JP3923035B2 - Biological condition analysis apparatus and biological condition analysis method - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a biological state analysis device and a biological state analysis method for analyzing data obtained by measuring a biological state, and more specifically, recognizing the biological state from data obtained by measuring the biological state. The present invention relates to a living body state analyzing apparatus and a living body state analyzing method.
[0002]
[Prior art]
Automatically recognizing biological behaviors and situations from time-series data obtained by measuring biological conditions is important in fields such as human health management and human body monitoring that immediately detects abnormalities. Technology. This technique can be applied to, for example, realizing health maintenance and appropriate treatment by managing meals and exercise, and detecting sudden changes in health. It is also possible to apply this technology to the provision of detailed services according to the state of the living body.
[0003]
Therefore, a technique for recognizing a behavior of a living body using an acceleration sensor attached to the body is disclosed in Japanese Patent Laid-Open No. 10-113343 (Patent Document 1).
[0004]
In this technique, a frequency component contained in the data is extracted from the measured acceleration data, and an action is recognized according to the level of correlation with a separately stored reference value.
[0005]
Japanese Patent Laid-Open No. 8-292939 (Patent Document 2) corrects temporal changes and individual differences as a time-series data analysis method, particularly as a time-series data analysis method in which the temporal collection interval is not constant. After that, a technique for performing statistical processing is disclosed.
[0006]
Japanese Patent Laid-Open No. 9-294727 (Patent Document 3) discloses an apparatus for measuring calorie consumption during rest or exercise from the pulse rate and body movement due to acceleration.
[0007]
[Patent Document 1]
Japanese Patent Laid-Open No. 10-113343 [Patent Document 2]
JP-A-8-292939 [Patent Document 3]
JP-A-9-294727 [0008]
[Problems to be solved by the invention]
In the action recognition technology using acceleration data represented by the first prior art, frequency components included in the data are extracted. In order to extract the frequency component, Fourier transform or wavelet transform is used. These Fourier transform and wavelet transform can obtain the spectrum intensity for each frequency component, and are effective for characterizing the behavior by frequency characteristics, but the calculation amount is large. Therefore, there is a problem that it takes a lot of processing time and the action cannot be recognized in real time. Further, since a large memory capacity is required when performing Fourier transform or wavelet transform, there is a problem that it is not suitable for realization with a small device.
[0009]
Even in the second prior art, since frequency components are used to represent temporal changes, there are still problems that the processing speed is slow and a large amount of memory is required. Was not suitable. Further, in the present technology, when the time change can be approximated by a specific function, a parameter representing the feature of the function is used. However, this technique cannot be applied when the function to be approximated is unknown.
[0010]
The third conventional technique is a technique for calculating the calorie consumption of the human body mainly for the purpose of health management, so actions such as whether the human body is walking, eating, or desk work, etc. I couldn't recognize it.
[0011]
The present invention has been made in view of the above problems, and an object of the present invention is to provide a biological state analysis apparatus and a biological state analysis method capable of accurately recognizing a state such as a behavior or a state of a living body with a small amount of calculation. It is to provide.
[0012]
[Means for Solving the Problems]
The biological state analyzer as one aspect of the present invention is:
A biological state measurement unit that measures a biological state and acquires time-series acceleration data and time-series biological data;
A feature quantity extraction unit that divides the acceleration data and the biological data into a plurality of sections, and extracts a feature quantity for each section;
A biological state recognizing unit that recognizes the state of the living body using the feature amount extracted from each section of the acceleration data and the feature amount extracted from each section of the biological data;
With
The feature amount extraction unit sets a length of a section for the biological data to be longer than a length of a section for the acceleration data.
[0013]
The method as one aspect of the present invention includes:
A method for processing time-series acceleration data and time-series biological data acquired by measuring a biological state in a biological state analyzer,
A feature quantity extraction step of dividing the acceleration data and the biological data into a plurality of sections in a feature quantity extraction unit, and extracting feature quantities for each section;
A biological state recognition step of recognizing the state of the biological body in the biological state recognition unit using the feature amount extracted from each section of the acceleration data and the feature amount extracted from each section of the biological data;
With
The feature amount extraction step includes setting the length of the section for the biometric data longer than the length of the section for the acceleration data by the feature amount extraction unit.
[0014]
The biological data used in the present invention is data such as pulse wave, skin temperature, GSR (Galvanic Skin Response), heart rate, myoelectricity, electrocardiogram, respiration, brain wave, blood flow, blood pressure, and the like.
