Family monitoring system

Abstract

To provide a family monitoring system for accurately estimating life behavior of family.SOLUTION: A family monitoring system includes: an adjustment factor calculation unit 50 for calculating an adjustment factor from human-sensitive data from a human-sensitive sensor installed in each home; a feature amount calculation unit for calculating a feature amount group of a plurality of types from consumed electric power section data indicating consumed electric power in time series for every home; an estimation model learning unit for generating an estimation model composed of a plurality of learned weak learning machines by having the plurality of weak learning machines output a temporary life activity classifier for the family by using learning data composed of actual life behavior classifier of the family in the home in time intervals and feature amount group; an estimation model execution unit for inputting a feature amount group of the home to the learned weak learning machine for the estimation model and having a temporary life behavior classifier output; and a life behavior classifier determination unit 51 for determining a final life behavior classifier for the family based on the temporary life behavior classifier and the adjustment factor.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a family member monitoring system that estimates the household activities of family members at home based on power consumption data. [Background technology]

[0002] Patent Document 1 discloses a monitoring device that detects the occurrence of an abnormality in a resident of a residence based on power usage data related to the power used in the residence. This monitoring device calculates the degree of deviation between an index related to the amount of change in power usage calculated in a time series from the power usage data acquired in a time series and a past index for the same time period, and judges whether an abnormality has occurred in the resident based on the calculated degree of deviation.

[0003] Patent Document 2 discloses a monitoring device that monitors residents in a house. This monitoring device calculates the amount of activity of the resident, which is the result of the resident's operation of electrical appliances, based on multiple frequency components extracted from power consumption, and determines the safety of the resident based on this amount of activity. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 6185683 [Patent Document 2] Patent No. 5452558 Summary of the Invention [Problem to be solved by the invention]

[0005] In Patent Document 1, the amount of change calculated from household power consumption data is compared for the same time period, and an abnormality in the resident is determined based on the degree of deviation. In Patent Document 2, the operation state of the electric appliance by the resident, that is, the state of the resident, is determined based on frequency analysis of the household power consumption data. That is, in Patent Document 1 and Patent Document 2, household power consumption data is used as basic data for monitoring. Furthermore, Patent Document 1 adopts the amount of change in the power consumption data as a feature, and Patent Document 2 adopts the result of frequency analysis of the power consumption data as a feature. However, to analyze the living behavior of a family member who is somewhat active, such as cooking or going out, features derived from specific data alone are insufficient, and it is difficult to accurately determine the state of the family member.

[0006] In view of the above circumstances, an object of the present invention is to provide a family member monitoring system that estimates the daily activities of family members more accurately. [Means for solving the problem]

[0007] a feature calculation unit that calculates a group of a plurality of types of feature values ​​from the power consumption group data; a learning data generation unit that generates learning data including an actual living activity classifier of the household member in the household for the time interval and the group of feature values ​​for the time interval; an estimation model learning unit that uses the learning data to generate an estimation model including a plurality of trained weak learners by ensemble learning of a plurality of weak learners that input the group of feature values ​​and output a provisional living activity classifier of the household member; an estimation model execution unit that inputs the group of feature values ​​of the household to the trained weak learners of the estimation model and causes them to output the provisional living activity classifier; and a living activity classifier determination unit that determines a final living activity classifier of the household member based on the provisional living activity classifier and the adjustment coefficient. a living activity pattern calculation unit that calculates a living activity pattern from the final living activity classifier group; Equipped with The adjustment coefficient calculation unit determines a specific living activity classifier that is most likely in the time interval based on the human detection data, and calculates the adjustment coefficient so as to increase a weight of the determined living activity classifier. .

