JPH05111474A - Sleep depth determination device - Google Patents

Abstract

(57) [Summary] [Purpose] To determine the sleep depth of the human body easily and at low cost. A body movement detecting means 1 for detecting a body movement of a human body and a sleep depth estimating means 10 for estimating a sleep depth of a human body based on an output of a heart rate counting means 9 for counting a heart rate of the human body are provided. The sleep depth estimating means 10 incorporates a neural network schematic means having a plurality of fixed connection weighting coefficients which have already been learned.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a sleep depth determining device for determining sleep depth.

[0002]

2. Description of the Related Art Conventionally, a sleep depth determining device of this type is equipped with various electrodes, as represented by a polygraph, on a human body to detect brain waves, eye movements, myoelectric potentials and perform waveform processing of detection signals. It was to determine the sleep depth.

[0003]

In such a conventional sleep depth determining apparatus, the measurement results of the electroencephalogram, the eye movement, and the myoelectric potential are analyzed by visual work, or by performing complicated signal processing. There has been a problem that complicated work is required, which requires a lot of labor and cost.

The present invention solves the above problems, and provides a sleep depth determination device that can be determined easily and at low cost by estimating sleep depth in real time from a physical quantity that can be actually measured and detected. It has the first purpose.
The second purpose is to easily estimate the sleep depth by a multidimensional information processing method.

[0005]

In order to achieve the above first object, the present invention provides a body movement detecting means for detecting body movement of a human body, and a heart rate of the human body receiving an output signal of the body movement detecting means. And a sleep depth estimating means for estimating a sleep depth based on the outputs of both the body motion detecting means and the heart rate counting means.

In order to achieve the second object, the sleep depth estimating means of the first problem solving means is obtained by a method simulating a neural network composed of a plurality of neural elements and the sleep depth is obtained. The neural network model means has therein a plurality of fixed connection weight coefficients of the neural network for estimating Furthermore, the sleep depth estimation means has a hierarchical neural network schematic means constructed by combining a number of layers composed of a plurality of neural elements.

[0007]

According to the present invention, by inputting each of the body movement information from the body movement detecting means and the heart rate information from the heart rate counting means into the sleep depth estimating means, the sleep depth estimating means
Estimate sleep depth from moment to moment.

Further, the neural network model means constituting the sleep depth estimation means estimates the sleep depth every moment by the already learned coupling weight coefficient.

[0009]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to FIGS.

FIG. 1 is a block diagram of the present embodiment, and FIG. 2 is an external perspective view when the present embodiment is applied to a bed. In the figure, reference numeral 1 is a body movement detecting means for detecting body movement of a human body, which is composed of a piezoelectric element 2, a filter 3, an amplifying means 4, a rectifying means 5, a smoothing means 6, and an AD converter 7. The piezoelectric element 2 is formed by forming a polymer piezoelectric material such as polyvinyl fluoride (PVDF) into a thin film into a tape shape by attaching a flexible electrode film on both surfaces. Here, the piezoelectric element 2 is mounted on the surface of the mattress 8. There is. Reference numeral 9 is a heart rate counting means for counting the heart rate of the human body, 10 is a body movement detecting means 1, a sleep depth estimating means for estimating a sleep depth based on the outputs of the heart rate counting means 9, and 11 is an output of the sleep depth estimating means 10. Is an LED for displaying. Filter 3, amplifying means 4, rectifying means 5, smoothing means 6, A
D converter 7, heart rate counting means 9, sleep depth estimating means 10
Is incorporated in the control unit 12. Sleep depth estimation means 1
0 uses a 4-bit microcomputer in this embodiment.

The means constituting the sleep depth estimating means 10 are
As a multi-dimensional information processing method, it is constructed by a method simulating an optimal neural network. In a method simulating a neural network, a method of using a plurality of connection weighting coefficients of the neural network for estimating sleep depth as a fixed table, a method of leaving a learning function and adapting to the environment and the user, There is. The present embodiment is provided with a sleep depth estimating means 10 having a neural network model means having therein a fixed connection weighting coefficient for estimating the sleep depth obtained by a method simulating a neural network.