[0015]
As the feature amount, it is preferable to use a statistic that can reduce the calculation amount without using a frequency component that requires Fourier transform or wavelet transform.
[0016]
The statistics used in the present invention include, for example, mean, variance, maximum value, minimum value, difference in level, difference between first and last values in a section, sum of fluctuations in a section, and other combinations thereof. Conceivable.
[0017]
As for the length of the section when extracting the feature amount, the length of the section for the biometric data is preferably longer than the length of the section for the acceleration data. Moreover, it is preferable to change the length of the section dynamically during processing.
[0018]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
[0019]
FIG. 1 shows a configuration diagram of a biological state analysis apparatus (behavior recognition apparatus) as an embodiment of the present invention.
[0020]
The behavior recognition apparatus according to the present embodiment includes a signal measurement unit 1, a preprocessing unit (noise removal unit) 2, a feature amount extraction unit 3, and a behavior recognition unit 4. The feature amount extraction unit 3 includes a section size determination unit 5 and a feature amount calculation unit 6.
[0021]
The signal measuring unit 1 measures time-series acceleration data and time-series biological data from a sensor attached to the living body.
[0022]
Here, the biological data is data such as pulse wave, skin temperature, GSR (Galvanic Skin Response), heart rate, myoelectricity, electrocardiogram, respiration, brain wave, blood flow, blood pressure, and the like. Various sensors that measure these data can be attached to multiple locations on the living body, in which case more various behaviors (for example, while eating, listening to music) and situations (for example, stress levels) Can be recognized. In addition to data related to a living body, the temperature and humidity of the environment in which the living body exists may be simultaneously measured.
[0023]
The preprocessing unit (noise removal unit) 2 removes extreme outliers in the measurement data. In addition, when the value is missing due to a data signal communication error or the like, the preprocessing unit 2 adds a value indicating that the value is missing to the missing part.
[0024]
The feature quantity extraction unit 3 extracts various statistics as feature quantities from the preprocessed data (preprocessed data).
[0025]
More specifically, the section size determination unit 5 determines the length of the section from which the feature amount is extracted in the pre-processed data, and the feature amount calculation unit 6 determines the feature amount such as a statistic for each determined section. To extract.
[0026]
The statistic may be an average value, variance, maximum value, minimum value, step difference, difference between the first value and the last value in the interval, the sum of fluctuations in the interval, and combinations thereof.
[0027]
As the feature value, a feature value obtained by calculating biometric data and data obtained from another system is also conceivable. For example, when the data of the environmental temperature where the living body exists and the skin temperature of the living body are obtained, the difference between them may be used as the feature amount.
[0028]
The behavior recognition unit 4 recognizes the behavior and situation of the living body using the calculated feature amount. For recognition, a model that outputs an action or the like based on a feature value is prepared in advance, and the action is classified using the feature value as an input to the model. As the model, a decision tree, a classification tree, a neural network, a regression equation, or the like can be used.
[0029]
In addition, it is possible to prepare feature quantity patterns for each action, obtain the degree of correlation and similarity between the calculated feature quantity and each pattern, and obtain the action with the highest degree of correlation and similarity as the recognition result. Good.
[0030]
Next, the processing operation of the action recognition device configured as described above will be described using a specific example.
[0031]
In the following, a sensor is attached to one wrist of a human body, and biaxial acceleration, skin temperature, pulse wave, and GSR (skin electrical reflection) data are measured by this sensor.
[0032]
FIG. 2 is a flowchart showing processing steps by the action recognition device.
[0033]
First, the signal measuring unit 1 acquires biaxial acceleration data, skin temperature, pulse wave, and GSR data using the above-described sensor (step S1).
[0034]
3 is a data example of measured biaxial acceleration (acceleration 1 and acceleration 2), FIG. 4 is a data example of measured skin temperature, FIG. 5 is a data example of measured pulse wave, FIG. FIG. 4 is a diagram illustrating an example of GSR data. 3 to 6 are obtained by connecting the respective data acquired in a sampling period of 50 ms ⁻¹ with straight lines, for example.
[0035]
Next, the preprocessing unit 2 performs processing such as noise removal and missing value detection from the measured data (step S2). Noise is generated, for example, by mistakes or disturbances at the measurement stage or signal communication stage.
[0036]
FIG. 7 is a diagram for explaining data loss and noise.
[0037]
As shown in part A of FIG. 7, a value extremely deviating from the reference value is regarded as noise, and the preprocessing unit 2 removes such a value as noise.