[0008] According to this configuration, power consumption interval data is extracted from the household power consumption data at a predetermined time interval, and multiple types of feature values ​​are calculated from the same power consumption interval data. An estimation model is generated through learning, which is composed of multiple weak learners that use the multiple types of feature values, i.e., a group of feature values, as input data and output a provisional living behavior classifier of the household member for this time interval. The learning data used during the learning includes an actual living behavior classifier (correct answer data) of the household member for the time interval and a group of feature values ​​for the time interval. The feature values ​​calculated for each power consumption interval data are input to each of the learned weak learners, and a provisional living behavior classifier is output from each of the weak learners for each of the power consumption interval data. Furthermore, the multiple provisional living behavior classifiers output from each of the multiple weak learners created using different combinations of feature values ​​randomly selected from the multiple types of feature values ​​are adjusted (weighted) by an adjustment coefficient calculated from human presence data of a human presence sensor installed in each household. A final living behavior classifier is output as a final estimation result based on the multiple provisional living behavior classifiers adjusted by the adjustment coefficient. In other words, the estimation results of an estimation model consisting of multiple weak learners are adjusted with an adjustment coefficient based on human detection data to arrive at a final conclusion, enabling more accurate estimation of household members' daily activities.

[0009] Many of the daily activities of household members depend on the time zone. Therefore, it is advantageous to associate the actual time with the time interval from which the power consumption section data is extracted. The time data (timestamp) is also adopted as a feature, and the time zone is used as one of the branching conditions in the weak learner, so that the estimation of daily activities becomes more accurate. The human presence data from the human presence sensor is also influenced by the time zone. For example, if no person is detected in the middle of the night, it is highly likely that the person is "asleep," and if no person is detected during the day, it is highly likely that the person is "outside the home." For this reason, it is preferable that the human presence data is accompanied by time data indicating the time when the data was acquired as an attribute value. For this reason, in one embodiment, the feature group and the human presence data are provided with time data (timestamp) that associates the actual time with the time interval.

[0010] In one preferred embodiment, the time interval is several tens of minutes, the living behavior classifier includes "sleeping," "at home," "going out," and "cooking," and the type of the feature includes at least one selected from the number of peaks, minimum value, maximum value, average value, standard deviation, difference value in consecutive short time periods, power spectrum in a specific cycle, previous feature calculated immediately before in a time series, and the living behavior classifier output immediately before in a time series. In this configuration, the living behavior classifier includes "sleeping," "at home," "going out," and "cooking" as important behaviors in knowing the living patterns of the family member to be monitored. As features related to the power consumption in the power consumption interval data related to these living behavior classifiers, the number of peaks, minimum value, maximum value, average value, standard deviation, difference value in consecutive short time periods, power spectrum in a specific cycle, previous feature calculated immediately before in a time series, and living behavior classifier output immediately before in a time series are used. By using such a variety of features, a large number of weak learners with various decision characteristics are created. This makes it possible to more accurately estimate the living behavior of the family member.

[0011] In one preferred embodiment of the present invention, a living behavior pattern calculation unit is provided that calculates a living behavior pattern from a tentative living behavior classifier group for an estimation period longer than the time interval output from the estimation model execution unit based on a feature group calculated from the power consumption interval data, and a final living behavior classifier group determined based on an adjustment coefficient calculated from the human detection data for the estimation period. If the estimation period is one day, the calculated living behavior pattern can be used to understand the state of a day, which is a basic unit of life for a family member. If the daily living behavior patterns of the family members are known, it can be determined whether the family members are in an abnormal state. Therefore, in one preferred embodiment of the present invention, an abnormal situation determination unit is provided that determines an abnormal situation of the family members for each household based on the living behavior patterns calculated by the living behavior pattern calculation unit.

[0012] In many cases, household members who need to be monitored are instructed to behave as regularly as possible every day. Therefore, if a lifestyle pattern that stands out from the average lifestyle pattern in the past is calculated, there is a high possibility that an abnormal situation has occurred. For this reason, it is preferable to store the daily lifestyle patterns of household members who need to be monitored as history data and compare this history data with a newly calculated lifestyle pattern to determine whether the current household member is in an abnormal state. For this reason, the abnormal situation determination unit is preferably configured to compare the lifestyle pattern with a history of lifestyle patterns calculated in the past to determine the abnormal situation.

