JP2004295861A - Method, device and program for detecting life abnormality - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a life abnormality detecting device capable of objectively detecting life abnormality without imposing a burden on residents. <P>SOLUTION: This life abnormality detecting device for detecting a resident's life behavior different from a normal state is provided with: a sensor information storing means 2; a sensor pattern generating means 4 for acquiring a sensor response pattern of each time by cutting out sensor information by a prescribed time window with respect to an optional time; a cluster analyzing means 6 for acquiring a cluster corresponding to each time by a cluster analysis based on each sensor response pattern; a standard cluster information generating means 8 for generating standard cluster information about a unit time on the basis of cluster transition in a plurality of unit times extracted in a prescribed time interval; a particular cluster information generating means 9 for generating particular cluster information on the basis of cluster transition in a particular unit time to be an evaluation object; and a life abnormality determining means 10 for determining the existence/absence of life abnormalities by comparing the particular cluster information with the standard cluster information. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a living abnormality detection method, apparatus, and program for detecting an unusual living behavior of a resident.

  In Japan, with the rapid aging of the population, the number of deaths due to accidents at home has increased rapidly, and in 2001 it was almost equal to the number of fatalities due to traffic accidents. The number of senior citizens living alone is also increasing, reaching 3 million in 2000. In such a great change in society, there is a need for life monitoring techniques such as a technique for automatically detecting abnormalities of consumers and a technique for knowing the lifestyle of a distant family. Therefore, a technology for automatically detecting a state different from usual based on information obtained from a plurality of sensors arranged in a house has been developed.

For example, Patent Literature 1 discloses a resident's health confirmation method described below. First, current values of a plurality of power supply systems in a house are measured, and characteristic amounts such as a duration of a constant current value and a current level are set. Then, data corresponding to each feature amount is input to the neural network, the actual health state is given as a teacher signal, and iterative calculation is performed while changing each connection weight until the difference between the output value and the teacher signal becomes sufficiently small. Determine each coupling load.
Further, an average value and a standard deviation of the learning feature amount data corresponding to the same time zone of the last several days are calculated, and virtual data is generated by adding and subtracting a real number times the standard deviation to the average value. Then, the virtual data is input to the neural network as boundary data, a teacher signal indicating an abnormality is given, and each connection weight is recalculated so that the output signal matches the teacher signal.

After performing the initial setting in this way, feature amount data is created at several health check timings every day based on the accumulated current value and the like, and when these data are input to the neural network, the health state is determined. , And the judgment result is output.
JP-A-2002-109663 (pages 4-8, FIG. 5)

  However, in the resident's health confirmation method disclosed in Patent Document 1, in an initial setting, feature amount data in a state recognized as a normal state of health is input to a neural network, and a teacher signal indicating the state of health is input. It is necessary to make the learning by giving the initial setting, so that not only the burden required for the initial setting is large, but also the initial setting must be redone each time the life pattern changes, which is troublesome. Furthermore, since it is determined whether or not the subject is healthy based on the subject's subjectivity, there is a problem that it is difficult to objectively determine a lifestyle change.

  In addition, the average value and the standard deviation of the learning feature amount data in the same time period of a plurality of days are used to make the neural network learn the non-healthy state. There was also a problem that it was difficult to set the belt specifically.

  The present invention has been made to solve such a problem, and provides a lifestyle change detection method, apparatus, and program that can objectively detect a lifestyle change without imposing a burden on a resident. With the goal.

The object of the present invention is a lifestyle change detection method for detecting a resident's unusual living behavior, wherein the living state of the resident is measured in time series by a plurality of sensors, and a step of acquiring sensor information. Acquiring a sensor response pattern at each time by cutting out the sensor information at a predetermined time window for an arbitrary time, and performing a cluster analysis based on each sensor response pattern to determine a cluster corresponding to each time. Obtaining, based on the transition of the cluster in a plurality of unit times extracted at a predetermined time interval, generating standard cluster information about the unit time, and in the specific unit time to be evaluated Generating specific cluster information based on the transition of the cluster; and transmitting the specific cluster information to the standard cluster. By comparing the data information is accomplished by living accident detection method comprising determining the presence or absence of life accident.