[0038]
Further, as shown in the part B, when the value is missing without being measured, the pre-processing unit 2 is able to distinguish the normal data such as “−1” so that the missing can be recognized. Enter a value that identifies Here, if the measured signal is simply recorded, there is a risk of missing data. Therefore, at the same time as the signal measurement, serial numbers and data measurement times are added at predetermined intervals. If the additional information is not continuous, it may be considered that there is a defect.
[0039]
Next, as shown in FIG. 2, the feature quantity extraction unit 3 extracts feature quantities from the preprocessed data (step S3).
[0040]
This feature amount extraction process (step S3) includes a section size determination process (step S3-A) and a feature amount calculation process (step S3-B).
[0041]
The section size determination process (step S3-A) is a process in which the section size determination unit 5 determines the length of a continuous section in which the feature amount is calculated. The feature amount calculation process (step S3-B) is a process in which the feature amount calculation unit 6 calculates a feature amount in each determined section.
[0042]
In the section size determination process (step S3-A), the section size is basically set to 10 seconds or less, particularly about 10 seconds as the initial value in the case of acceleration data (section size as will be described later). Change dynamically). On the other hand, in the case of biometric data (skin temperature, pulse wave, GSR data), 40 seconds to 2 minutes, especially about 40 seconds is used as an initial value. This is because acceleration data generally changes immediately after the movement, whereas biometric data generally changes slowly.
[0043]
For example, in the case of jogging, the acceleration changes from the moment when jogging starts, but the pulse (pulse wave) and skin temperature rise (change) only after jogging is started and continued for a while.
[0044]
In particular, the section size in the case of biometric data is longer than the acceleration and is set to about 40 seconds based on experimental results. I will explain this in more detail.
[0045]
FIG. 8 shows the misclassification rate (misrecognition rate) based on the result of an experiment in which the feature amount is calculated by changing the section size of the biometric data and the human body's meal motion is recognized using the feature amount. FIG.
[0046]
Although there are some fluctuations depending on the experimental situation, according to the figure, it can be confirmed that the error decreases when the length of the section exceeds 60 seconds. Therefore, although it is desirable from the aspect of accuracy to set it to 60 seconds or more, it seems that if the section size is set to be long, the change in data cannot be extracted depending on the behavior of the recognition target, and the accuracy may be adversely affected. Therefore, in the present embodiment, as described above, the section size is set to 40 seconds or more.
[0047]
However, depending on the target living body and the characteristics of the action to be recognized, it is preferable to consider so as to determine a more appropriate section size. For example, if the purpose of behavior recognition is to investigate the relationship between behavior in action and instantaneous heart rate changes, the interval for extracting feature values of heart rate data is short, for example about 15 seconds It can also be considered.
[0048]
The section size is not fixed, and the section size can be dynamically changed according to the situation. In particular, in the case of biometric data, since the data change is gradual, it is easier to extract features if the section size is increased to some extent, but if it is too long, there is a risk that a short change in the situation may be overlooked. On the other hand, if the section size is too short, there is a risk of overlooking a long change in the situation. For example, the period of change in biometric data differs between sleeping and active during the day. Therefore, for example, during sleep, since the period of change is generally long, the section size is increased.
[0049]
FIG. 9 is a flowchart showing detailed processing steps of the feature amount extraction processing (step S3).
[0050]
First, the feature amount calculation unit 6 (see FIG. 1) calculates a feature amount in a section with the same section size as the previous one (step S11). Here, the average acceleration, the dispersion of acceleration, the pulse converted from the pulse wave, and the difference in skin temperature (difference between the first value and the last value in the section) are obtained as feature amounts.
[0051]
Next, the section size determination unit 6 checks whether or not the calculated feature value has a change of a predetermined width or more as compared with the feature value of the previous section (step S12). Here, the "previous section" may be the immediately preceding section, 5 sections before, 10 sections before, 1 hour before, etc., but usually one of the sections from the previous section to about 5 minutes before, Or a combination of these is often appropriate.
[0052]
If the section size determination unit 6 determines that the feature amount has changed (Yes in step S12), the section size determination unit 6 determines that the state of the living body has changed and makes it easier to capture the change in the living body. The section size is made shorter than the size (step S13). Then, the feature amount calculation unit 6 calculates a feature amount within the new section size, and the section size determination unit 6 adopts this feature amount instead of the feature amount obtained in step S11 described above, It is sent to the action recognition unit 4 (step S14).
