generation unit 13 relates to a value of time series data related to acceleration, a value obtained by multiplying a value of time series data related to heart rate by a predetermined coefficient, and a heart rate corresponding to the predetermined coefficient for each measurement timing. A set of correct data relating to fluctuations in the value of the time-series data is further set as statistical learning data as additional learning data.

  As the predetermined coefficient, a value larger than 0 and smaller than 1 (for example, 0.9), a value larger than 1 (for example, 1.1), and the like are conceivable. The correct answer data corresponding to the predetermined coefficient may be increased when a value greater than 0 and smaller than 1 is set as the predetermined coefficient, and decreased when a value greater than 1 is set as the predetermined coefficient. It is done.

For example, when the predetermined coefficient is 0.9, the correct answer data that is a learning parameter is determined to rise. The correct data for the rise is defined by the following formula, and f = 1. Note that x represents an acceleration value at a predetermined timing, and r represents a heart rate value (usually at the same timing) corresponding to the timing related to the acceleration value. Sin _up (Δr) is a variable indicating that the heart rate increases.

For example, when the predetermined coefficient is 1.1, correct data that is a learning parameter is determined to fall. The descending correct answer data is defined by the following formula, and f = -1. Note that x represents an acceleration value at a predetermined timing, and r represents a heart rate value (usually at the same timing) corresponding to the timing related to the acceleration value. Further, sin _down (Δr) is a variable indicating that the heart rate decreases.

In addition, the estimation function generation unit 13 sets the acceleration norm of the acceleration measured by the three-axis acceleration sensor when the time-series data value related to the acceleration input by the learning data input unit 11 is used as the statistical learning data. calculate. Then, processing for converting the data into frequency axis data by performing Fourier analysis (FFT analysis), filtering processing, trend removal, and the like may be performed. These processes may all be performed continuously, or may be performed independently of each other. Conventional methods can be used for the Fourier analysis, filtering processing, and trend removal. Note that the above-described processing such as Fourier analysis is a data processing process for creating data (parameters) having various characteristics from one measured acceleration. As an example of such a data processing process, an acceleration norm conversion value n _t (acceleration norm conversion value n _t-1 at time t-1 which is a timing at which acceleration is measured before time t) is used as an acceleration norm at time t. it is considered to use a calculation) by summing the measured values _n't of the acceleration norm at time t at a predetermined ratio. Acceleration norm conversion value n _t is by changing the adding tailor ratio measurements _n't of the acceleration norm in the acceleration norm conversion value n _t-1 and time t at time t-1, to create various data (parameters) be able to. As described above, a plurality of data (parameters) can be created from the measured acceleration, but the measured acceleration may be simply used as the statistical learning data regarding acceleration. Any one of a plurality of data (parameters) created from one acceleration may be employed, or all of the plurality of data (parameters) may be employed.

  The input unit 14 is an input unit that receives an acceleration input value that is a first type and a heart rate input value that is a second type. The input value of acceleration received by the input unit 14 is a value obtained by the input unit 14 receiving the acceleration measured by the acceleration sensor 15.

  The acceleration sensor 15 is a device having the same function as the acceleration sensor 24 of the learning data measurement unit 12. The acceleration sensor 15 has a communication function like the wristwatch type terminal 23 and transmits the measured acceleration to the time-series data processing server 2.

  Further, the input value of the heart rate received by the input unit 14 can be a heart rate value output by the output unit 19 described later. That is, the input unit 14 can use the value of the heart rate at the next timing output from the output unit 19 at a certain timing as the input value of the heart rate at the next timing.

  The input unit 14 may input the acceleration value measured by the acceleration sensor 15 for each measured acceleration value, or may input a plurality of acceleration values collectively. The input unit 14 inputs a plurality of acceleration values in consideration of the measurement order of the plurality of accelerations. For the acceleration measured by the acceleration sensor 15, after the acceleration norm is calculated, the data is subjected to Fourier analysis (FFT analysis) and converted to frequency axis data, filtering processing, trend removal, etc. May be.