  In addition, the object of the present invention is a lifestyle change detection device for detecting a resident's unusual living behavior, wherein the sensor information stores sensor information on a resident's living state measured by a plurality of sensors. A storage unit, a sensor pattern generation unit that obtains a sensor response pattern at each time by cutting out the sensor information at a predetermined time window for an arbitrary time, and a cluster analysis based on each sensor response pattern, Cluster analysis means for acquiring a cluster corresponding to time, and standard cluster information generation means for generating standard cluster information about the unit time based on a transition of the cluster in a plurality of unit times extracted at predetermined time intervals And a specific cluster information based on a transition of the cluster in the specific unit time to be evaluated. A particular cluster information generating means for generating, by comparing the specific cluster information and the standard cluster information is accomplished by living accident detection apparatus and a living accident judging means for judging whether the living accident.

  In addition, the above object of the present invention is achieved by a lifestyle change detection program for causing a computer to function as the above-mentioned lifestyle change detection device.

  As is apparent from the above description, according to the present invention, it is possible to objectively detect a lifestyle change without burdening the resident.

  Hereinafter, an embodiment of the present invention will be described with reference to the accompanying drawings. FIG. 1 is a block diagram showing a lifestyle change detection device according to an embodiment of the present invention. As shown in FIG. 1, the lifestyle change detection device 1 is connected to a plurality of sensors S1, S2,..., Sn, and includes a sensor information storage unit 2, a sensor pattern generation unit 4, a cluster analysis unit 6, The system includes a standard cluster information generation unit 8, a specific cluster information generation unit 9, and a lifestyle change determination unit 10. It is preferable that the plurality of sensors S1, S2,..., Sn can be constantly monitored without disturbing the living condition of the resident, and examples thereof include an infrared sensor and an electric energy sensor. The number of sensors is not particularly limited as long as it is plural, but it is preferable that the sensors are arranged so that the movement of the occupants in the whole house can be grasped. The input of the sensor information to the lifestyle change detection device 1 can be performed by, for example, wireless communication.

  The sensor information storage unit 2 stores, as sensor information, measurement data that is constantly input from a plurality of connected sensors S1, S2,..., Sn. The sensor pattern generation means 4 acquires a sensor response pattern at each time based on the sensor information. The cluster analysis unit 6 acquires a cluster corresponding to each time by performing a cluster analysis based on the sensor response pattern. The standard cluster information generating means 8 generates standard cluster information of a unit time based on cluster transitions at a plurality of unit times extracted at predetermined time intervals. The specific cluster information generating means 9 generates specific cluster information based on a transition of a cluster in a specific unit time to be evaluated. The lifestyle change determination means 10 determines the presence or absence of a lifestyle change by comparing the specific cluster information with the standard cluster information.

  In the present embodiment, the cluster analysis unit 6 performs a component analysis such as a principal component analysis on each sensor response pattern, thereby extracting feature information including a plurality of main components, A feature amount generating means 62 for respectively obtaining a feature amount such as a principal component load amount corresponding to each time from the reaction pattern based on the feature information, and performing a cluster analysis of the feature amount to obtain a cluster at each time. Is configured to obtain.

Further, in the present embodiment, the standard cluster information generating means 8 sets the unit time to one day and sets each unit time to be continuous (that is, the time interval between each unit time is 1 second). In addition, based on clusters at the same time on each day obtained from cluster transitions on a plurality of days, a frequency distribution creating unit 81 for creating a cluster occurrence frequency distribution at the time is provided.
Next, an embodiment of a method of detecting a lifestyle change using the above-described lifestyle change detection apparatus 1 will be described. In the present embodiment, a case will be described in which a room sensor and a power sensor are used as the sensors S1, S2,..., And Sn are installed in a house of a family of four (a couple and two children). As each of the installation locations and measurement targets, for example, those shown in FIG. 2 can be mentioned.