[0053]
On the other hand, when the section size determination unit 6 determines that the feature amount has not changed (No in step S12), the section size determination unit 6 checks whether the current feature amount is a value within a predetermined range (abnormal range). (Step S15). This is because even if there is no change from the previous section, if the living body is abnormal, it is considered that there is a high possibility that the biological data will change in the long run.
[0054]
When the section size determination unit 6 determines that the feature amount is within the predetermined range (Yes in step S15), it can be considered that the change in the living body is gradual, so the section size is changed to be longer ( Step S16). Then, the feature amount calculation unit 6 calculates a feature amount within the new section size, and the section size determination unit 6 adopts this feature amount instead of the feature amount calculated in step S11 described above, It is sent to the action recognition unit 4 (step S14).
[0055]
On the other hand, when the section size determination unit 6 determines that the feature amount is outside the predetermined range (No in step S15), the section size determination unit 6 uses the feature amount calculated in step S11.
[0056]
It is conceivable that the section length changing process in step S13 and step 16 is performed with a width of about ± 10%, for example.
[0057]
Although the feature amount extraction process including the process of dynamically changing the section size has been described above, the section size changing method is not limited to that described above. For example, when knowledge about the measured time and the behavior of the living body is obtained, the section size may be changed using these data as supplementary information. For example, it is possible to lengthen the section at the time when it is apparently sleeping.
[0058]
The feature amount is calculated from the preprocessed data by the feature amount extraction process (step S3 in FIG. 2) described above.
[0059]
FIG. 10 is a chart showing a partial example of the calculated feature amount.
[0060]
As described above, since the section size dynamically changes, as shown in FIG. 10, the feature amount is used after being converted into a time per unit time (for example, 1 minute), for example.
[0061]
Next, as shown in FIG. 2, the action recognition unit 4 performs action recognition based on the extracted feature amount and outputs a recognition result (step S4).
[0062]
In this embodiment, a classification tree model is used for action recognition. Various methods for constructing a classification tree or a decision tree from tabular data have been proposed. For example, in this embodiment, a binary classification tree called CART is used. CART is well-known for books such as “Applied binary tree analysis method by CART” (Atsushi Otaki, Satoshi Horie, Dan Steinberg, published by Nikka Techen Publishers, 1997).
[0063]
In action recognition using a classification tree, the branch determined by the value of the feature amount is traced in order from the root to the last end (leaf). The type of action is assigned to the end. Therefore, data having this feature amount can be interpreted as this behavior or recognition.
[0064]
FIG. 11 is a diagram showing an example of a classification tree used in the present embodiment.
[0065]
For example, the feature amount data on the first line in the table of FIG. 10 can be interpreted as data obtained in the event of music appreciation according to the classification tree of FIG. However, the classification tree in FIG. 11 is extracted for explaining the present embodiment, and values and terms actually used are not limited thereto.
[0066]
As a method used for recognition processing of actions and the like, the classification tree has an advantage in that the amount of calculation is small and the configuration is easy for the user to understand among types in which data is classified using the constructed model. However, the realization of recognition of actions and the like according to the present invention is not limited to the classification tree. Many methods have been proposed for constructing a model from tabular data as shown in FIG. 10 and performing classification using the model. For example, a neural network or a regression equation may be used for recognition processing of actions or the like. Conceivable. Also, it is possible to calculate a similarity or correlation by comparing the calculated feature quantity with a typical feature quantity of each action, and to make an action with a high similarity or correlation degree as a recognition result.
[0067]
In this way, various applications can be considered for recognizing the behavior and situation of a living body. In particular, in the present embodiment, since a reduction in calculation amount and high-accuracy recognition are compatible, a small device can be attached to a living body. Therefore, the burden on the living body can be minimized and used in many situations. For one thing, it is considered effective in terms of health management. For example, by knowing the state of meals, exercise, sleep, etc., it is possible to provide advice for maintaining the health of the living body or for treatment or rehabilitation. Further, in terms of a living body monitoring device, it is possible to detect a sudden change in the state of the living body and issue a warning to another person, or to communicate with a distant party via a network. In particular, according to the present embodiment, it is possible to determine the state of the living body in real time, so it can be said that it is suitable for the monitoring device. In addition, it can be applied to fine-grained services that depend on biological behavior. For example, a service such as showing a place where a person can sit or take a drink when he / she recognizes that his / her living body is tired in the place of going out can be considered. In addition, since the present invention also targets the recognition of the mental state of the living body (for example, the degree of stress), for example, services such as knowing the direction of interest of the living body and providing the content desired by the living body can be considered.