  The initial value input unit 16 gives an initial value of the input value of the heart rate to the input unit 14. The initial value of the heart rate received from the initial value input unit 16 may be set so that an average heart rate of the user can be set for each user using the time-series data processing system 1, or an average heart rate assuming all users. The number may be defaulted. The initial value input unit 16 may be able to set an initial value by accepting a user input, or may set an initial value stored in advance. The input unit 14 receives the initial value of the heart rate from the initial value input unit 16 as the initial value of the heart rate, and the heart rate output by the output unit 19 described above can be used as the input value of the subsequent heart rate. .

  The input value fluctuation estimation unit 17 converts the input value of the acceleration that is the first type, the input value of the heart rate that is the second type, and the estimation function into the input value of the heart rate that is the heart rate value. It is an input value fluctuation estimating means for estimating an input value fluctuation that is a fluctuation toward the next timing of the timing. Here, the input value of acceleration, which is the first type used by the input value fluctuation estimation unit 17 to estimate the fluctuation, refers to the acceleration input by the input unit 14, and the heart rate, which is the second type. The input value refers to the heart rate input by the input unit 14.

  As described above, the estimation function f can estimate the fluctuation of the heart rate value toward the next timing, that is, the rise or fall of the heart rate value, from the acceleration value and the heart rate value. Is. The input value fluctuation estimation unit 17 uses the estimation function f generated by the estimation function generation unit 13 so that the value of the heart rate toward the next timing from the acceleration and heart rate input values received by the input unit 14. Is estimated to rise or fall (input value fluctuation). The acceleration value and the heart rate value input to the estimation function are values related to timings corresponding to each other as in the learning. In the estimation, only the acceleration value is actually measured, and the heart rate value is not measured. Therefore, the heart rate value corresponding to the acceleration value measured first (input first) is set as the heart rate value input from the initial value input unit 16. As described later, when the heart rate at the next timing is calculated, the heart rate value is set as the heart rate value corresponding to the acceleration value at the next timing.

  As in the present embodiment, the estimation function for estimating the fluctuation of the heart rate at the next timing from the acceleration and the heart rate is usually a value with a relatively high acceleration input value input by the input unit 14. When the input value of the number is a relatively low value, the fluctuation is estimated to be an increase, the input value of the acceleration input by the input unit 14 is a relatively low value, and the input value of the heart rate is a relatively high value In this case, the fluctuation is estimated to be down.

  The deriving unit 18 is deriving means for deriving the value of the heart rate at the timing next to the timing related to the input value of the heart rate from the input value variation and the input value of the heart rate that is the second type.

Deriving heart rate values by deriving unit 18 is defined by the formula shown below, x is the value of the input by the input unit 14 to the nearest acceleration, r _t is the heart rate of the input to the last by the input unit 14 Value, r _{t + 1} is an estimated value of the heart rate at the timing next to the timing related to the input value of the heart rate, and f is an estimation function. Further, ξ is a parameter determined from the sampling rate of acceleration and the maximum value of heart rate change per second, and is a parameter determining fluctuation. For example, assuming that the acceleration sampling rate is 50 Hz and the maximum change in heart rate per second is 5, ξ is set to 0.1 (fixed value). Thus, for example, when it is estimated to rise by the estimation function f, since f = 1 becomes, xi] is 0.1, a value obtained by summing the 0.1 heart rate value r _t is the input heart rate It becomes an estimated value r _{t + 1} of the heart rate at the timing next to the timing related to the value.

  The output unit 19 is an output unit that outputs the value of the heart rate that is the second type derived by the deriving unit 18. The output may be, for example, displayed on a monitor so that it can be confirmed in real time, or output to a storage medium and stored. The heart rate value output by the output unit 19 can be the input value of the heart rate at the input unit 14 at the timing next to the timing at which the input at the input unit 14 relating to the output of the heart rate is received.

  Next, the hardware configuration of the time-series data processing server 2 will be described with reference to FIG. FIG. 4 is a diagram showing a hardware configuration of the time-series data processing server 2 according to the embodiment of the present invention. As shown in FIG. 4, the time-series data processing server 2 includes a CPU 101, a RAM 102 and a ROM 103 that are main storage devices, a communication module 104 for performing communication, and hardware such as an auxiliary storage device 105 such as a hard disk. It is comprised as containing. Each of the above-described functions is exhibited when these components are operated by a program or the like.

