JP2019032676A - State estimation device, model learning device, method, and program - Google Patents

Abstract

To provide a state estimation device that can estimate the state of a person, using human-derived data obtained by a person.SOLUTION: A state estimation device 20 acquires human-derived data for estimation. The state estimation device 20 estimates the state of a person to be estimated, based on the human-derived data for estimation and a learned model learned in advance from human-derived data for learning.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a state estimation device, a model learning device, a method, and a program.

  Conventionally, a technique for modeling the state of a building based on building facility data (for example, electricity, water, gas, heat, etc.) is known (for example, Non-Patent Document 1). In addition, a technique for determining the relationship between a plurality of people existing in a space is known (for example, Patent Document 1). Moreover, the technique which estimates the individuality of the person who acts in space is known (for example, patent document 2).

JP 2016-149076 A JP 2016-189073 A

Yuji Takai, “Building Activity Profile: Modeling and Learning of Activities in Buildings”, 2016, National Congress of the Japanese Society for Artificial Intelligence (2016)

  However, in the said nonpatent literature 1, only the state of a building is modeled from the equipment data of a building, and the state of the person in a building cannot be estimated. In addition, the techniques described in Patent Document 1 and Patent Document 2 only estimate a person's relationship or individuality existing in the space, and as in Non-Patent Document 1, estimate the state of a person. I can't do it. For this reason, for example, the state of a person existing in a specific space cannot be estimated. For example, it is difficult to estimate a state related to stress in a closed space or a state related to health damage in an extreme region such as space, the deep sea, or the South Pole.

  In view of the above facts, an object of the present invention is to estimate a person's state using human-derived data representing data obtained from a person.

  In order to achieve the above object, the state estimation apparatus of the present invention includes an information acquisition unit that acquires estimation human-derived data representing data obtained from a person to be estimated, and the estimation person-derived data acquired by the information acquisition unit. And an estimation unit for estimating the state of the person to be estimated based on data and a learned model for humans that has been learned in advance from learning-derived data representing data obtained from a learning person. The Thereby, a person's state can be estimated using the human origin data obtained from a person.

  The state estimation method of the present invention acquires estimation person-derived data representing data obtained from a person to be estimated, and obtains the obtained estimation person-derived data and learning person-derived data representing data obtained from a learning person The computer is caused to execute a process of estimating the state of the person to be estimated based on the learned model for human being learned in advance.

  The program of the present invention includes a computer, an information acquisition unit that acquires estimation human-derived data representing data obtained from a person to be estimated, the estimation human-derived data acquired by the information acquisition unit, and a learning person Is a program for functioning as an estimation unit that estimates the state of the person to be estimated based on a human learned model that has been learned in advance from learning-person-derived data that represents data obtained from.

  The state estimation apparatus of the present invention further includes a control unit that changes a space in the building where the estimation target person exists according to the state of the estimation target person estimated by the estimation unit. Can do. Thereby, the space suitable for a person can be created by changing the space in a building according to human origin data.

  The information acquisition unit of the present invention further acquires equipment-derived data for estimation representing data obtained from equipment of a building where the person to be estimated exists, and the estimation unit is acquired by the information acquisition unit Estimating the state of the building where the person to be estimated exists, based on the facility-derived data for estimation and the learned model for building previously learned from the data derived from the learning facility representing the state of the building for learning, The control unit may change a space in the building where the estimation target person exists according to the state of the estimation target person and the state of the building estimated by the estimation unit. it can. Thereby, the space suitable for a person and a building can be created by changing the space in a building according to human origin data and equipment origin data.

  The information acquisition unit of the present invention acquires the estimation person-derived data of the estimation target person existing in a specific space, and the estimation unit learns the estimation person-derived data acquired by the information acquisition unit, and learning Based on the human learned model previously learned from the learning person-derived data of the learning person existing in the specific space for the estimation person, the state of the estimation target person It is possible to estimate a state relating to stress or a state relating to the health damage of the estimation target person. As a result, it is possible to accurately estimate the state relating to the stress or health damage of a person existing in the specific space.

  The model learning device of the present invention learns a learning model for estimating a person's state from human-derived data representing data obtained from a person, based on learning-derived data representing data obtained from a learning person. Consists of a learning unit. Thereby, the learned model for estimating a person's state from human origin data can be obtained.

  According to the present invention, there is an effect that it is possible to estimate a person's state using human-derived data representing data obtained from a person.

It is a block diagram which shows schematic structure of the state estimation system which concerns on 1st Embodiment. It is explanatory drawing for demonstrating the specific structural example of the state estimation system which concerns on 1st Embodiment. It is explanatory drawing for demonstrating the production | generation process of mode distribution. It is a figure which shows an example of the learning process routine of 1st Embodiment. It is a figure which shows an example of the estimation process routine of 1st Embodiment. It is a figure which shows the experimental example regarding human origin data. It is a block diagram which shows schematic structure of the state estimation system which concerns on 2nd Embodiment. It is explanatory drawing for demonstrating the specific structural example of the state estimation system which concerns on 2nd Embodiment. It is a figure which shows an example of the learning process routine of 2nd Embodiment. It is a figure which shows an example of the control processing routine of 2nd Embodiment. It is a figure which shows an example of the modification of 2nd Embodiment.