The connection weight coefficient fixed in the neural network for estimating the sleep depth can be obtained by making the neural network schematic means learn the correlation between the body movement data, the heart rate data and the sleep depth data. As a neural network schematic means to be used, reference 1 (D.E. Lamelhart et al., 2 authors, translated by Shunichi Amari “PDP Model” 1989)
, 2 (Kaoru Nakano and 7 others, "Basics of Neurocomputers", Corona Publishing Co., Ltd., P102, 1990),
There is one disclosed in Japanese Examined Patent Publication No. 63-55106. Hereinafter, the configuration and operation of a concrete neural network schematic means will be described by taking a multilayer perceptron using an error backpropagation method as the most well-known learning algorithm described in Document 1 as an example.

FIG. 3 is a conceptual diagram of a neural element which is a constituent unit of the neural network model means. In FIG. 3, 21 to 2
N is a pseudo synapse coupling converter simulating the synaptic coupling of nerves, and 2a is the pseudo synapse coupling converters 21 to 2N.
2b is a set nonlinear function, for example, a sigmoid function whose threshold is h, f (y, h) = 1 / (1 + exp (-y + h)) (Equation 1 ) Is a non-linear converter for non-linearly converting the output of the adder 2a. Although omitted because the drawing is complicated, an input line for receiving the correction signal from the correction means is connected to the pseudo synapse coupling converters 21 to 2N and the non-linear converter 2b. Further, the pseudo synapse coupling converters 21 to 2N serve as the coupling weight coefficient of the neural network schematic unit. This neural element has two types of operating modes, signal processing modes.

The operation of each mode of the neural element will be described below with reference to FIG. First, the operation of the signal processing mode will be described. Neural elements are N inputs X _{1 to} X
Receives _n and outputs one output. i-th input signal X
_i is converted into W _i · X _i in the i-th pseudo synapse coupling converter 2 i shown by a square. N signals W ₁ · X ₁ converted by the pseudo synapse coupling converters 21 to 2N
The ~W _n · X _n enters the adder 2a, sum y is sent to the non-linear converter 2b, the final output f (y, h). Next, the operation of the learning mode will be described. In the learning mode, the conversion parameters W _{1 to} W _n and h of the pseudo synapse coupling converters 21 to 2N and the non-linear converter 2b are corrected by the correction signals representing the correction amounts ΔW _{1 to} ΔW _n and Δh of the conversion parameters. Then, it is corrected to W ₁ + ΔW _i ; i = 1, 2, ..., N h + Δh (Equation 2).

FIG. 4 is a conceptual diagram of a signal converting means constituted by connecting four neural elements in parallel. It should be noted that the following description does not specify that the number of neural elements forming the signal converting means is four. In FIG. 4, 211 to 24
Reference numeral 4 is a pseudo synapse coupling converter,
Is an addition nonlinear converter in which the adder 2a and the nonlinear converter 2b described in FIG. 3 are combined. Although omitted because the drawing becomes complicated as in FIG. 3, the input lines for receiving the correction signal from the correction means are pseudo synapse coupling converters 211 to 211.
44 and the addition nonlinear converters 201 to 204. The pseudo synapse coupling converters 211 to 244 also serve as coupling weight coefficients and weight coefficients. Regarding the operation of this signal conversion means, the operation of the neural element described in FIG. 3 is performed in parallel.

FIG. 5 is a block diagram showing the configuration of the signal processing means when the error back-propagation method is adopted as the learning algorithm, and 31 is the above-mentioned signal converting means. However,
Here, M neural elements for receiving N inputs are arranged in parallel. Reference numeral 32 is a correction means for calculating the correction amount of the signal conversion means 31 in the learning mode. Hereinafter, the operation when learning the signal processing means will be described with reference to FIG. The signal converting means 31 has N inputs S
Upon receiving _in (X), it outputs M outputs S _out (X).
The correction means 32 includes an input signal S _in (X) and an output signal S _out.
After receiving (X), the process waits until M error signals δ _j (X) are input from the error calculating means or the signal converting means in the subsequent stage. The error signal δ _j (X) is input and the correction amount is set to ΔW _ij = δ _j (X) · S _jout (X) · (1−S _jout (X)) · S _iin (X) (i = 1 to N, j = 1 to M) (Equation 3) is calculated, and the correction signal is sent to the signal conversion means 31. The signal conversion means 31 modifies the conversion parameters of the internal neural elements according to the learning mode described above.