  Subsequently, a flow related to learning will be described with reference to the flowchart of FIG. FIG. 5 is a diagram showing a learning flow of the time-series data processing system 1 according to the embodiment of the present invention.

  First, acceleration time-series data is input by the learning data input unit 11 (S101, learning data input step). The time series data of the acceleration is measured by the acceleration sensor 24 of the learning data measuring unit 12.

  Next, the norm of the acceleration is calculated by the estimation function generation unit 13 for the time series data of the acceleration input by the learning data input unit 11 (S102, estimation function generation step). As the acceleration sensor 24, a three-axis acceleration sensor is usually used, and acceleration in a three-dimensional space represented by the X, Y, and Z axes is measured separately. The norm of is calculated.

Next, the estimation function generation unit 13 performs processing for conversion to the frequency axis by Fourier analysis (FFT analysis), filtering processing, trend removal, and the like (S103, estimation function generation step). These processes may all be performed continuously, or may be performed independently of each other. The process is a data processing process for creating data (parameters) having various characteristics. As an example of such a data processing process, an acceleration norm conversion value n _t (time calculated by adding the acceleration norm conversion value n _t-1 at time t-1 which is the timing when acceleration is measured before t and the acceleration norm measurement value n ′ _t at time _t at a predetermined ratio). It is possible.

  The learning data input unit 11 inputs time series data of heart rate (S104, learning data input step). The time-series data of the heart rate is measured by the heart rate monitor 21 of the learning data measurement unit 12.

  The time-series data of the heart rate input by the learning data input unit 11 is subjected to filtering processing by the estimation function generation unit 13 to remove noise components (S105, estimation function generation step).

  Then, for each value of the time-series data related to the filtered heart rate by the estimation function generation unit 13, the fluctuation of each value toward the value of the timing at which the value was measured next to the timing at which each value was measured. (Increase or decrease) is determined as the correct answer data regarding the variation of each value of the time series data regarding the heart rate (S106, estimation function generation step).

  In addition, the estimation function generation unit 13 calculates a value obtained by multiplying each value at each measurement timing of the time series data of the heart rate by 0.9 (S107, estimation function generation step). And the correct data regarding the fluctuation | variation of the value which multiplied each value by 0.9 is made to rise (S108, an estimation function production | generation step).

  Further, the estimation function generation unit 13 calculates a value obtained by multiplying each value at each measurement timing of the time series data of the heart rate by 1.1 (S109, estimation function generation step). And the correct data regarding the fluctuation | variation of the value which multiplied each value by 1.1 is made to fall (S110, an estimation function production | generation step).

  Then, after finishing the processing of S101 to S110, the estimation function generation unit 13 relates to the time-series data value related to acceleration, the time-series data value related to heart rate, and the heart rate at each timing when the learning data is measured. The correct answer data regarding the change in the value of the time series data and the set data are prepared as a data set (S111, estimation function generation step). Here, the value of the time series data relating to the heart rate includes a value obtained by multiplying the predetermined coefficient calculated in S107 and S109. Further, the correct answer data relating to the fluctuation of the value of the time series data relating to the heart rate includes the data that is the correct answer data of rising and falling in S108 and S110. The correct data obtained in S108 is multiplied by the value obtained by multiplying the predetermined coefficient calculated in S107 and the value of time series data related to acceleration, and the correct answer data obtained in S110 is multiplied by the predetermined coefficient calculated in S109. A data set is associated with each value and time-series data value related to acceleration.

  Then, the estimation function generation unit 13 statistically learns the data set prepared in S111 (S112, estimation function generation step), and calculates the heart rate from the value of time series data related to acceleration and the value of time series data related to heart rate. An estimation function for estimating the fluctuation (increase or decrease) of the value is generated (S113, estimation function generation step). The above is the description of the learning flow of the time-series data processing system 1 according to the embodiment of the present invention.

  Subsequently, a flow relating to estimation will be described using the flow of FIG. FIG. 6 is a diagram showing an estimation flow of the time-series data processing system 1 according to the embodiment of the present invention.