[0013] As the weak learner, various classifiers, such as a neural network, a support vector machine (SVM), a decision tree, etc., can be used. A decision tree is suitable as a weak learner for ensemble learning. Ensemble learning using multiple decision trees is called a random forest model. In a preferred embodiment of the present invention, the weak learner is a decision tree, and the estimation model is a random forest model.

[0014] Other features, operations, and advantages of the present invention will become apparent from the following description of the invention using the accompanying drawings. [Brief description of the drawings]

[0015] [Figure 1] FIG. 1 is a schematic diagram of a family monitoring system. [Diagram 2] FIG. 2 is a functional block diagram showing functions of the family monitoring system. [Diagram 3] FIG. 1 is a functional block diagram showing the basic configuration of an ensemble learning unit configured as a random forest unit. [Figure 4] FIG. 1 is an explanatory diagram illustrating a learning state of a decision tree as a weak learner that constitutes a random forest unit. [Diagram 5] 11 is an explanatory diagram illustrating a state in which a final living activity classifier is output by a random forest unit and a living activity classifier determination unit. FIG. [Figure 6] 1 is a functional block diagram showing the configuration of a preferred embodiment of a family member monitoring system. [Figure 7] 13 is a table showing the activity ratios between reference households and estimated target households using each sub-random forest. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] The family member monitoring system WS according to the present invention will be described below with reference to the drawings. Fig. 1 shows an overall view of the family member monitoring system WS.

[0017] Households participating in the household monitoring system WS receive electricity from the power supply company EC. Household power consumption data, which indicates the power consumption of each household in a time series measured by a meter (power meter) M installed in the house, is sent to the power supply company EC via a communication line. Furthermore, at least one human presence sensor PS is installed in the house in each household participating in the household monitoring system WS. The human presence data from the human presence sensor PS is sent to the management center CC via the communication line. The management computer system CS that constitutes the household monitoring system WS receives household power consumption data associated with time data from the power supply company EC, and receives the human presence data associated with time data from the management center CC. The management computer system CS also includes a user management unit UC that notifies the monitoring information of registered users (household members) via the Internet or a public line.

[0018] As shown in FIG. 2, the family member monitoring system WS includes an ensemble learning unit ALU, an adjustment coefficient calculation unit 50, a living behavior classifier determination unit 51, a living behavior pattern calculation unit 52, an abnormal situation determination unit 60, and a notification unit 70. The ensemble learning unit ALU estimates and outputs a plurality of tentative living behavior classifiers at a predetermined time interval based on the household power consumption data. The adjustment coefficient calculation unit 50 calculates an adjustment coefficient at a predetermined time interval based on the human detection data. The human detection data includes the time when a person is detected. That is, from the human detection data, it is possible to read information on whether a person is detected at a predetermined time interval, and, if a person is detected, information including the frequency of detection. The adjustment coefficient calculation unit 50 determines the most likely living behavior classifier at each time interval based on such human detection data, and calculates an adjustment coefficient that increases the weight of the determined living behavior classifier. For example, a configuration can be adopted in which a living behavior classifier is determined using a lookup table or the like based on the time period of human detection and its frequency, and an adjustment coefficient that increases the weight of the determined living behavior classifier is calculated. If ranking is given to the determination of the living activity classifiers, the adjustment coefficient is calculated so that the weight of each living activity classifier is assigned according to the ranking. The living activity classifier determination unit 51 outputs a final living activity classifier based on the multiple provisional living activity classifiers (provisional living activity classifier group) output from the ensemble learning unit ALU and the adjustment coefficient calculated by the adjustment coefficient calculation unit 50. Here, "sleeping", "at home", "going out", and "cooking (cooking)" are handled as the living activity classifiers.