  As the occupancy sensor, a configuration in which four infrared sensors having detection areas in four different directions are mounted on the lower part of the lighting fixture and a configuration in which one infrared sensor is mounted are used and emitted from the human body. It can be configured to detect a change in the intensity of infrared rays and transmit an ON / OFF signal every second. As for a device mounted on a lighting fixture, it can be configured to also transmit an ON / OFF signal of the lighting fixture. Further, the electric energy sensor can be configured to output the integrated value of the electric power consumption of various connected home appliances at one minute intervals.

  After connecting such sensors S1, S2,..., Sn to the life-change detection device 1 so as to enable wireless communication, time-series measurement is started. The measured data is stored in the sensor information storage unit 2 as time-series information of the sensor information. FIG. 3 shows an example of the sensor information. The storage of the sensor information is performed for a predetermined period (for example, about two weeks).

  The time-series information shown in FIG. 3 is information that changes at intervals of a minimum of one second, and the sensor response is different even in the same living condition (such as eating and relaxing). Is difficult to detect. Therefore, in the present embodiment, the cluster analysis unit 6 performs cluster analysis based on the sensor response pattern generated by the sensor pattern generation unit 4 to convert the sensor information into information representing a living state.

  More specifically, the sensor pattern generation unit 4 cuts out sensor information by a time window w corresponding to the measurement start time, and generates a sensor response pattern. The time window w is set to 10 seconds, and the sensor response patterns are sequentially acquired while moving the time window every second. The feature information extracting means 61 performs a principal component analysis of each of the obtained sensor reaction patterns (see FIG. 3).

The principal component analysis method can be performed using various known algorithms. In the present embodiment, the sensor response pattern per day obtained every second in a time window of 10 seconds is about 86400 patterns. Because of the huge number, it is difficult to find the main components of the sensor information over several weeks by a direct solution method that calculates eigenvalues. Therefore, a generalized Hebbian algorithm is used to find the principal components using the learning function of the neural network.
Can be preferably exemplified. According to this algorithm, when n number of principal components are obtained, the i-th principal component vector c _i (i = 0,..., N−1) is calculated by the sensor response pattern x after k seconds extracted by the time window w. By successively inputting ^(k) , it is obtained recursively as in the following equation 1.

Here, γ is a constant, and the left superscript t indicates transposition of a vector.

  The number n of the main components can be automatically set by setting the sum of the contribution rates to a minimum value exceeding a predetermined value (for example, about 0.8), whereby only the main components in the sensor response can be set. Can be extracted, and feature information composed of a plurality of principal components can be obtained. FIG. 4 shows an example of the feature information obtained from the sensor information for 14 days, and is composed of ten main components 1 to 10.

Next, the feature amount generation means 62 calculates the principal component load amount (p _i = ^t c _i x / | c _i |) obtained by projecting the sensor response pattern cut out in the time window w onto each principal component vector at each time. Ask for it. The cluster analysis is performed on the principal component load.

There is no particular limitation on the method of cluster analysis. For example, a method based on ART (Adaptive Resonance Theory), which can generate non-hierarchical clusters in a self-organizing manner, can be mentioned. That is, when the vector p having the input principal component load as an element is different from the direction of the already generated m clusters q _i (i = 0,..., M−1), that is, all q _When the following equation 2 holds for _i , a new cluster p is generated.

FIG. 5 is an example of a life information cluster generated by the cluster classification (where α = 0.6) of the principal component load amounts based on the ten principal components shown in FIG. Types of clusters are shown. Note that the principal component load amounts at the time of cluster classification are arranged on the horizontal axis, and the numbers on the horizontal axis in the figure correspond to the numbers of the respective principal components. These clusters are not necessarily strongly related to the meaning of life such as “meal” and “relax”, but are considered to represent the main living state of the family. For reference, direct cluster analysis of sensor information without performing principal component analysis resulted in the generation of a large number of clusters corresponding to fine sensor responses, as shown in FIG. Was. From this result, the effectiveness of performing the cluster analysis of the principal component load can be confirmed.