  First, an input value of acceleration and an input value of heart rate are received by the input unit 14 (S201, input step). In this case, the input value of acceleration is a value measured by the acceleration sensor 15. The initial value of the heart rate input value is set by the initial value input unit 16. The acceleration measured by the acceleration sensor 15 is similar to S102 and S103. After the acceleration norm is calculated, the data is subjected to Fourier analysis (FFT analysis) and converted to frequency axis data, filtering processing, Trend removal or the like may be performed.

  Next, the input value fluctuation estimating unit 17 changes the heart rate value from the acceleration input value, the heart rate input value, and the estimation function toward the timing next to the timing related to the heart rate input value. An input value fluctuation that is (up or down) is estimated (S202, input value fluctuation estimation step).

  Then, the derivation unit 18 derives the value of the heart rate at the timing next to the timing related to the input value of the heart rate from the fluctuation of the input value as the estimation result and the heart rate (S203, derivation step). .

  Then, the heart rate value derived by the deriving unit 18 is output by the output unit 19 (S204, output step). Subsequently, the processes of S201 to S204 are repeated again (for example, every acceleration measurement timing). At this time, the heart rate value output in S204 is used as the input value of the heart rate at the next timing, and the input is accepted by the input unit 14. The above is the description of the estimation flow of the time-series data processing system 1 according to the embodiment of the present invention.

  Next, operations and effects of the time-series data processing system 1 according to the present embodiment will be described.

  In the time-series data processing system 1 according to the present embodiment, in the learning stage, the estimation function generation unit 13 learns data obtained by measuring acceleration and heart rate values and fluctuations in each value of the heart rate. An estimation function for estimating the fluctuation of the heart rate value at the time of estimation is generated. Since the state model is not taken into consideration, it is possible to reduce the data acquisition cost for acquiring the learning data and facilitate learning of the learning data. In the estimation stage, the input value fluctuation estimation unit 17 can estimate the fluctuation (increase or decrease) of the heart rate at the next timing from the input value of acceleration, the input value of heart rate, and the estimation function. . Since the state model is not used, it is possible to appropriately estimate the heart rate regardless of conformity / nonconformity with the state model.

  Although the heart rate value is not estimated directly, but the fluctuation is estimated, the heart rate value at each time point is used to estimate the heart rate value at the next time point. Not only estimation but also specific values can be estimated appropriately.

  In the time-series data processing system 1 according to this embodiment, the heart rate value can be used when estimating the heart rate value at the next timing. As a result, information on the state of two or more classes is provided in the two-class classification of ascending / descending. Conventionally, the heart rate is estimated from the measured acceleration value and the state model. In the time-series data processing system 1 according to the present embodiment, the next timing is determined from the acceleration and the heart rate. Heart rate can be estimated. That is, accurate data estimation with high accuracy is realized by estimating the heart rate at the next timing using the heart rate value to be estimated.

  Further, in the time-series data processing system 1 according to the present embodiment, the estimation function generation unit 13 uses a value obtained by multiplying the acceleration coefficient and the heart rate value by a predetermined coefficient at each measurement timing, and the predetermined coefficient. It is possible to prepare a plurality of sets of statistical learning data from the data related to one measurement timing by further setting the correct answer data regarding fluctuations in the value of the corresponding heart rate as statistical learning data. It is possible to efficiently increase the amount of data to be subjected to statistical learning and improve the accuracy at the time of estimation. Also, the value obtained by multiplying the measured heart rate value at one timing by a predetermined coefficient is considered to fluctuate toward the measured heart rate value at one timing at the next timing. Accurate correct data can be used as statistical learning data, and the estimation accuracy can be further improved.

  Here, the relationship between the estimated value and the actually measured value when the heart rate is estimated using the estimation function will be described with reference to FIGS. FIG. 7 is a diagram showing a relationship between an estimated value of the heart rate and an actually measured value of the heart rate using an estimation function when statistical learning is performed without using additional learning data. FIG. 8 is a diagram showing the relationship between the estimated value of the heart rate and the actually measured value of the heart rate using the estimation function when statistical learning is performed using the additional learning data. 7 and 8, the vertical axis indicates the heart rate (in FIG. 7, the left axis indicates the actual heart rate and the right axis indicates the estimated heart rate), and the horizontal axis indicates time.