FIG. 6 shows a multi-layer using a neural network schematic means.
FIG. 31 is a block diagram showing a configuration of a perceptron.
X, 31Y, and 31Z are K, L, and M, respectively.
32X, 32, which is a signal conversion means composed of neural elements
Y and 32Z are correction means, and 33 is an error calculation means.
Is. In the multi-layer perceptron configured as above
The operation will be described with reference to FIG. signal
In the processing means 34X, the signal conversion means 31X receives the input
S_iin(X) (i = 1 to N) and output S_jout(X)
(J = 1 to K) is output. The correction means 32X uses the signal S
_iin(X) and signal S _jout(X), the error signal δ
_jWait until (X) (j = 1 to K) is input. Since
The same processing as below is performed in the signal processing means 34Y and 34Z.
Performed, the final output S from the signal conversion means 31Z_hout(Z)
(H = 1 to M) is output. Final output S _hout(Z)
Is also sent to the error calculation means 33. Error calculation means 33
Is based on the squared error evaluation function COST (Equation 4).
Based on the ideal output T (T₁・ ・ ・ ・ ・ ・ ・ ・, T_M) With
The error is calculated and the error signal δh (Z) is corrected by the correction means 32Z.
Sent to.

[0018]

[Equation 1]

However, η is a parameter that determines the learning speed of the multilayer perceptron. Next, when the evaluation function is a square error, the error signal is δh (Z) = − η · (S _hout (Z) −T _h ) (Equation 5). The correction unit 32Z calculates the correction amount ΔW (Z) of the conversion parameter of the signal conversion unit 31Z according to the procedure described above, calculates the error signal to be sent to the correction unit 32Y based on (Equation 6), and the correction signal ΔW (Z) is sent to the signal converting means 31Z, and the error signal δ (Y) is corrected by the correcting means 32.
Send to Y. The signal conversion means 31Z uses the correction signal ΔW (Z).
Modify internal parameters based on. The error signal δ (Y) is given by (Equation 6).

[0020]

[Equation 2]

[0021] Here, W _ij (Z) signal converting means 31Z
Is a conversion parameter of the pseudo synapse coupling converter of. Hereinafter, similar processing is performed in the signal processing means 34X and 34Y. By repeating the above procedure called learning, the multi-layer perceptron produces an output that closely approximates the ideal output T when given an input. In the above description, the three-stage multi-layer perceptron is used, but this may be any number of stages. In addition, reference 1
The modification method of the conversion parameter h of the non-linear conversion means in the signal conversion means and the method of accelerating learning known as the inertia term are omitted for simplification of description, but this omission is described below. It does not bind the invention.

In this way, the neural network schematic means learns the relationship between body movement data, heart rate data, and sleep depth data,
The method of sleep depth estimation, which is not easy to describe by simple rules, can be expressed in a natural form. In this embodiment, the sleep depth estimating means 10 is configured by incorporating the information thus obtained. Specifically, the signal converting means 31 of the multi-layer perceptron after sufficiently learning is completed.
The sleep depth estimating means 10 is configured by using only X, 31Y, and 31Z as the neural network schematic means.

Data actually learned will be described. FIG. 7 shows changes in body movement information, heart rate information, and sleep depth. To the neural network model means, current body movement information from the body movement detecting means 1 and body movement information before a certain time Δt, heart rate information from the heart rate counting means 9 at present and a heartbeat before a certain time Δt. 4 pieces of numerical information and sleep depth information of the human body as an ideal output are input and learned, and signal conversion means 31X, 31Y, 31Z in the neural network schematic means are established, and these are used as the neural network schematic means for sleep. It is incorporated in the depth estimation means 10. It should be noted that although the input to the neural network model means is 4 information, the above description does not limit the input to the neural network model means to 4 information. The estimation accuracy can be further improved.

Next, the operation will be described based on the block diagram shown in FIG. First, when the piezoelectric element 2 is deformed by the movement of a human body, a voltage is generated from the piezoelectric element 2 according to the degree of the deformation. This output signal is filtered by the filter 3, amplified by the amplification means 4, rectified by the rectification means 5, and input to the smoothing means 6 and the heart rate counting means 9.
The filter 3 has a characteristic of filtering signals having frequencies other than 1 to 5 Hz. FIG. 8 shows output waveforms of the amplification means 4, the rectification means 5, and the smoothing means 6. Amplifying means 4 to FIG.
When the signal of (4) is input to the rectifying means 5, it is rectified and becomes the signal of FIG. When the signal is input to the smoothing means 6, the signal is smoothed and the signal of FIG. 8C is output. The signal smoothed by the smoothing means 6 is AD-converted by the AD converter 7 and input to the sleep depth estimating means 10. FIG. 9 shows an output waveform of the smoothing means 6 when actually going to bed. Bed, turn over,
In the case of a coarse body movement such as getting out of bed, the piezoelectric element 2 is greatly deformed and the smoothing means 6 outputs a high level. Further, when the human body is in a resting state, a relatively low level output is produced from the smoothing means 6 due to a small body movement of the human body. When the human body is absent, the output of the smoothing means 6 is zero. The heart rate counting means 9 counts the heart rate from the output signal of the rectifying means 5,
It is output to the sleep depth estimation means 10. The sleep depth estimating means 10 estimates the sleep depth from moment to moment based on these input signals and information. The LED 11 displays the result estimated by the sleep depth estimating means 10 on a 7-segment LED.