  Hereinafter, embodiments of the present invention will be described in detail.

<System Configuration of State Estimation System According to First Embodiment>

  FIG. 1 is a block diagram illustrating an example of a configuration of a state estimation system 100 according to the first embodiment. The state estimation system 100 can be functionally represented by a configuration including an information terminal 10, a state estimation device 20, and an output device 40, as shown in FIG.

  FIG. 2 shows a specific configuration example of the state estimation system 100 according to the first embodiment. As shown in FIG. 2, in the state estimation system 100, the information terminal 10 attached to the person U acquires person-derived data representing data obtained from the person U. The information terminal 10 transmits the human-derived data to the state estimation device 20 by a predetermined communication means. And the state estimation apparatus 20 produces | generates the learned model for estimating the state of the person U according to the human origin data acquired by the information terminal 10. FIG. Further, the state estimation device 20 of the state estimation system 100 estimates the current state of the person U according to the human-derived data sequentially acquired by the information terminal 10.

  The information terminal 10 is attached to a person. The information terminal 10 acquires human-derived data representing data obtained from a person. Then, the information terminal 10 transmits the human-derived data to the state estimation device 20.

  In the present embodiment, a case where a person's activity (kilocalories), the number of steps, and the heart rate (value converted to the number of beats per minute) are used as human-derived data will be described as an example. The information terminal 10 is realized by, for example, APPLE WATCH (registered trademark).

  The information terminal 10 acquires estimation-derived data representing human-derived data obtained from a person to be estimated and learning-person-derived data representing data obtained from a learning person. The estimation person-derived data and the learning person-derived data will be described later.

  The state estimation apparatus 20 includes a CPU (Central Processing Unit), a ROM (Read Only Memory) that stores programs for realizing each processing routine, a RAM (Random Access Memory) that temporarily stores data, and a storage unit. And a computer including a network interface and the like. Functionally, the state estimation device 20 includes an information acquisition unit 22, a learning data storage unit 24, a learning unit 26, a learned model storage unit 28, and an estimation unit 30.

  The information acquisition unit 22 acquires the learning person-derived data and the estimation person-derived data acquired by the information terminal 10. The learning person-derived data is used when a learned model is generated by the learning unit 26 described later. Further, the estimation person-derived data is used when estimating the state of a specific person.

  Each of the learning person-derived data acquired by the information acquisition unit 22 is stored in the learning data storage unit 24. Specifically, the learning data storage unit 24 stores the amount of activity, the number of steps, and the heart rate at each time, which is data derived from the learning person acquired at each time, as time series data.

  In the first embodiment, the state of a person in space is modeled, that is, profiled, from time-series data representing learning-derived data at each time. A profile is expressed by a model representing a combination of modes and transition information described later. In the first embodiment, a human state is estimated using a learned model that is an obtained profile.

  The learning unit 26 learns a learning model for estimating a person's state from human-derived data based on a plurality of learning-derived human data stored in the learning data storage unit 24. In this embodiment, a mixed Gaussian distribution is used as a learning model. In this embodiment, since the amount of activity, the number of steps, and the heart rate of the person are used as the data derived from the person, the mode distribution for each of the amount of activity, the number of steps, and the heart rate of the person is expressed using a mixed Gaussian distribution. Let them learn.

  FIG. 3 is an explanatory diagram for explaining the processing of the learning unit 26 of the present embodiment. In the present embodiment, the following processes (1) to (4) are executed.

(1) The learning-derived data that is time-series data is segmented to generate segmented time-series data for humans.
(2) The dimension reduction of human time-series data is performed to obtain a feature vector group.
(3) Clustering is performed on the feature vector group to estimate the mode distribution.
(4) The transition probability between modes is acquired.

  First, the learning unit 26 divides the time series data for one day every three minutes with respect to the time series data that is the data derived from the learning person stored in the learning data storage unit 24, and classifies the 240-dimensional person. Time-series data. In the present embodiment, since the amount of human activity, the number of steps, and the heart rate are used as the data derived from the person, the learning unit 26 uses the segmented time series for humans for each of the amount of human activity, the number of steps, and the heart rate. Generate data.

Further, as shown in FIG. 3, the learning unit 26 performs dimension reduction on the 240-dimensional segmented time series data X _i for each of the human activity, the number of steps, and the heart rate, and the feature vector X _i. 'Get. A feature vector corresponds to one vector in the feature space. Note that i is an index number for identifying the segmented time series data.

  As a specific dimension reduction method, for example, principal component analysis, multi-dimensional scaling (MDS), or t-SNE method (t-distributed stochastic neighbor embedding) can be used.