  As shown in FIG. 7, when additional learning data is not provided as statistical learning data, there are few misrecognitions in the two-class classification of ascending and descending (rising / decreasing classification misrecognition rate: 0.0268. ) However, there are places where heart rate values are not estimated correctly, and the scale is not appropriate.

  On the other hand, as shown in FIG. 8, when additional learning data is given as statistical learning data, the classification of ascending and descending is correctly performed (classification error recognition rate of ascending and descending: 0.0159). At the same time, the estimation of the value of the heart rate can be performed with high accuracy.

  In the time-series data processing system 1 according to this embodiment, the input unit 14 receives the heart rate value output from the output unit 19 as an input value of the heart rate at the next timing, thereby continuously accelerating. Is measured, it is possible to appropriately estimate the value of the heart rate that is time-series data.

  Further, in the time series data processing system 1 according to the present embodiment, the fluctuation is made up of two classes of ascending or descending, thereby eliminating the necessity of designing as a multi-class and reducing the amount of processing appropriately. Data estimation is possible. In addition, learning of correct data can be simplified.

  In addition, the time-series data processing system 1 according to the present embodiment can use a flexible estimation model without requiring fine parameter adjustment work by using a conventional statistical classification method at the time of estimation. Become.

  In addition, the time-series data processing system 1 according to the present embodiment learns a large amount of data in various situations, for example, vibration when riding a bus, specific movement when riding a bicycle, etc. It becomes possible to appropriately estimate the heart rate in a detailed state.

  The preferred embodiments of the present invention have been described above. However, the present invention is not limited to the above-described embodiments. The present invention is modified without departing from the scope described in the claims or applied to others. It may be.

  For example, in the present embodiment, an example of acceleration as the first type and heart rate as the second type has been described. However, the present invention is not limited to this, and the first type and the second type are correlated. As long as there is a relationship, the first type and the second type may be any event.

  Further, although the learning data measurement unit 12 has been described with reference to FIG. 3 as measuring the first learning data and the second learning data input by the learning data input unit 11, the learning data The measurement unit 12 is not limited to the configuration shown in FIG. For example, as shown in FIG. 9, the heart rate measured by the heart rate monitor 21 is transmitted to the mobile terminal application 27 of the mobile terminal 26, and the acceleration measured by the acceleration sensor 28 of the mobile terminal 26 is the mobile terminal application. 27, the mobile terminal application 27 can store the transmitted heart rate and acceleration in the storage 29 of the mobile terminal 26. For example, as shown in FIG. 10, the heart rate measured by the heart rate monitor 21 is transmitted to the mobile terminal application 31 of the mobile terminal 30, and the measurement data of the heart rate is the micro device of the micro device 32. The acceleration measured by the acceleration sensor 34 of the micro device 32 and transmitted to the application 33 is transmitted to the micro device application 33, and the acceleration and heart rate data received by the micro device application 33 can be displayed on the display 35. The data can be stored in the storage 36. For example, as shown in FIG. 11, the heart rate measured by the heart rate monitor 21 is transmitted to the portable terminal application 38 of the portable terminal 37, and the acceleration measured by the acceleration sensor 41 of the micro device 40 is portable. The acceleration and heart rate data transmitted to the mobile terminal application 38 of the terminal 37 and received by the mobile terminal application 38 of the mobile terminal 37 can be displayed on the display 42 and stored in the storage 39. Can do. Note that the mobile terminal app and the micro device app may be executed on the same open source operating system.

  Note that the above-described microdevice includes a display and an acceleration sensor, and has a function of wireless communication such as mobile communication, and can be a device that can be worn on an arm like a wristwatch. The mobile terminal application and the micro device application are applications that are executed on the mobile terminal and the micro device, respectively, and cause the mobile terminal and the micro device to function as the above-described learning data measurement unit. Developed on an open platform can be considered.

  Further, the first learning data and the second learning data input by the learning data input unit 11 have been described as being measured by the learning data measurement unit 12, but the present invention is not limited to this. There is no need to be measured data.

  Further, the statistical learning at the time of generation of the estimation function by the estimation function generation unit 13 is performed by a random forest. However, the present invention is not limited to this, and is performed by another statistical learning method such as Adaboost. It may be.

