CN106388779A - A sleep state sample data type marking method and system - Google Patents

Abstract

The invention relates to a sleep state sample data type marking method and system. The method comprises the steps of collecting sample data of a user in sleep state analysis, establishing feature vectors of the sample data of multiple sleep state types and a cluster center formed through feature vector aggregation, and establishing a target function according to the feature vectors and the cluster center thereof, wherein the target function represents minimizing the distances between the sample data of the same type and dictionary atoms and maximizing the distances between atoms of different types; selecting multiple feature vectors from the sample data of multiple sleep state types as the initial values of the atoms and allocating each piece of sample data to the atoms and solving the target function to obtain a classification dictionary; inputting the sample data into the classification dictionary, comparing the types and the distances of the atoms closest to the sample data, and if the distances are less than a preset threshold value, marking the types of the sample data as the same as those of the atoms. The method and the system can mark the types of sample data accurately.

Description

The mask method of sleep state sample data type and system

Technical field

The present invention relates to assisting sleep technical field, more particularly to a kind of mark side of sleep state sample data type Method and system.

Background technology

In sleep, human body has carried out the process self loosened and recover, and therefore good sleep is to maintain healthy A primary condition；But due to the reason such as operating pressure is big, daily life system is irregular, result in the sleep matter of part population Amount is not good enough, shows as insomnia, midnight wakes up with a start.

There are some equipment at present on the market to help people to fall asleep, improved sleep quality.For example specific sleep a certain Pass through the manual intervention such as sound, optical signal, it is to avoid wake user etc. under the state of sleeping soundly under dormancy state.Assisting sleep is set For standby, in order to be really achieved the purpose improving user's sleep quality, the sleep state of correct detection user is extremely important 's.

Clinically mainly adopt at present polysomnogram to identify sleep state, mainly use EEG signals come to sleep into Row analysis, identifies the sleep state of measured by training sleep state model, for example, judges which of sleep user be in Stage, but because the specificity of EEG signals is stronger, and intensity is very weak is easily subject to external interference.What prior art was trained divides There is error to the detection of a lot of users it is difficult to be labeled to the type of sleep state sample data in class device.

Content of the invention

Based on this it is necessary to be directed to the problems referred to above, a kind of mask method of sleep state sample data type is provided and is System, effectively improves the accuracy of sleep state classification device identification.

A kind of mask method of sleep state sample data type, including：

The EEG signals that collection user produces in sleep state analysis, obtain sample data；

Build the cluster center that the characteristic vector of sample data of multiple sleep state types and characteristic vector are assembled, Object function is set up according to described characteristic vector and its cluster center；Wherein, described object function characterizes and minimizes same type Sample data and the distance of dictionary atom, and maximize the distance between different types of atom；

Select several characteristic vectors as the initial value of atom from the sample data of multiple sleep state types respectively, Each sample data is distributed to described atom and solves described object function, obtain classifying dictionary；

Sample data is inputted classifying dictionary, compares type and the distance of the atom nearest with sample data, if apart from little In default threshold value, then the type of this sample data is labeled as consistent with the type of this atom.

A kind of labeling system of sleep state sample data type, including：

Sample collection module, for gathering the EEG signals that user produces in sleep state analysis, obtains sample data；

Dictionary builds module, for building the characteristic vector of sample data and the characteristic vector of multiple sleep state types The cluster center assembled, sets up object function according to described characteristic vector and its cluster center；Wherein, described object function characterizes Minimize the sample data of same type and the distance of dictionary atom, and maximize the distance between different types of atom；

Dictionary training module, for selecting several characteristic vectors respectively from the sample data of multiple sleep state types As the initial value of atom, each sample data is distributed to described atom and solves described object function, obtain classifying dictionary；

Sample labeling module, for sample data is inputted classifying dictionary, compares the class of the atom nearest with sample data Type and distance, if distance, less than default threshold value, the type of this sample data is labeled as consistent with the type of this atom.

The mask method of above-mentioned sleep state sample data type and system, based on have the clustering algorithm of classification capacity Lai Training dictionary, in dictionary, atom corresponds to a kind of sleep state respectively, the parameter with the number of atom as algorithm, passes through during training The sample of littleization same type and the distance of dictionary atom, maximize the mode of the distance between different types of atom simultaneously, To train corresponding atom for every kind of sleep state, then utilize corresponding to different types of atom and distance to sample Type is judged, such that it is able to mark the type of sample data exactly so that being subsequently used for oneself of sleep state classification device More accurate sleep state classification device, the accuracy of lifting sleep state detection can be trained in learning process.