Next, as illustrated in FIG. 3, the learning unit 26 clusters a plurality of feature vectors in the feature space to obtain each mode q _j . Note that j is an index number for identifying a mode.

  When performing clustering using a mixed Gaussian distribution, first, the learning unit 26 presets the number of modes. Note that the number of modes is set in advance by, for example, a user or automatically set based on an information amount criterion.

Then, the learning unit 26 uses an EM (Expectation-Maximization) algorithm based on the feature vector obtained from the learning-derived data and a plurality of segmented time-series data to average the Gaussian distribution representing each mode q _j. Find μ _j and the variance-covariance matrix Σ _j . One Gaussian distribution corresponds to one mode. Thereby, a person's mode can be obtained from person origin data. Using these modes, a person's state (for example, a weekday state or a holiday state, a commuting state, etc.) can be estimated.

  In addition, the learning unit 26 calculates a transition probability between modes corresponding to the mode distribution obtained based on each of the divided time series data for humans. Specifically, for example, for each of the segmented time series data for humans, transition between modes depending on the frequency of the mode to which the segmented time series data for humans belonged at the previous time and the mode to which the next time belonged Calculate the probability.

  The learned model storage unit 28 stores a human learned model obtained by the learning unit 26. Specifically, the mode distribution for each of the person's activity amount, the number of steps, and the heart rate is stored as a human learned model.

  Based on the estimation human-derived data acquired by the information acquisition unit 22 and the mode distribution that is a human learned model stored in the learned model storage unit 28, the estimation unit 30 determines the state of the person to be estimated. presume.

  Specifically, the estimation unit 30 estimates which mode of the mode distribution each of the activity amount, the number of steps, and the heart rate acquired by the information acquisition unit 22 belongs to. The estimation unit 30 determines whether the activity amount acquired by the information acquisition unit 22 belongs to each mode of the mode distribution of the activity amount stored in the learned model storage unit 28, thereby obtaining the information acquisition unit The mode of the amount of activity acquired by 22 is estimated.

  In this case, for example, the estimation unit 30 calculates the average value of the Gaussian distribution and the value of the activity amount acquired by the information acquisition unit 22 for each of a plurality of Gaussian distributions of the mixed Gaussian distribution of activity amounts. The difference absolute value between is calculated. And the estimation part 30 estimates that the activity amount acquired by the information acquisition part 22 belongs to the mode of the Gaussian distribution with the smallest difference absolute value.

  Note that the estimation unit 30 may estimate in which mode the estimation person-derived data acquired by the information acquisition unit 22 belongs in consideration of the variance of the Gaussian distribution. For example, the estimation unit 30 calculates the Mahalanobis distance with the value of the estimation-person-derived data acquired by the information acquisition unit 22 using the average of the Gaussian distribution and the variance covariance matrix. Then, the estimation unit 30 estimates that the amount of activity acquired by the information acquisition unit 22 belongs to a Gaussian distribution mode with the smallest Mahalanobis distance. This is synonymous with estimating the mode with the highest probability.

  The output device 40 outputs the human mode estimated by the estimation unit 30 as a result. Specifically, the output device 40 outputs each mode of the activity amount, the number of steps, and the heart rate of the estimation person-derived data estimated by the estimation unit 30 as a result. For example, the output device 40 is realized by a display.

<Operation of state estimation system>

  Next, the operation of the state estimation system 100 will be described. The state estimation system 100 executes a learning process routine and an estimation process routine.

<Learning processing routine>
The information terminal 10 of the state estimation system 100 collects learning person-derived data, and the information acquisition unit 22 acquires the learning person-derived data acquired by the information terminal 10 and stores it in the learning data storage unit 24. And the state estimation apparatus 20 will perform the learning process routine shown in FIG. 4, if the some data derived from a person for learning is stored in the data storage part 24 for learning, and the instruction signal of a learning process is received.

  In step S <b> 100, the learning unit 26 acquires a plurality of learning person-derived data stored in the learning data storage unit 24.

  In step S102, the learning unit 26 divides the time series data for one day every three minutes from the time series data that is the plurality of learning person-derived data acquired in step S100, and uses 240-dimensional human data. Generate segmented time series data.

  In step S104, the learning unit 26 performs dimension reduction on the human segmented time series data generated in step S102 to obtain a feature vector.

In step S106, the learning unit 26 obtains the average μ _j and the variance covariance matrix Σ _j of each mode q _{j in} the mixed Gaussian distribution using the EM algorithm based on the feature vector obtained in step S104. . Further, the learning unit 26 calculates a transition probability between modes corresponding to the Gaussian distribution based on each of the segmented time series data belonging to each Gaussian distribution in the mixed Gaussian distribution.