[0019] The living activity pattern calculation unit 52 calculates a living activity pattern for an estimated period (preferably one day) from a time-dependent collection of final activity classifiers (final living activity classifier group) at a predetermined time interval determined by the living activity classifier determination unit 51. The abnormal situation determination unit 60 determines an abnormal situation for each household member based on the daily living activity pattern. For example, if an event occurs in which the "sleeping" time or "outing" time is abnormally long, or "cooking" time is missing, it can be determined that an abnormal situation has occurred. The notification unit 70 notifies the corresponding registered user of the abnormal situation determined by the abnormal situation determination unit 60.

[0020] Figure 3 shows the functional blocks and information flow of the ensemble learning unit ALU. The ensemble learning unit ALU is constructed using the random forest algorithm. In other words, the weak learner in the ensemble learning is a decision tree.

[0021] The ensemble learning unit ALU includes a household power consumption data input unit 11, an interval data extraction unit 12, a feature calculation unit 13, a learning data generation unit 20, an estimation model learning unit 30, an estimation model execution unit 31, and a model creation management unit 40. The model creation management unit 40 controls the operation of each of the household power consumption data input unit 11, the interval data extraction unit 12, the feature calculation unit 13, the learning data generation unit 20, the estimation model learning unit 30, and the estimation model execution unit 31.

[0022] The household power consumption data input unit 11 inputs household power consumption data of registered users. In this embodiment, the household power consumption data is time-series power data in 30-second units. Furthermore, this household power consumption data indicates the relationship between actual time and power. The section data extraction unit 12 extracts power consumption section data at a predetermined time interval from the household power consumption data. In this embodiment, 30 minutes is used as the predetermined time interval. Of course, other time intervals, or time intervals at which a household member's living behavior classifier can be stably obtained and an abnormal situation of a household member can be appropriately determined through a living behavior pattern based on the living behavior classifier, may be used. Since each power consumption section data is assigned a time (time stamp), the behavior of power consumption from a specific time to a specific time on a specific date can be obtained from the power consumption section data.

[0023] The feature amount calculation unit 13 calculates a plurality of types of feature amounts related to power consumption from the power consumption interval data. The feature amount used in this embodiment includes the number of peaks in the power consumption interval data, the minimum value in the power consumption interval data, the maximum value in the power consumption interval data, the average value in the power consumption interval data, the standard deviation in the power consumption interval data, the difference value in consecutive minute intervals (minute time) in minute intervals further divided into the power consumption interval data, and the power spectrum in the power consumption interval data. The power spectrum here includes a power spectrum in a specific period, for example, a power spectrum in a one-minute period, a power spectrum in a two-minute period, and a power spectrum in a 30-minute period. Furthermore, the feature amount includes the previous feature amount (past feature amount) calculated from past power consumption interval data including at least the power consumption interval data immediately before the power consumption interval data to be processed, for example, the minimum value, maximum value, average value, standard deviation, etc. for the past three hours. Furthermore, the preceding tentative living behavior classifier output (estimated) by the ensemble learning unit ALU using these previous feature amounts, for example, the tentative living behavior classifier obtained in the previous 30 minutes, can also be treated as a feature amount.

[0024] The estimation model learning unit 30 generates a random forest model that inputs a group of features and outputs a provisional living behavior classifier of the family member through machine learning. The random forest model is composed of a plurality of decision trees, as shown in FIG. 4. The learning data generating unit 20 generates learning data for learning each decision tree of the random forest model, and provides the learning data to the estimation model learning unit 30. The learning data includes explanatory variables and correct answer data. The explanatory variables are extracted from a group of features calculated from the power consumption section data. The correct answer data is the type of actual living behavior (referred to as a living behavior classifier) ​​performed by the family member in the time interval (time interval) extracted from the power consumption section data used for calculating the group of features. The type of actual living behavior performed by the family member is input by the family member through an online user interface constructed by the user management unit UC. In this embodiment, as shown in FIG. 4, the living behavior classifiers are "sleeping", "at home", "going out", and "cooking".