Brief description

Fig. 1 is the flow chart of the mask method of sleep state sample data type of an embodiment；

Fig. 2 is the labeling system structural representation of the sleep state sample data type of an embodiment.

Specific embodiment

Illustrate the mask method of sleep state sample data type of the present invention and the embodiment of system below in conjunction with the accompanying drawings.

With reference to shown in Fig. 1, Fig. 1 is the flow chart of the mask method of sleep state sample data type of an embodiment, Including：

Step S101, the EEG signals that collection user produces in sleep state analysis, obtain sample data；

In this step, when user is carried out with assisting sleep analysis, related transducer equipment is worn by user, detect user EEG signals, gather EEG signals when, can be acquired with 30s for a frame.

Carry out the task of sleep state identification as needed, determine feature data types, extract therewith from EEG signals Corresponding sample data；For example, 1～N kind sleep state to be identified, extract the sample data for carrying out this N kind state recognition.

Step S102, builds the characteristic vector of sample data of multiple sleep state types and characteristic vector gathering forms Cluster center, object function is set up according to described characteristic vector and its cluster center；Wherein, described object function characterizes and minimizes phase The sample data of same type and the distance of dictionary atom, and maximize the distance between different types of atom；

In this step, on the basis of KMeans (K average) and KNN (K is closest) algorithm, design has classification capacity Training dictionary, in dictionary, atom corresponds to a kind of sleep state (such as waking state, sleep state etc.), atom to clustering algorithm respectively Number be algorithm parameter.

When the distance of input sample and atom sufficiently small (when similarity is sufficiently large), then it is considered that the type of sample with former The type of son is consistent；Set up object function, in training by minimize same type sample and dictionary atom away from From maximizing the mode of the distance between different types of atom, all to train accordingly for every kind of particular state simultaneously Atom.

If many classification problems, it is provided with the sample of a total t type,Be characterized to Amount,It is characterized the cluster center of vector gathering, the common version of object function can be expressed as Following form：

In formula, it is provided with the sample data of t kind sleep state type,It is characterized vector,It is characterized the cluster center of vector gathering.

Taking waking state and dormant two class problems as a example, ifFor waking state type (wake) Characteristic vector,The cluster center assembled for the characteristic vector of waking state type,For The characteristic vector of sleep state type (sleep),The cluster assembled for the characteristic vector of sleep state type Center, its object function can be expressed as form：

Described object function is：

In formula,For the characteristic vector of clear-headed type,Spy for clear-headed type Levy the cluster center of vector gathering,For the characteristic vector of sleep pattern,For sleep The cluster center that the characteristic vector of type is assembled, wake represents clear-headed type, and sleep represents sleep pattern.

Step S103, selects several characteristic vectors as atom respectively from the sample data of multiple sleep state types Initial value, each sample data is distributed to described atom and solves described object function, obtain classifying dictionary；

In this step, based on described object function, train classifying dictionary, in training, this programme is in classical KMeans Improve on the basis of algorithm, training process can be as follows taking waking state and dormant two classification problem as a example：

(1) when initializing, if random from the sample data of clear-headed type and the sample data of sleep pattern respectively set Dry characteristic vector is as atom；Each sample data is distributed to the atom away from its nearest neighbours；

(2) update atom, if all sample datas belonging to this atom are consistent with the type of atom (is waking state Type or be sleep state type), then calculate the average of all sample datas belonging to this atom, and in this, as new Atom；

If there is the sample data inconsistent with atomic type, calculate sample data and the sleep class of clear-headed type respectively The average of the sample data of type, calculating process can include equation below：

In formula, c'_wakeFor the average of the sample data of clear-headed type, c'_sleepAverage for the sample data of sleep pattern；

According to the quantity of sample data (negative sample) inconsistent with atomic type and its position of position correction atom, By the location updating of atom in the farther position apart from negative sample data, calculating process can include equation below：

In formula, c is the position of atom after correction, and g is discriminant function, and w is weighted value；

Further, the computing formula of described weighted value w can be as follows：

In formula, w_wakeFor the weighted value of clear-headed type, w_sleepWeighted value for sleep pattern.

As another embodiment, the computing formula of described weighted value w can also be as follows：

(3) if all sample datas belonging to this atom are all inconsistent with the type of atom, change the class of this atom Type, and calculate the average belonging to all sample datas of this atom, and using this average as new atom；

(4) repeated execution of steps (2) and (3) are iterated, and the difference of the atom before and after iteration is less than set point (foot Enough little), or do not have sample data to be allocated to the new atomic time, store current classifying dictionary and exit training.