In step S108, the learning unit 26 _obtains the average μ _j of each mode q _j of the mixed Gaussian distribution generated in step S106, the variance covariance matrix Σ _j, and the transition probability between the modes corresponding to the Gaussian distribution. Then, it is stored in the learned model storage unit 28 as a human learned model, and the learning processing routine is terminated.

<Estimation processing routine>
When the mixed Gaussian distribution as the human learned model is stored in the learned model storage unit 28 and the information acquisition unit 22 receives the estimation human-derived data, the state estimation device 20 performs the estimation processing routine shown in FIG. Run.

  In step S <b> 200, the information acquisition unit 22 acquires presumed person-derived data acquired by the information terminal 10.

  In step S <b> 202, the estimation unit 30 reads a mixed Gaussian distribution that is a human learned model stored in the learned model storage unit 28.

  In step S204, the estimation unit 30 estimates to which Gaussian distribution of the mixed Gaussian distributions read in step S202 the estimated person-derived data acquired in step S200 belongs. By estimating which Gaussian distribution among the mixed Gaussian distributions, the mode of the estimation person-derived data is estimated.

  In step S206, the estimation unit 30 outputs the mode of the estimation-derived data estimated in step S204.

  The output device 40 outputs the estimation person-derived data mode output from the estimation unit 30 as a result.

  As described above in detail, in the first embodiment, the state of the person to be estimated is estimated based on the estimation person-derived data and the human learned model previously learned from the learning person-derived data. Thereby, a person's state can be estimated using the human origin data obtained from a person.

  In addition, changes in space or changes in the environment can be found from human-derived data. Specifically, since a person is influenced by the space to which the person belongs, this profiling makes it possible to read the relationship between the person and the space.

  Moreover, a learned model for estimating a human state from human-derived data can be obtained by learning a learning model for estimating a human state from human-derived data based on the human-derived data for learning. . Thereby, a person's profile can be obtained based on the learned model obtained from the data derived from the person.

<Experimental example>

  Next, an experimental example of this embodiment will be described. FIG. 6 shows an example of a mode obtained from actual learning-person-derived data.

In the example shown in FIG. 6, the “original data” of (a), (d), and (g) are human activity data (a), the number of steps (d), and the heart rate (g ). The horizontal axis of “original data” is time, and the vertical axis is day. The brighter the color, the larger the value. The “dynamics” of (b), (e), and (h) represents each mode that is a learned mixed Gaussian distribution, and the dark line represents the average μ _j of each mode q _j , and the light color Represents the variance Σ _j of each mode q _j . The “mode calendar” in (c), (f), and (i) represents the mode of each day. The horizontal axis of the “mode calendar” is the week number, the vertical axis is the day of the week (month to day from the top), and the color corresponds to the learned mode. Note that (a) to (c) are activity amounts, (d) to (f) are steps, and (g) to (i) are heart rate data.

  Looking at the original data (a) of the activity amount, a large value column appears at about 8:00 in the morning, and it can be seen that the daily commuting activity is at a substantially constant time. However, the amount of activity after the day varies from day to day, and there is no particular activity.

  The results of mode estimation using the original data (a) of activity data are (b) and (c). (B) Referring to the dynamics of the activity amount, it can be seen that the number of activity modes of the subject in the data collection period is four. In addition, referring to (b) dynamics of activity data, it can be seen that mode 1 and 4 show activity during morning commuting, and that no clear peak appears after the evening. It is consistent with the results obtained by visualizing the data. Also, modes 2 and 3 appear as patterns that are difficult to intuition from (a). As shown in (c), it can be seen that these modes appear mainly on weekends and holidays, and that mode 3 has more morning activities than mode 2.

  Next, referring to the results of the number of steps (FIGS. 6D to 6F), it can be seen that a peak appears in the morning and tends to be similar to the result of the activity amount described above. Further, when looking at the original data (d), there is a tendency for a larger value to appear after July, but this is not estimated as an independent mode. If this trend continues even after the current data acquisition period, there is a possibility of being recognized as a new mode. With respect to mode 4, referring to the mode calendar (f), since it appears only on one day, it can be considered that it is a mode of a peculiar day in which there is no similar day in this period.

  Finally, the heart rate results (FIGS. 6 (g) to (i)) will be considered. Referring to the original data (g), the trend is clearly divided between the end of March and April. Since there is no change in the amount of activity or the number of steps during this period, it is considered that some kind of change that greatly affects the heart rate occurred. One possibility is that there was a change in the workplace layout of the test subject during this period. In other words, it seems that this layout change has caused changes in work style and environment, which has influenced the heart rate. For the factors that have influenced this, more detailed analysis and data are required, but it is considered that changes in space or changes in the environment can be found from human-derived data.