  The frequency distribution creating means 81 expresses daily life as a transition pattern between clusters by using clusters corresponding to each time. FIG. 7 illustrates an example of a transition pattern, and illustrates a transition state between clusters on each date. From this result, based on the clusters at the same time on a plurality of days, a cluster occurrence frequency distribution at that time is created. If the maximum value of the cluster occurrence frequency at a certain time is normalized to be 1, as shown in FIG. 8, the standard life template (standard cluster Information). In FIG. 8, the higher the frequency, the brighter the display.

The specific cluster information generating means 9 generates a cluster transition pattern (specific cluster information) on a specific day to be evaluated. The lifestyle change determination means 10 evaluates the ordinary degree of the day by comparing the standard lifestyle template including the above-described cluster occurrence frequency distribution with the cluster transition pattern of the day to be evaluated.

The standard life template at time t is P (t; c), and the time t on the i-th day at which the normal degree
The cluster transition pattern X _i in; When (t c), ordinary index O _i of i-th day is given by the following equation 3.

Here, c represents a cluster number.

The ordinary degree O _i indicates the maximum value 1 on the day when only the cluster most likely to occur, which is indicated in the standard life template, occurs. The smaller the value, the larger the deviation from the normal life state. The lifestyle change determination means 10 determines the presence or absence of a lifestyle change by comparing the thus-obtained ordinary degree of the evaluation target day with a preset reference value (for example, 0.75). When the usual degree for 14 days was calculated based on the cluster occurrence frequency distribution shown in FIG. 8, the result shown in FIG. 9 was obtained.

  Inspections based on actual situations for the four days when the degree of normality was below the standard value determined that the mental state was unstable by the psychological test (POMS), the day when the child rested in kindergarten, the day when he went out for a long time It was confirmed that it was a day when there were 5 visitors and that a day different from the usual one could be effectively detected.

  As described above, according to the lifestyle abnormality detection device of the present embodiment, by accumulating a plurality of pieces of sensor information for a predetermined period, it is possible to automatically detect a lifestyle abnormality without imposing a burden on a resident. . Further, even if the layout of the house, the arrangement of sensors, the members of the family, and the like change, the sensor information is automatically reflected in the standard living template (that is, the standard cluster information such as the cluster frequency distribution) by accumulating the sensor information. Therefore, it is possible to easily respond to changes in the environment.

  In the present embodiment, as an example of the standard life template (standard cluster information), a case is described in which a standard cluster pattern including a cluster occurrence frequency distribution using a cluster occurrence frequency (occurrence probability) as an evaluation index is used. It is also possible to use the evaluation index of.

  FIGS. 10 to 12 show the normality of each day when the evaluation index is the transition probability, the duration probability, and the period probability of the cluster, respectively. When the ordinary degree of the day falls below a preset reference value, it is determined that there has been a lifestyle change.

The ordinary degree when the transition probability of the cluster shown in FIG. 10 is used as the evaluation index is determined by calculating the transition frequency distribution between the clusters within a predetermined time width from the cluster at each time by using the standard life template (standard cluster information shown in FIG. 13). ) Is generated and compared with a cluster transition pattern (specific cluster information) on a specific day to be evaluated. In FIG. 13, both the vertical axis and the horizontal axis indicate cluster numbers, and a matrix display is made such that the higher the transition probability from the cluster number on the vertical axis to the cluster number on the horizontal axis, the brighter the area.

  When the duration probability of the cluster shown in FIG. 11 is used as the evaluation index, the standard life template (standard cluster information) shown in FIG. 14 is generated by obtaining the duration distribution of each cluster from the cluster at each time. By comparing this with a cluster transition pattern (specific cluster information) on a specific day to be evaluated, evaluation can be performed. In FIG. 14, the vertical axis indicates the cluster number, and the horizontal axis indicates the duration. The time zone in which the percentage of the duration included is larger is displayed brighter.