Step S104, sample data is inputted classifying dictionary, compares type and the distance of the atom nearest with sample data, If distance, less than default threshold value, the type of this sample data is labeled as consistent with the type of this atom.

In this step, using sample data, classifying dictionary is tested, by comparing the atom nearest with sample data Type and distance carry out the type of judgment sample data, if distance less than threshold value then it is assumed that the type of sample data and this atom Type consistent, the judgement of output "true", the type of this sample data is labeled as consistent with the type of this atom, otherwise then refuses Judge absolutely.

Type sample data being marked using such scheme.Can be used for training grader, for example, (supported from svm Vector machine) grader, neural metwork training disaggregated model.

With reference to shown in Fig. 2, Fig. 2 is the labeling system structural representation of the sleep state sample data type of an embodiment Figure, including：

Sample collection module 101, for gathering the EEG signals that user produces in sleep state analysis, obtains sample number According to；

Dictionary builds module 102, for building the characteristic vector of sample data and the feature of multiple sleep state types The cluster center that vector is assembled, sets up object function according to described characteristic vector and its cluster center；Wherein, described object function Characterize the distance of sample data and the dictionary atom minimizing same type, and maximize between different types of atom away from From；

Dictionary training module 103, for selecting several features respectively from the sample data of multiple sleep state types Vector, as the initial value of atom, each sample data is distributed to described atom and solves described object function, classified Dictionary；

Sample labeling module 104, for sample data is inputted classifying dictionary, compares the atom nearest with sample data Type and distance, if distance, less than default threshold value, the type of this sample data is labeled as consistent with the type of this atom.

The labeling system of sleep state sample data type of the present invention and the sleep state sample data type of the present invention Mask method correspond, above-mentioned sleep state sample data type mask method embodiment illustrate technical characteristic And its advantage, all be applied to the embodiment of the labeling system of sleep state sample data type, hereby give notice that.

Each technical characteristic of embodiment described above can arbitrarily be combined, for making description succinct, not to above-mentioned reality The all possible combination of each technical characteristic applied in example is all described, as long as however, the combination of these technical characteristics is not deposited In contradiction, all it is considered to be the scope of this specification record.

Embodiment described above only have expressed the several embodiments of the present invention, and its description is more concrete and detailed, but simultaneously Can not therefore be construed as limiting the scope of the patent.It should be pointed out that coming for those of ordinary skill in the art Say, without departing from the inventive concept of the premise, some deformation can also be made and improve, these broadly fall into the protection of the present invention Scope.Therefore, the protection domain of patent of the present invention should be defined by claims.

Claims (10)

1. a kind of mask method of sleep state sample data type is it is characterised in that include：

The EEG signals that collection user produces in sleep state analysis, obtain sample data；

Build the cluster center that the characteristic vector of sample data of multiple sleep state types and characteristic vector are assembled, according to Object function is set up at described characteristic vector and its cluster center；Wherein, described object function characterizes the sample minimizing same type Data and the distance of dictionary atom, and maximize the distance between different types of atom；

Select several characteristic vectors respectively as the initial value of atom from the sample data of multiple sleep state types, will be each Individual sample data is distributed to described atom and is solved described object function, obtains classifying dictionary；

Sample data is inputted classifying dictionary, compares type and the distance of the atom nearest with sample data, if distance is less than pre- If threshold value, then the type of this sample data is labeled as consistent with the type of this atom.

2. the mask method of sleep state sample data type according to claim 1 is it is characterised in that described target letter Number is：

$min\underset{p=1}{\overset{t}{\Σ}}\underset{i=1}{\overset{n^p}{\Σ}}\underset{j=1}{\overset{k^p}{\Σ}}\left|\right|y_i^p-c_j^p|{|}_2^2$

$\begin{array}{cc}s.t.& max\underset{l=1}{\overset{t}{\Σ}}\underset{s=1}{\overset{t}{\Σ}}\underset{m=1}{\overset{k^l}{\Σ}}\underset{n=1}{\overset{k^s}{\Σ}}\left|\right|c_m^l-c_n^s|{|}_2^2,l\≠s\end{array}$

In formula, it is provided with the sample data of t kind sleep state type,I=1 ..., n^p, p=1 ..., t be characterized vector, J=1 ... k^p, p=1 ..., t is characterized the cluster center of vector gathering.