[0025] As shown in FIG. 5, the random forest model is composed of decision trees, which are weak learners with low correlation with each other. Selected learning data randomly extracted from the learning data is given to each decision tree. At that time, downsampling is performed to adjust the number of extractions according to the number of the smallest living behavior classifiers in the learning data. In other words, in the selected learning data extracted for learning, the number of data in which the correct answer data is "sleeping", "at home", "going out", and "cooking" is approximately equal. For example, if the learning data for one day includes 40 "sleeping", 50 "at home", 30 "going out", and 30 "cooking", the selected learning data actually applied will include 30 "sleeping", 30 "at home", 30 "going out", and 30 "cooking" by downsampling. The selected learning data is given to each decision tree constituting the random forest model.

[0026] In this embodiment, the feature set includes time data (time stamp) that associates the actual time with the time interval of the power consumption section data, so the time data is also used as an explanatory variable. For example, a conditional expression such as "after a certain time" or "before a certain time" is used as an option for a node of a decision tree.

[0027] When the learning of each decision tree is completed and the random forest model is completed, the estimation model training unit 30 outputs model parameters that define the random forest model. In this embodiment, the estimation model training unit 30 provides the model parameters to the estimation model execution unit 31, whereby a random forest model having a plurality of trained decision trees (trained weak learners) is constructed in the estimation model execution unit 31 based on the model parameters.

[0028] In a general random forest model, the output result (final living activity classifier) ​​of the random forest model is determined based on the output results of multiple decision trees, for example, by majority vote. However, in the present invention, the provisional living activity classifiers output from each decision tree are weighted in the living activity classifier determination unit 51 using the adjustment coefficients calculated by the adjustment coefficient calculation unit 50.

[0029] An example of weighting using this adjustment coefficient will be described below. In this example, the random forest model constructed in the estimation model execution unit 31 has five decision trees. For example, the time interval for estimation is from 20:00 to 20:30, and the provisional living behavior classifier output from the first decision tree is "sleeping", the provisional living behavior classifier output from the second decision tree is "going out", the provisional living behavior classifier output from the third decision tree is "sleeping", the provisional living behavior classifier output from the fourth decision tree is "going out", and the provisional living behavior classifier output from the fifth decision tree is "going out". The adjustment coefficient calculation unit 50 calculates an adjustment coefficient that increases the weight of a specific living behavior classifier from the human sensation data in the estimation target time interval. In this example, if the human sensation data indicates no detection between 20:00 and 4:00 the next morning, it is determined that the possibility of "sleeping" is high, and if the human sensation data indicates no detection between 9:00 and 18:00, it is determined that the possibility of "going out" is high, and the adjustment coefficient is determined accordingly. For example, if the human presence data indicates no detection between 8 p.m. and 4 a.m. the following morning, the adjustment coefficient is a coefficient that gives double the weight to "sleeping." Therefore, in this estimation target time interval, the output results from the five decision trees are 3 points for "going out" and 2 points for "sleeping," so that "going out" is the estimation result by simple majority vote. However, since the adjustment coefficient gives double the weight to "sleeping," adjustment by the adjustment coefficient results in 3 points for "going out" and 4 points for "sleeping." Therefore, the living behavior classifier determination unit 51 determines the final living behavior classifier for this estimation target time interval to be "sleeping." Also, if the human presence sensor PS located in the kitchen detects a human, the adjustment coefficient is a coefficient that gives, for example, double the weight to "cooking."

[0030] FIG. 6 shows a functional block diagram showing the main functions of the household monitoring system WS in a further preferred embodiment. In this household monitoring system WS, the configurations of the daily behavior pattern calculation unit 52 and the abnormal situation determination unit 60 are more complicated than those in the functional block diagram shown in FIG. 3. That is, the daily behavior pattern calculation unit 52 includes a sub-model storage unit 31a and a sub-model selection unit 31b, and the abnormal situation determination unit 60 includes a pattern history storage unit 61. The pattern history storage unit 61 stores previously calculated daily behavior patterns in chronological order as daily behavior pattern history. The abnormal situation determination unit 60 calculates the similarity between the daily behavior pattern of the estimated day and the past daily behavior patterns read from the pattern history storage unit 61, and determines the occurrence of an abnormal situation when the similarity is smaller than a threshold level. When the similarity is calculated as the distance between comparison targets, Euclidean distance, Mahalanobis distance, or the like is used.