  When the periodic probability of the cluster shown in FIG. 12 is used as the evaluation index, the standard life template (standard cluster information) shown in FIG. 15 is generated by obtaining the periodic distribution in which the same cluster appears from the cluster at each time. Then, by comparing this with a cluster transition pattern (specific cluster information) of a specific day to be evaluated, evaluation can be performed. In FIG. 15, the vertical axis represents the cluster number, and the horizontal axis represents the amplitude distribution of the time frequency obtained by Fourier transforming the periodic distribution in which the same cluster occurs. are doing. The horizontal axis representing the time frequency is 0 at both ends, and the time frequency increases in the positive direction from the left end toward the center, while the time frequency increases in the negative direction from the right end toward the center. It is displayed so that the absolute value becomes maximum.

  Further, in addition to the above, for example, the transition frequency between clusters within a predetermined time width can be used as the evaluation index. In FIG. 16, the vertical axis represents the transition frequency between clusters on each day, and a day on which the transition frequency exceeds the upper reference value or falls below the lower reference value is determined to be a day on which a lifestyle change has occurred. The upper limit reference value and the lower limit reference value can be set based on, for example, the standard deviation when the transition frequency distribution of each day is approximated by a normal distribution, and there are more people on a day exceeding the upper limit reference value than usual. It is a day (for example, a visitor day), and a day below the lower limit reference value can be determined to be a day with fewer people than usual (for example, a day out).

  As is clear from FIGS. 10 to 12 and FIG. 16, even when different evaluation indices are used, the date determined to be a life abnormal is common to some extent, and instead of using the cluster occurrence probability as the evaluation index, In addition, even if the transition probability, duration time, period probability, transition frequency, and the like of the cluster are used, the life change can be effectively detected.

  In addition, instead of using a single evaluation index to determine a lifestyle abnormality, a lifestyle abnormality can be determined by combining the evaluation results using a plurality of evaluation indexes as described above. For example, all of the above-mentioned cluster occurrence probability, transition probability, duration probability, period probability, and transition frequency are used as evaluation indices, and the date (April 8, 10, 14, 14) determined to be a life anomaly in any of the evaluation indices , 17, 19 and 20), it is possible to reduce the omission of grasping the day when there is a possibility that some abnormality has occurred. In addition, by extracting the day (April 17) determined to be a life abnormality in all the evaluation indices, it is possible to accurately grasp the day when a significant abnormality occurs.

  Further, by devising an evaluation method based on the evaluation result for each evaluation index, a specific lifestyle change can be detected. For example, the day when the life is determined to be abnormal in one of the evaluation indices of the occurrence probability, the transition probability, the duration probability, and the cycle probability of the cluster, and the life is determined to be the life in the evaluation index of the transition frequency of the cluster (April 18) And 20th), it is possible to accurately grasp the day with the most visitors.

  As described above, according to the method for detecting a lifestyle change using a plurality of evaluation indices, it is possible to grasp not only the presence / absence of a lifestyle change but also the content and degree of the lifestyle change.

  As described above, the embodiments of the present invention have been described, but specific aspects of the present invention are not limited to the above embodiments. For example, in the present embodiment, the unit time of the standard cluster information and the specific cluster information is one day, and the standard cluster information is generated based on each cluster transition on a plurality of consecutive days. It is not always necessary to be one day, and for example, one week can be set as a unit time. Further, each unit time does not necessarily have to be continuous, and for example, a unit of time having a fixed width extracted at predetermined time intervals, such as the same day of the week or a predetermined time zone, can be used as the unit time.

  Further, in the above embodiment, the cluster analysis unit 6 performs principal component analysis on the sensor response pattern to extract characteristic information, and obtains a principal component load amount from each sensor response pattern based on the characteristic information. Although it is configured, the method of component analysis is not necessarily limited to principal component analysis, and factor analysis, independent component analysis, and the like can be applied.

  In addition, it is also possible to create a program for causing a general-purpose computer to function as the above-mentioned life incident detection device, and to distribute the program in the market.