<System Configuration of State Estimation System According to Second Embodiment>

  Next, a second embodiment will be described. In the second embodiment, the equipment-derived data representing data obtained from the equipment of the building where the person to be estimated exists is further acquired, and the person to be estimated is based on the equipment-derived data and the learned model for the building. The point which estimates the state of the existing building differs from 1st Embodiment. Furthermore, the second embodiment is different from the first embodiment in that the space in the building where the estimation target person exists is changed according to the state of the estimation target person and the state of the building. In addition, about the part which becomes the structure similar to 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  FIG. 7 is a block diagram illustrating an example of a configuration of the state estimation system 200 according to the second embodiment. The state estimation system 200 can be functionally represented by a configuration including an information terminal 10, a building facility 210, and a state estimation device 220, as shown in FIG.

  FIG. 8 shows a specific configuration example of the state estimation system 200 according to the second embodiment. As shown in FIG. 8, in the state estimation system 200, the information terminal 10 attached to the person U acquires person-derived data. Moreover, in the state estimation system 200, the equipment origin data showing the data obtained from the building equipment 210 which is the equipment of the building where the person U exists are acquired. And in the state estimation system 200, by controlling the building equipment 210 according to the human-derived data and the equipment-derived data, and the learned model for building and the learned model for building previously learned from the learning equipment-derived data. The space in the building where the person U exists is changed. Specifically, for example, in the space where the person U exists, the space where the person U exists is changed by changing the intensity of light output from the building equipment 210 or by outputting sound from the building equipment 210. Let

In the first embodiment, time series data for one day is targeted. However, in the second embodiment, in order to correspond to a shorter time interval, data in a shorter time interval is converted to human-derived data or equipment. Obtain as origin data. Therefore, for example, 1-minute data is set as segmented time-series data X _i every 3 seconds.

  The building equipment 210 is installed in a building. In the present embodiment, a case where the building facility 210 is a lighting facility will be described as an example. The building facility 210 is provided with a sensor (not shown) and the like, and the sensor acquires facility-derived data representing data obtained from the building facility 210.

  The state estimation device 220 functionally includes an information acquisition unit 222, a learning data storage unit 224, a learning unit 226, a learned model storage unit 228, an estimation unit 230, and a control unit 232. Yes.

  The information acquisition unit 222 further acquires equipment-derived data from the building equipment 210 of the building where the person to be estimated exists. In this embodiment, the information acquisition part 222 acquires the illumination intensity data of lighting equipment as equipment origin data. Specifically, the information acquisition unit 222 acquires the learning equipment-derived data and the estimation equipment-derived data acquired from the building equipment 210. The learning equipment-derived data is used when a learning model for building is generated by the learning unit 226 described later. Moreover, the equipment-derived data for estimation is used when estimating the state of the building.

  The learning data storage unit 224 stores each of the learning person-derived data acquired by the information acquisition unit 222 and each of the learning facility-derived data. Specifically, the learning data storage unit 224 further stores illuminance data at each time, which is data derived from the learning equipment acquired at each time, as time-series data.

  The learning unit 226 generates the building learned model from the learning facility-derived data in the same manner as the generation of the human learned model in the first embodiment.

Specifically, first, the learning unit 226 generates time-series data for buildings by dividing the time-series data, which is the learning-derived data stored in the learning data storage unit 224, into predetermined time intervals. Next, the learning unit 226 performs dimension reduction on the building time-series data to obtain a feature vector. Then, the learning unit 226 obtains each mode Q _j by clustering a plurality of feature vectors in the feature space. Note that j is an index number for identifying a mode.

When performing clustering using a mixed Gaussian distribution, first, the learning unit 226 presets the number of modes. Note that the number of modes is preset by the user, for example. Then, the learning unit 226 uses the EM algorithm based on the feature vector obtained from the learning equipment-derived data and the segmented time series data for a plurality of buildings, to calculate the average μ of the Gaussian distribution representing each mode Q _{j. j} and the variance-covariance matrix Σ _j are obtained. Thereby, the mode of a building can be obtained from the equipment-derived data.

  Further, the learning unit 226 calculates the transition probability between modes corresponding to the Gaussian distribution based on each of the segmented time series data for buildings belonging to each Gaussian distribution among the mixed Gaussian distributions. Specifically, for example, for each piece of building time series data, transition between modes according to the frequency of the mode to which the building time series data belongs at the previous time and the mode to which the next time belongs. Calculate the probability.

  The learned model storage unit 228 stores the human learned model and the building learned model obtained by the learning unit 226. As the building learned model, the mode distribution for the illuminance data is stored in the learned model storage unit 228.

  The estimation unit 230 has an estimation target person based on the estimation facility-derived data acquired by the information acquisition unit 222 and the mode distribution that is the learned model for buildings stored in the learned model storage unit 228. Estimate the state of the building.

  Specifically, the estimation unit 230 estimates which mode of the mode distribution the illumination data acquired by the information acquisition unit 222 belongs to. For example, the estimation unit 230 specifies which mode of the mode distribution stored in the learned model storage unit 228 the illuminance data acquired by the information acquisition unit 222 belongs to. The mode of the illuminance data acquired by is estimated. In this case, the estimation unit 230 estimates the illuminance mode of the lighting equipment that is the building equipment 210 as in the mode estimation of the first embodiment.