[0031] The sub-model storage unit 31a stores a plurality of types of sub-random forest models that are learned by the estimation model learning unit 30 for each of a plurality of household groups with different lifestyles and that are constructed by the estimation model executing unit 31 using model parameters as the learning results. That is, in this embodiment, the estimation model learning unit 30 has a function of learning and generating a plurality of types of sub-random forest models for each of a plurality of household groups with different lifestyles, and the estimation model executing unit 31 has a function of outputting a tentative living activity classifier for a household to be estimated using any one of the plurality of trained sub-random forest models.

[0032] The sub-model selection unit 31b selects an optimal sub-random forest model that is optimal for the household to be estimated from the multiple types of sub-random forest models stored in the sub-model storage unit 31a. The selected optimal sub-random forest model is constructed in the estimation model execution unit 31.

[0033] The technique for selecting the optimal sub-random forest model by the sub-model selection unit 31b is described below. First, as a premise of this technique, the households of registered users are classified into several categories based on the characteristics of the households and the characteristics of the household members to be monitored, and the standard household is set for each classification as the reference household. The household characteristics that are the conditions for this classification include regional characteristics (climate, etc.) and dwelling size. The household members' characteristics include gender, age, health condition, etc. A combination of household characteristics and household member characteristics may be handled as the household characteristics. Learning data is generated from the reference household, a sub-random forest model is generated, and stored in the sub-model storage unit 31a. Furthermore, a reference household that is a standard in the household group used to create the selected sub-random forest model is set. The reference households set for each sub-random forest model are managed by the user management unit UC.

[0034] In the monitoring process, sub-random forest models are sequentially read out from the sub-model storage unit 31a. A group of features of a reference household set in the read sub-random forest model is used as input data, and a provisional living behavior classifier is output from the sub-random forest model. Furthermore, a distribution ratio is obtained by statistically processing the living behavior classifiers for one day. This distribution ratio is the activity ratio of the reference household, that is, the reference living behavior pattern. Next, a group of features of the household to be estimated is input into the sub-random forest model, and the living behavior classifier to be estimated is output, and a distribution ratio of the living behavior classifiers for one day is obtained. This distribution ratio is the activity ratio of the household to be estimated, that is, the behavior pattern. Such a process is performed on all sub-random forest models stored in the sub-model storage unit 31a, or on a sub-random forest model selected therefrom. An example of the result of this process is shown in FIG. 7. In FIG. 7, each sub-random forest model is simply named Model No. 1 to Model No. 8. In Fig. 7, the activity ratios estimated for the target household (estimated daily activity pattern) and the activity ratios for the reference household (reference daily activity pattern) are shown for each sub-random forest model. Next, the reference daily activity pattern and the estimated daily activity pattern are compared to calculate the goodness of fit. The sub-random forest model with the highest goodness of fit is the optimal random forest model.

[0035] [Another embodiment] (1) In the above-described embodiment, the daily living activity classifiers were “sleeping,” “at home,” “going out,” and “cooking,” but other daily living activities such as “bathing,” “cleaning,” and “watching television” may be added or may be reduced.

[0036] (2) In the above embodiment, the family member monitoring system WS and the user management unit UC are built in a management computer system CS that manages many homes, but they may be built in separate computer systems. Also, the family member monitoring system WS may be placed in each home as a stand-alone system and configured to be accessible by family members and the like.

[0037] (3) In this application, the terms "home" and "family members" are to be interpreted broadly. For example, a separate living space such as a nursing home may be treated as a home, and the people living there may be treated as family members.

[0038] (4) The human presence sensor PS in this application can be interpreted in a broad sense. The human presence sensor PS may have not only a function of simply detecting a person, but also a function of detecting temperature, humidity, gas usage status, and the like. A collection of dedicated sensors having such different functions may be called the human presence sensor PS. Furthermore, the human presence data may include not only instantaneous values ​​but also data indicating human movement based on human detection over time.