FIG. 1 is a block diagram illustrating a lifestyle change detection device according to an embodiment of the present invention. It is a figure showing an example of an installation place of a sensor connected to the above-mentioned life abnormal detection device, and a measuring object. It is a figure showing an example of sensor information. FIG. 4 is a diagram illustrating an example of feature information. FIG. 3 is a diagram illustrating an example of a cluster. It is a figure showing a comparative example of a cluster. FIG. 6 is a diagram illustrating an example of a transition pattern between clusters. It is a figure showing an example of a standard life template. It is a figure showing an example of an evaluation result of usual degree. It is a figure showing other examples of a usual evaluation result. It is a figure which shows the further another Example of the evaluation result of usual degree. It is a figure which shows the further another Example of the evaluation result of usual degree. It is a figure showing other examples of a standard life template. It is a figure showing still another example of a standard life template. It is a figure showing still another example of a standard life template. It is a figure showing an example of an evaluation result of transition frequency between clusters.

Explanation of reference numerals

DESCRIPTION OF SYMBOLS 1 Life change detection device 2 Sensor information storage means 4 Sensor pattern generation means 6 Cluster analysis means 8 Standard cluster information generation means 9 Specific cluster information generation means 10 Life change determination means S1, S2, ..., Sn sensor

Claims (8)

A living abnormality detection method for detecting an unusual living behavior of a resident,
A step of measuring the living condition of the resident by a plurality of sensors in time series and acquiring sensor information;
Obtaining a sensor response pattern at each time by cutting out the sensor information at a predetermined time window for an arbitrary time;
Acquiring a cluster corresponding to each time by cluster analysis based on each sensor reaction pattern;
Generating standard cluster information about the unit time based on the transition of the cluster in a plurality of unit times extracted at a predetermined time interval;
Generating specific cluster information based on a transition of the cluster in the specific unit time to be evaluated;
Comparing the specific cluster information with the standard cluster information to determine the presence or absence of a lifestyle change. The step of acquiring the cluster is a step of extracting feature information composed of a plurality of components by performing a component analysis on each of the sensor reaction patterns,
Obtaining a feature amount corresponding to each time based on the feature information from each of the sensor reaction patterns,
Acquiring a cluster at each time by performing a cluster analysis of the feature amount. 3. The method according to claim 2, wherein the component analysis is a principal component analysis, and the feature amount is a principal component load amount. The step of generating the standard cluster information includes a step of generating a plurality of standard cluster patterns based on different evaluation indexes for the transition of the cluster,
The step of generating the specific cluster information includes a step of generating a plurality of specific cluster patterns based on the evaluation index,
The step of determining the presence or absence of a lifestyle change includes comparing the specific cluster patterns with the corresponding standard cluster patterns and determining the presence or absence of a lifestyle change based on a combination of the comparison results. 2. The life abnormality detection method according to 1. The method according to claim 4, wherein the evaluation index includes at least two of an occurrence probability, a transition probability, a duration probability, a cycle probability, and a transition frequency of the cluster. The method of claim 1, wherein the plurality of sensors include a plurality of infrared sensors and a plurality of electric energy sensors. A living abnormality detection device for detecting an unusual living behavior of a resident,
Sensor information storage means for storing sensor information about a living condition of a resident measured by a plurality of sensors;
By cutting out the sensor information in a predetermined time window for an arbitrary time, a sensor pattern generation unit that acquires a sensor reaction pattern at each time,
Cluster analysis means for obtaining a cluster corresponding to each time by cluster analysis based on each of the sensor reaction patterns,
Standard cluster information generating means for generating standard cluster information for the unit time based on transitions of the cluster in a plurality of unit times extracted at predetermined time intervals,
A specific cluster information generation unit that generates specific cluster information based on a transition of the cluster in the specific unit time to be evaluated;
A lifestyle change detection device comprising: a lifestyle change determination unit that determines whether there is a lifestyle change by comparing the specific cluster information with the standard cluster information. A lifestyle change detection program for causing a computer to function as the lifestyle change detection device according to claim 7.