  The control unit 232 changes the space in the building where the estimation target person exists, according to the estimation target person state and the building state estimated by the estimation unit 230.

  For example, as shown in FIG. 8 above, when the state of the estimation target person U is estimated to be “sleepiness mode” and the lighting equipment that is the building equipment 210 is estimated to be “normal lighting mode”, The control unit 232 controls the lighting equipment so that the lighting mode of the lighting equipment is set to “drowsiness wake-up mode” (more intense light is applied) so that the person sleeps. Change.

<Operation of state estimation system>

  Next, the operation of the state estimation system 200 will be described.

<Learning processing routine>
The information terminal 10 of the state estimation system 200 collects the learning person-derived data, and the information acquisition unit 222 acquires the learning person-derived data acquired by the information terminal 10 and stores it in the learning data storage unit 224. In addition, the information acquisition unit 222 acquires learning-derived data obtained from the building facility 210 and stores it in the learning data storage unit 224. Then, when the plurality of learning person-derived data and the plurality of learning facility-derived data are stored in the learning data storage unit 224 and the instruction signal of the learning process is received, the state estimation device 20 receives the learning process illustrated in FIG. Run the routine.

  In step S <b> 300, the learning unit 226 acquires a plurality of learning person-derived data stored in the learning data storage unit 224. In addition, the learning unit 226 acquires a plurality of facility-derived data stored in the learning data storage unit 224.

  In step S302, the learning unit 226 divides the time series data, which is a plurality of learning-derived data acquired in step S300, into predetermined time intervals, and generates human divided time series data. In addition, the learning unit 226 generates time-series data for buildings by dividing the time-series data that is a plurality of learning-derived data acquired in step S300 into predetermined time intervals.

  In step S304, the learning unit 226 performs dimension reduction on the human segmented time series data generated in step S302 to obtain a human feature vector. In addition, the learning unit 226 performs dimension reduction on the building segmented time series data generated in step S302 to obtain a building feature vector.

In step S306, the learning unit 226 uses the EM algorithm based on the human feature vector obtained in step S304, and the mean μ _j and variance covariance matrix Σ of each mode q _{j in} the mixed Gaussian distribution. _{Find j} . In addition, the learning unit 226 calculates transition probabilities between modes corresponding to the Gaussian distribution based on each of the segmented time series data belonging to each Gaussian distribution in the mixed Gaussian distribution.

In step S307, the learning unit 226 uses the EM algorithm based on the building feature vector obtained in step S304, and the mean μ _j and variance covariance matrix Σ of each mode Q _{j in} the mixed Gaussian distribution. _{Find j} . Further, the learning unit 226 calculates the transition probability between modes corresponding to the Gaussian distribution based on each of the segmented time series data for buildings belonging to each Gaussian distribution among the mixed Gaussian distributions.

In step S308, the learning unit 226 obtains the average μ _j of each mode q _j of the mixed Gaussian distribution generated in step S306, the variance covariance matrix Σ _j, and the transition probability between the modes corresponding to the Gaussian distribution. Then, it is stored in the learned model storage unit 228 as a human learned model. Also, the learning unit 226, and a transition probability between modes corresponding to a Gaussian distribution and mean mu _j and the variance-covariance matrix sigma _j of each mode Q _j of the Gaussian mixture that is generated in step S307, the building Is stored in the learned model storage unit 228 as a learned model, and the learning processing routine is terminated.

<Estimation processing routine>
When the state learning device 20 stores the human learned model and the building learned model in the learned model storage unit 228, and the information acquisition unit 222 receives the estimation human-derived data and the estimation facility-derived data, 10 is executed. The control processing routine is executed each time estimation human-derived data and estimation facility-derived data are received.

  In step S <b> 400, the information acquisition unit 222 acquires estimation person-derived data acquired by the information terminal 10. The information acquisition unit 222 acquires the equipment-derived data for estimation obtained from the building equipment 210.

  In step S <b> 402, the estimation unit 230 reads a mixed Gaussian distribution corresponding to the human learned model stored in the learned model storage unit 228 and a mixed Gaussian distribution corresponding to the building learned model.

  In step S404, the estimation unit 230 belongs to which Gaussian distribution among the mixed Gaussian distributions corresponding to the human learned model read in step S402, the estimation human-derived data acquired in step S400. The mode of the estimation person-derived data is estimated by estimating. In addition, the estimation unit 230 determines which Gaussian distribution of the mixed Gaussian distribution corresponding to the building learned model read in Step S402 belongs to the estimation facility-derived data acquired in Step S400. By estimating, the mode of the equipment-derived data for estimation is estimated.

  In step S406, the control unit 232 controls the building equipment 210 in accordance with the mode of estimation person-derived data and the mode of estimation equipment-derived data estimated in step S404, and ends the control processing routine.