[0068]
As described above, according to the present embodiment, since both acceleration data and biological data are used, it is possible to recognize various biological actions and situations with high accuracy. That is, compared to the case of recognizing actions and situations using only acceleration data, it is possible to correctly recognize situations involving changes in the living body such as meals when used together with biological data. In addition, it is difficult to recognize mental situations from only acceleration data, for example, situations such as tension or relaxation, but it also recognizes mental situations by using biometric data. It is possible. On the other hand, when using only biometric data, there are few actions and situations that can be recognized, for example, when sweating was measured, it was difficult to distinguish whether it was due to exercise or due to tension, By analyzing it together with acceleration data, it is possible to recognize various behaviors and situations in daily life.
[0069]
In addition, according to the present embodiment, since the statistics for each section are used as feature quantities, it is possible to perform a process with a small amount of calculation and less influenced by noise and missing values.
[0070]
Furthermore, according to the present embodiment, since the length of the section for the biometric data is set to be longer than the section for the acceleration data, it is possible to recognize the living body with high accuracy.
[0071]
Further, according to the present embodiment, since the length of the section is dynamically changed, it is possible to recognize the living body with high accuracy.
[0072]
【The invention's effect】
According to the present invention, since both acceleration data and biological data are used, various actions and situations can be easily recognized with high accuracy.
[Brief description of the drawings]
FIG. 1 is a block diagram showing an example of an action recognition apparatus as an embodiment of the present invention.
FIG. 2 is a flowchart showing processing steps by the action recognition device.
FIG. 3 is a graph showing an example of acceleration data.
FIG. 4 is a graph showing an example of skin temperature data.
FIG. 5 is a graph showing an example of pulse wave data.
FIG. 6 is a graph showing an example of GSR (skin electrical reflection) data.
FIG. 7 is a diagram for explaining noise and missing values.
FIG. 8 is a diagram for explaining a relationship between a section size and an erroneous recognition rate.
FIG. 9 is a flowchart showing processing steps by a feature amount extraction unit;
FIG. 10 is a chart showing an example of feature amounts calculated in the present embodiment.
FIG. 11 is an example of a classification tree used in an action recognition unit.
[Explanation of symbols]
1 Signal Measurement Unit 2 Pre-processing Unit (Noise Removal Unit)
3 feature amount calculation unit 4 action recognition unit 5 section size determination unit 6 feature amount calculation unit

Claims (6)

A biological state measurement unit that measures a biological state and acquires time-series acceleration data and time-series biological data;
A feature quantity extraction unit that divides the acceleration data and the biological data into a plurality of sections, and extracts a feature quantity for each section;
A biological state recognizing unit that recognizes the state of the living body using the feature amount extracted from each section of the acceleration data and the feature amount extracted from each section of the biological data;
With
The biological quantity analysis apparatus characterized in that the feature quantity extraction unit sets a length of a section for the biological data to be longer than a length of a section for the acceleration data.   The biological state analysis apparatus according to claim 1, wherein the feature amount extraction unit dynamically changes at least one of a length of a section for the biological data and a length of a section for the acceleration data.   The feature amount extraction unit changes a length of a section for the biometric data based on a feature amount extracted from a section in front of the section, and changes a length of the section for the acceleration data of the section. The living body state analysis apparatus according to claim 2, wherein the biological state analysis apparatus is changed based on a feature amount extracted from a forward section. A method for processing time-series acceleration data and time-series biological data acquired by measuring a biological state in a biological state analyzer,
A feature quantity extraction step of dividing the acceleration data and the biological data into a plurality of sections in a feature quantity extraction unit, and extracting feature quantities for each section;
A biological state recognition step of recognizing the state of the biological body in the biological state recognition unit using the feature amount extracted from each section of the acceleration data and the feature amount extracted from each section of the biological data;
With
The feature amount extraction step includes setting the length of the section for the biometric data longer than the length of the section for the acceleration data by the feature amount extraction unit.   5. The feature amount extraction step includes dynamically changing at least one of a length of a section for the biometric data and a length of a section for the acceleration data by the feature amount extraction unit. The method described in 1.   In the feature amount extraction step, the feature amount extraction unit changes the length of the section for the biometric data based on the feature amount extracted from the section in front of the section, and the section of the section for the acceleration data The method according to claim 5, comprising changing the length based on a feature amount extracted from a section in front of the section.

Images