  Further, in the measurement of the acceleration and the heart rate by the learning data measurement unit 12, it has been described that the sampling rate of the acceleration and the heart rate are the same. However, the present invention is not limited to this, and both sampling rates are different. It is good as a thing. For example, if the acceleration sampling rate is 50 Hz and the heart rate sampling rate is 1 Hz, the acceleration is 50 per second and the heart rate is 1 per second. In this case, when the first measurement timing is synchronized, the first acceleration data is the first heart rate data and the 51st acceleration data is the second heart rate data and the measurement timing. It is sufficient to respond. In this case, since there is no measurement value of the heart rate corresponding to the second to 50th acceleration data, there is a problem during statistical learning by the estimation function generation unit 13. When such sampling rates are different, the estimated learning generation unit 13 compares the value of a certain measurement timing with the value of the next measurement timing for time series data with a small sampling rate, so that the sampling rate is large. Corresponding to each value of the time series data, the value of the time series data having a small sampling rate is determined. In the above example, if the value of the first data of the heart rate is 50 and the value of the second data of the heart rate is 100, for example, the second to 50th accelerations. For example, the values of the heart rate corresponding to the data are set to values 51 to 99 by performing interpolation (interpolation) from the values of the data, for example. By doing in this way, even when the sampling rates of acceleration and heart rate are different, statistical learning by the estimation function generation unit 13 can be appropriately performed.

  Further, it is assumed that the time series data of acceleration and heart rate measured by the learning data measuring unit 12 is not measured at the set sampling rate due to a software bug or the like. For example, let us consider a case where data is measured at 1 Hz for a certain period for an acceleration set at a sampling rate of 10 Hz. In this case, the estimation function generator 13 statistically learns the acceleration value, the heart rate value corresponding to the measurement timing of each acceleration, and the fluctuation of the heart rate during the period when the data measurement is performed at 1 Hz. On the other hand, the priority (weight) at the time of learning is lowered and statistical learning is compared with other periods in which data measurement is performed appropriately (periods in which data measurement is performed at a set sampling rate of 10 Hz). I do. By performing statistical learning by lowering the priority of data that is not properly measured and has a wide measurement interval, it is possible to generate an estimation function that enables more accurate estimation. In addition, it is possible to avoid such a situation that data having a wide measurement interval cannot be learned data due to the influence of garbage collection.

  DESCRIPTION OF SYMBOLS 1 ... Time series data processing system, 2 ... Time series data processing server, 11 ... Data input part for learning, 12 ... Data measuring part for learning, 13 ... Estimation function production | generation part, 14 ... Input part. 15, 24, 28, 34, 41 ... acceleration sensor, 16 ... initial value input unit, 17 ... input value fluctuation estimation unit, 18 ... derivation unit, 19 ... output unit, 21 ... heart rate monitor, 22, 26, 30, 37 ... Mobile terminal, 23 ... Wristwatch type terminal, 25, 29, 36, 39 ... Storage, 27, 31, 38 ... Mobile terminal app, 32, 40 ... Micro device, 33 ... Micro device app, 35, 42 ... Display, 101 ... CPU, 102 ... RAM, 103 ... ROM, 104 ... communication module, 105 ... auxiliary storage device.

Claims (9)