  As described above in detail, in the second embodiment, based on the estimation equipment-derived data and the building learned model, the state of the building where the estimation target person exists is further estimated, and the estimation target person The space in the building where the person exists is changed according to the state of the building and the state of the building. Thereby, the space suitable for a person and a building can be created by changing the space in a building according to human origin data and equipment origin data. In addition to the human-derived data, the space in the building can be changed in consideration of the equipment-derived data of the building.

  Moreover, a space suitable for each person can be created by changing the space in the building based on the profile obtained from the data derived from the person.

  The present invention is not limited to the above-described embodiment, and various modifications and applications can be made without departing from the gist of the present invention.

  For example, in the above-described embodiment, the case where the mixed Gaussian distribution as an example of the learning model is learned by the EM algorithm has been described as an example. For example, another learning model different from the mixed Gaussian distribution may be learned by another learning method different from the EM algorithm.

  Moreover, although the case where the state estimation apparatus performs the learning process and the estimation process has been described as an example in the above embodiment, the apparatus that performs the learning process and the apparatus that performs the estimation process may be configured as separate apparatuses. . In this case, for example, a model learning device including a learning data storage unit, a learning unit, and a learned model storage unit, and a state estimation device including a learned model storage unit 228 and an estimation unit 230 are configured. You may do it. Further, the learned model storage unit 228 may be installed in an external server, and the state estimation device may read the learned model from the learned model storage unit 228 of the external server.

  Moreover, although the said embodiment demonstrated to the case where the mode was calculated | required from human origin data and equipment origin data as an example, you may cluster the data of the obtained mode further. For example, when a plurality of building facilities are installed for one building, a building mode can be assigned to the building by associating a plurality of building equipment modes with one building. In this case, since there is a high possibility that the mode 1 of the building A and the mode 1 of the building B do not correspond to each other, a mode corresponding to the building can be acquired by classifying a plurality of building modes. .

  In the above embodiment, the case where the time series data is divided every predetermined time when generating the time series data has been described as an example. However, the time interval to be divided may be variable. For example, by using scale space filtering, it is possible to automatically detect a part of the time series data where the activity pattern changes greatly. When the time-series data is divided using the detected portion, the divided time-series data divided according to the characteristics of the time-series data can be obtained.

  Further, in the above embodiment, the case where the state of the person is estimated from the learned model for humans or the state of the building is estimated from the learned model for buildings is described as an example, but the present invention is not limited to this. For example, the mode at the next time may be further estimated using the transition probability between modes.

  Moreover, although the case where the person's state was simply estimated from the estimation person-derived data was described as an example in the above embodiment, the present invention is not limited to this. For example, a state relating to human stress or a state relating to human health damage may be estimated as the state of the person existing in the specific space. In this case, the information acquisition unit acquires estimation person-derived data of the estimation target person existing in the specific space. Then, the estimation unit is based on the estimation human-derived data acquired by the information acquisition unit and the human-learned model previously learned from the learning human-derived data of the learning person existing in the specific learning space. The state relating to stress or the state relating to health damage is estimated as the state of the person to be estimated. For example, the degree of stress or the degree of health damage is estimated. Examples of the specific space include a closed space in an extreme region such as the universe, the deep sea, or the South Pole. In this case, for example, a state related to stress or a state related to health damage is estimated in accordance with each mode estimated as the state of the person to be estimated. Then, in accordance with each estimated mode, a sign of a change in the person to be estimated is detected.

  In the first embodiment, the case where the state of the estimation target person is estimated based on the estimation person-derived data and the human learned model has been described as an example. However, depending on the estimated person state Thus, the state of the building where the person exists may be further estimated. For example, as shown in the above experimental example, it is considered that a change in space or a change in environment can be found from human-derived data, so it is considered that a change in the state of a building where a person exists can be detected. .

  In the second embodiment, the state of the building where the estimation target person exists is further estimated based on the estimation facility-derived data and the building learned model, and the estimation target person state and the building state Although the case where the space in the building where a person exists is changed according to the state has been described as an example, the present invention is not limited to this. For example, the space in the building in which the person to be estimated exists may be changed according to only the mode representing the estimated state of the person to be estimated. Thereby, the space suitable for a person can be created by changing the space in a building according to human origin data.

  FIG. 11 is an explanatory diagram for explaining an example in the case where the space in the building in which the person to be estimated exists is changed according to only the mode representing the estimated state of the person to be estimated. For example, as shown in FIG. 11, in a meeting or the like in which a plurality of people participate, human-derived data is acquired in real time from a plurality of people, and a mode of a plurality of people is obtained from the human-derived data and a human learned model. It is possible to estimate and visualize the relationship between a plurality of people using the estimated mode.