As described above, according to the present embodiment, the sleep depth incorporating the body movement detecting means, the heart rate counting means, and the neural network schematic means having a plurality of fixed coupling weight coefficients of the already learned neural network. Since the estimation means is provided, the sleep depth can be easily determined at low cost.

A plurality of piezoelectric elements 2 may be used, and the accuracy of detecting body movement is improved. Further, the piezoelectric element 2 may be mounted in the body axis direction of the mattress 8 (from the head to the legs).

If the body movement can be detected, another pressure sensitive sensor such as a capacitance sensor or a weight sensor may be used instead of the piezoelectric element.

Further, instead of the piezoelectric element, an infrared sensor, an ultrasonic sensor, an optical sensor or the like may be used to detect the body movement.

Further, the estimation result of the sleep depth estimating means 10 may be output to the recording means and recorded instead of being displayed by the LED.

Further, the present invention may be installed on a sofa, a chair or the like.

[0031]

As is apparent from the above embodiments, according to the present invention, the body movement detecting means for detecting the body movement of the human body, the heart rate counting means for counting the heart rate of the human body, and the body movement detecting means. By including the means and the sleep depth estimating means for estimating the sleep depth based on the output of the heart rate counting means, the sleep depth can be determined easily and at low cost.

Further, the sleep depth estimating means internally stores a plurality of connection weight coefficients of a fixed neural network for estimating a sleep depth obtained by a method simulating a neural network composed of a plurality of neural elements. Having a neural network model means to have,
Alternatively, since it has a hierarchical neural network schematic means constructed by combining multiple layers composed of a plurality of neural elements, the accuracy of sleep depth estimation can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram of a sleep depth determination device according to an embodiment of the present invention.

FIG. 2 is an external perspective view of the same device when it is arranged on a mattress.

FIG. 3 is a conceptual diagram of a neural element which is a constituent unit of a neural network schematic means used in the device.

FIG. 4 is a conceptual diagram of a signal conversion unit composed of neural elements used in the device.

FIG. 5 is a block diagram of a signal processing unit that employs an error backpropagation method as a learning algorithm used in the device.

FIG. 6 is a block diagram showing a configuration of a multilayer perceptron using a neural network schematic means used in the device.

FIG. 7 is a diagram showing an example of body movement information, heart rate information, and sleep depth data given to the apparatus.

FIG. 8 is a diagram showing output waveforms after processing of an amplifying means, a rectifying means, and a smoothing means of the same device.

FIG. 9 is a waveform diagram showing the output from the smoothing means of the device.

[Explanation of symbols]

 1 Body Motion Detection Means 9 Heart Rate Counting Means 10 Sleep Depth Estimating Means

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁵ Identification number Office reference number FI technical display location A61M 21/02 8119-4C A61B 5/04 312 U 7831-4C A61M 21/00 330 A (72) Inventor Kazunari Nishii 1006 Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd.

Claims (3)

[Claims] 1. A body movement detecting means for detecting a body movement of a human body, a heart rate counting means for counting a heart rate of a human body by receiving an output signal of the body movement detecting means, the body movement detecting means and the heart rate. A sleep depth determining device comprising a sleep depth estimating means for estimating sleep depth based on both outputs of the counting means. 2. The sleep depth estimation means internally stores a plurality of fixed connection weighting factors of a neural network which is obtained by a method simulating a neural network composed of a plurality of neural elements and which estimates a sleep depth. The sleep depth determining device according to claim 1, further comprising a neural network model possessing unit. 3. The sleep depth determining apparatus according to claim 1, wherein the sleep depth estimating means has a hierarchical neural network schematic means constructed by combining a number of layers composed of a plurality of neural elements.