[0039] (5) The division of each functional unit in the functional block diagrams shown in Figures 2 and 3 is an example to make the explanation easier to understand, and various functional units may be integrated or a single functional unit may be divided into multiple units.

[0040] In addition, the configurations disclosed in the above embodiments (including other embodiments, the same applies below) can be applied in combination with configurations disclosed in other embodiments, provided no contradiction arises. Furthermore, the embodiments disclosed in this specification are illustrative, and the embodiments of the present invention are not limited thereto, and can be appropriately modified within the scope that does not deviate from the purpose of the present invention. [Industrial Applicability]

[0041] The present invention is applicable to a householder monitoring system that estimates the daily activities of householders based on household power consumption data and human detection data. [Explanation of symbols]

[0042] 11: Household power consumption data entry section 12: Section data extraction section 13: Feature calculation unit 20: Learning data generation unit 30: Estimation model learning section 31: Estimation model execution unit 31a: Submodel storage section 31b: Submodel selection section 40: Model Creation Management Department 50: Adjustment coefficient calculation unit 51: Living behavior classifier determination unit 52: Life activity pattern calculation unit 60: Abnormal Situation Judgment Department 61: Pattern history memory section 70: Information Department ALU: Ensemble Learning Unit CC: Management Center CS: Control Computer System EC: Electricity supply company PS: Human presence sensor WS: Family monitoring system

Claims

1. A family monitoring system, A human presence sensor installed in each home, an adjustment coefficient calculation unit that calculates an adjustment coefficient from the human presence data from the human presence sensor; a section data extraction unit that extracts power consumption section data at a predetermined time interval from the household power consumption data that indicates the power consumption of each household in a time series; a feature amount calculation unit that calculates a group of a plurality of types of feature amounts from the power consumption section data; a learning data generation unit that generates learning data including an actual living activity classifier of the household member in the home for the time interval and the group of features for the time interval; an estimation model learning unit that uses the learning data to perform ensemble learning on a plurality of weak learners that input the group of features and output a tentative living behavior classifier of the family member, thereby generating an estimation model consisting of a plurality of trained weak learners; an estimation model execution unit that inputs the group of household features to the trained weak learner of the estimation model and outputs the provisional living activity classifier; a living activity classifier determination unit that determines a final living activity classifier of the family member based on the provisional living activity classifier and the adjustment coefficient; a living activity pattern calculation unit that calculates a living activity pattern from the final living activity classifier group, The adjustment coefficient calculation unit determines the specific living activity classifier that is most likely in the time interval based on the human presence data, and calculates the adjustment coefficient so as to increase the weight of the determined living activity classifier.

2. The family member monitoring system according to claim 1 , wherein the feature amount group and the human presence data are provided with time data that associates the time interval with a real time.

3. A family monitoring system as described in claim 1 or 2, wherein the daily living activity classifiers include sleeping, being at home, going out, and cooking, and the types of features include at least one selected from the number of peaks, minimum value, maximum value, average value, standard deviation, difference value over successive small time periods, power spectrum at a specific period, the most recent feature calculated immediately before in a time series, and the daily living activity classifier output immediately before in a time series.

4. A family monitoring system as described in any one of claims 1 to 3, wherein the daily living behavior pattern calculation unit calculates a daily living behavior pattern from a final daily living behavior classifier group determined based on a provisional daily living behavior classifier group for an estimation period longer than the time interval output from the estimation model execution unit based on a group of features calculated from the power consumption interval data, and the adjustment coefficient calculated from the human presence data for the estimation period.

5. The householder monitoring system according to claim 4 , further comprising an abnormal situation determination unit that determines an abnormal situation occurring to a household member of each household based on the daily activity pattern calculated by the daily activity pattern calculation unit.

6. The family monitoring system according to claim 1 , wherein the weak learner is a decision tree, and the estimation model is a random forest model.

Images