  Specifically, in the example shown in FIG. 11, the information terminal 10 is attached to a plurality of people U1, U2, and U3, and the relationship between people on the display device embedded in the desk that is the building facility 210 Is displayed. The state estimation device 220 acquires data derived from each of the plurality of persons U1, U2, and U3 acquired from the information terminal 10. Moreover, the state estimation apparatus 220 estimates the mode showing each state of the persons U1, U2, and U3 according to the data derived from each of the plurality of persons U1, U2, and U3. And the state estimation apparatus 220 changes the display screen of the display apparatus which is the building equipment 210 according to each mode of person U1, U2, and U3.

  For example, as shown in FIG. 11, when the mode of the person U1 and the mode of the person U2 are the same (or the distance between the modes is equal to or less than a predetermined threshold, etc.), on the display screen displayed on the display device, The display device is controlled so that the line connecting the person U1 and the person U2 becomes thick. Further, when the mode of the person U1 is different from the mode of the person U3 (or the distance between the modes is larger than a predetermined threshold), a line connecting the person U1 and the person U2 on the display screen displayed on the display device. The display device is controlled so as to be thin. Thus, by visualizing the state of the conference, it is possible to create an appropriate space for a plurality of people participating in the conference.

  In the said 2nd Embodiment, although the case where the building equipment 210 was lighting equipment was demonstrated to the example, it is not limited to this, Other equipment may be sufficient. For example, when the building equipment 210 is an air conditioner, a mode in which a person is in a state (for example, a hot state) and a mode in which a building is in a state (moderate air conditioning) are Control such as strengthening the air conditioning can be performed.

  Moreover, although the said 2nd Embodiment demonstrated as an example the case where one building installation of one building was made into object, it is not limited to this. For example, based on each learning equipment-derived data obtained from each of a plurality of building facilities installed in one building, a learned model for building corresponding to each of a plurality of construction facilities is generated, and a plurality of buildings A mode representing the state of each facility may be estimated and used when controlling each of the plurality of building facilities. Moreover, you may make it estimate the mode about each of the equipment origin data obtained from the some building equipment installed in the said building about each of several buildings. In this case, for example, a plurality of building equipment modes correspond to one building, such as a lighting equipment mode, an air conditioning equipment mode, and a calling equipment mode. Also, from these modes, a year-end mode, year-end mode, holiday mode, weekday mode, or the like can be estimated.

  In the above description, the program is stored (installed) in a storage unit (not shown) in advance. However, the program is recorded on any recording medium such as a CD-ROM, a DVD-ROM, and a micro SD card. It is also possible to provide it in the form in which it is provided.

DESCRIPTION OF SYMBOLS 10 Information terminal 20,220 State estimation apparatus 22,222 Information acquisition part 24,224 Learning data storage part 26,226 Learning part 28,228 Learned model storage part 30,230 Estimation part 40 Output apparatus 100,200 State estimation system 210 Building Equipment 232 Control Unit

Claims (7)

An information acquisition unit for acquiring estimation-derived data representing data obtained from a person to be estimated;
Based on the estimation human-derived data acquired by the information acquisition unit and a human-learned model previously learned from learning human-derived data representing data obtained from a learning person, the estimation target person's A state estimation device including an estimation unit for estimating a state. In accordance with the state of the estimation target person estimated by the estimation unit, the control unit further includes a control unit that changes a space in the building where the estimation target person exists.
The state estimation apparatus according to claim 1. The information acquisition unit further acquires estimation equipment-derived data representing data obtained from the equipment of the building where the person to be estimated exists.
The estimation unit is based on the estimation facility-derived data acquired by the information acquisition unit, and a building learned model previously learned from learning facility-derived data representing the state of a building for learning. Estimate the state of the building where the person being estimated exists,
The control unit is configured to change a space in the building where the estimation target person exists, according to the state of the estimation target person and the state of the building estimated by the estimation unit.
The state estimation apparatus according to claim 2. The information acquisition unit acquires the estimation person-derived data of the estimation target person existing in a specific space,
The estimation unit is the human learned model previously learned from the estimation human-derived data acquired by the information acquisition unit and the learning human-derived data of the learning person existing in a specific learning space. Based on the above, the state of the person to be estimated is estimated, the state relating to the stress of the person to be estimated or the state of health damage of the person to be estimated
The state estimation apparatus of any one of Claims 1-3. A model learning device including a learning unit that learns a learning model for estimating a human state from human-derived data representing data obtained from a person based on learning human-derived data representing data obtained from a learning person. Get estimator-derived data representing data obtained from the person being estimated,
Estimating the state of the person to be estimated based on the acquired data derived from the estimation person and a human learned model previously learned from data derived from a learning person representing data obtained from the person for learning;
A state estimation method for causing a computer to execute processing. Computer
An information acquisition unit that acquires estimation human-derived data representing data obtained from a person to be estimated, and the estimation human-derived data acquired by the information acquisition unit, and a learning human origin that represents data obtained from a learning person The program for functioning as an estimation part which estimates the state of the said estimation object person based on the human learned model learned beforehand from data.