This is data obtained by measuring values of a first type and a second type, which are different types, and is a time-series data in which the value measurement timing is associated with the first type and the second type. Learning data input means for inputting the first learning data and the second learning data;
The variation of each value of the second learning data toward the value of the timing at which the value is measured next to the timing at which the value is measured is related to the variation of the value of the second learning data. Correct answer data, and for each measurement timing, a set of a value of the first learning data, a value of the second learning data, and correct data relating to fluctuations in the value of the second learning data Is used for statistical learning, and by performing statistical learning, an estimation function is generated that estimates a variation in the value of the second type from the value of the first type and the value of the second type. Function generation means;
Input means for receiving the input value of the first type and the input value of the second type;
The timing next to the timing related to the input value of the second type of the value of the second type from the input value of the first type, the input value of the second type, and the estimation function. Input value fluctuation estimation means for estimating fluctuations in the input value that are fluctuations toward
Derivation means for deriving a value of the second type at a timing next to a timing related to the input value of the second type from the input value variation and the input value of the second type;
Output means for outputting the value of the second type derived by the derivation means;
A time-series data processing system comprising:   The estimation function generating means, for each measurement timing, a value obtained by multiplying a value of the first learning data, a value of the second learning data by a predetermined coefficient, and the predetermined coefficient The time-series data processing system according to claim 1, wherein a set of correct data relating to fluctuations in the value of the second learning data is further set as statistical learning data.   3. The time-series data processing system according to claim 1, wherein the input unit accepts the value of the second type output by the output unit as the input value of the second type at the next timing.   The time series data processing system according to any one of claims 1 to 3, wherein the fluctuation is either ascending or descending. This is data obtained by measuring values of a first type and a second type, which are different types, and is a time-series data in which the value measurement timing is associated with the first type and the second type. Learning data input means for inputting the first learning data and the second learning data;
The variation of each value of the second learning data toward the value of the timing at which the value is measured next to the timing at which the value is measured is related to the variation of the value of the second learning data. Correct answer data, and for each measurement timing, a set of a value of the first learning data, a value of the second learning data, and correct data relating to fluctuations in the value of the second learning data Is used for statistical learning, and by performing statistical learning, an estimation function is generated that estimates a variation in the value of the second type from the value of the first type and the value of the second type. Function generation means;
A time-series data processing system comprising: Input means for receiving input values of a first type and a second type, which are different types;
From the input value of the first type, the input value of the second type, and a predetermined estimation function, the timing of the second type value next to the timing of the input value of the second type Input value fluctuation estimation means for estimating input value fluctuation which is a fluctuation toward timing;
Derivation means for deriving a value of the second type at a timing next to a timing related to the input value of the second type from the input value variation and the input value of the second type;
Output means for outputting the value of the second type derived by the derivation means;
A time-series data processing system comprising: This is data obtained by measuring values of a first type and a second type, which are different types, and is a time-series data in which the value measurement timing is associated with the first type and the second type. A learning data input step for inputting the first learning data and the second learning data;
The variation of each value of the second learning data toward the value of the timing at which the value is measured next to the timing at which the value is measured is related to the variation of the value of the second learning data. Correct answer data, and for each measurement timing, a set of a value of the first learning data, a value of the second learning data, and correct data relating to fluctuations in the value of the second learning data Is used for statistical learning, and by performing statistical learning, an estimation function is generated that estimates a variation in the value of the second type from the value of the first type and the value of the second type. A function generation step;
An input step of receiving the input value of the first type and the input value of the second type;
The timing next to the timing related to the input value of the second type of the value of the second type from the input value of the first type, the input value of the second type, and the estimation function. An input value fluctuation estimation step for estimating an input value fluctuation that is a fluctuation toward
A derivation step of deriving a value of the second type at a timing next to a timing related to the input value of the second type from the input value variation and the input value of the second type;
An output step for outputting the value of the second type derived in the derivation step. This is data obtained by measuring values of a first type and a second type, which are different types, and is a time-series data in which the value measurement timing is associated with the first type and the second type. A learning data input step for inputting the first learning data and the second learning data;
The variation of each value of the second learning data toward the value of the timing at which the value is measured next to the timing at which the value is measured is related to the variation of the value of the second learning data. Correct answer data, and for each measurement timing, a set of a value of the first learning data, a value of the second learning data, and correct data relating to fluctuations in the value of the second learning data Is used for statistical learning, and by performing statistical learning, an estimation function is generated that estimates a variation in the value of the second type from the value of the first type and the value of the second type. A function generation step;
Time-series data processing method including An input step for receiving input values of a first type and a second type which are different types;
From the input value of the first type, the input value of the second type, and a predetermined estimation function, the timing of the second type value next to the timing of the input value of the second type An input value fluctuation estimating step for estimating an input value fluctuation which is a fluctuation toward timing;
A derivation step of deriving a value of the second type at a timing next to a timing related to the input value of the second type from the input value variation and the input value of the second type;
An output step of outputting the value of the second type derived in the derivation step;
Time-series data processing method including