3. the mask method of sleep state sample data type according to claim 2 is it is characterised in that described sleep shape State type includes clear-headed type and sleep pattern；

Described object function is：

$min\underset{i=1}{\overset{n^{wake}}{\Σ}}\underset{j=1}{\overset{k^{wake}}{\Σ}}\left|\right|y_i^{wake}-c_j^{wake}|{|}_2^2+\underset{i=1}{\overset{n^{sleep}}{\Σ}}\underset{j=1}{\overset{k^{sleep}}{\Σ}}||y_i^{sleep}-c_j^{sleep}|{|}_2^2$

$\begin{array}{cc}s.t.& max\underset{m=1}{\overset{k^{wake}}{\Σ}}\underset{n=1}{\overset{k^{sleep}}{\Σ}}\left|\right|c_m^{wake}-c_n^{sleep}|{|}_2^2\end{array}$

In formula,I=1 ..., n^wakeFor the characteristic vector of clear-headed type,J=1 ..., k^wakeSpy for clear-headed type Levy the cluster center of vector gathering,I=1 ..., n^sleepFor the characteristic vector of sleep pattern,J=1 ..., k^sleepThe cluster center assembled for the characteristic vector of sleep pattern, wake represents clear-headed type, and sleep represents sleep pattern.

4. sleep state sample data type according to claim 2 mask method it is characterised in that described respectively from Several characteristic vectors are selected as the initial value of atom, by each sample data in the sample data of multiple sleep state types Distribute to described atom and solve described object function, the step obtaining classifying dictionary includes：

(1) set several characteristic vectors respectively from the sample data of clear-headed type and the sample data of sleep pattern at random to make For atom；Each sample data is distributed to the atom away from its nearest neighbours；

(2) if all sample datas belonging to this atom are consistent with the type of atom, all samples belonging to this atom are calculated The average of notebook data, and in this, as new atom；

If there is the sample data inconsistent with atomic type, calculate the sample data of type of regaining consciousness and sleep pattern respectively The average of sample data；And the position of the quantity according to the sample data inconsistent with atomic type and its position correction atom, By the location updating of atom in the farther position apart from negative sample data；

(3) if all sample datas belonging to this atom are all inconsistent with the type of atom, change the type of this atom, and Calculate the average belonging to all sample datas of this atom, and using this average as new atom；

(4) repeated execution of steps (2) and (3) are iterated, and the difference of the atom before and after iteration is less than set point, or does not have There is sample data to be allocated to the new atomic time, store current classifying dictionary and exit training.

5. the mask method of sleep state sample data type according to claim 4 is it is characterised in that described step (2) the average process of the sample data settling the sample data of awake type and sleep pattern of falling into a trap includes equation below：

$c_{wake}^{\′}=\frac{1}{n^{wake}}{\mathrm{\Σy}}_i^{wake}$

$c_{sleep}^{\′}=\frac{1}{n^{sleep}}{\mathrm{\Σy}}_j^{sleep}$

In formula, c'_wakeFor the average of the sample data of clear-headed type, c'_sleepAverage for the sample data of sleep pattern.

6. the mask method of sleep state sample data type according to claim 3 is it is characterised in that described step (2) position revising atom in includes equation below：

$c=\left\{\begin{array}{ccc}{c}_{wake}+g\left(c_{wake}\right|w_{sleep},c_{sleep})\·w_{sleep},c_{sleep},\left|\right|c-c_{sleep}|{|}_2^2>||c_{wake}-c_{sleep}|{|}_2^2& if& type=wake\\ c_{sleep}+g\left(c_{sleep}\right|w_{wake},c_{wake})\·w_{wake},c_{wake},\left|\right|c-c_{wake}|{|}_2^2>||c_{sleep}-c_{wake}|{|}_2^2& if& type=sleep\end{array}\right.$

In formula, c is the position of atom after correction, and g is discriminant function, and w is weighted value.

7. the mask method of sleep state sample data type according to claim 3 is it is characterised in that described weighted value The computing formula of w is as follows：

$w_{wake}=\frac{n_{wake}}{n_{wake}+n_{sleep}},g\left(c_{wake}\right|w_{sleep},c_{sleep})=\left\{\begin{array}{ccc}1& if& w_{sleep}{c}_{sleep}^T(2c_{wake}-c_{sleep})>0\\ -1& if& w_{sleep}{c}_{sleep}^T(2c_{wake}-c_{sleep})<0\end{array}\right.$

$w_{sleep}=\frac{n_{sleep}}{n_{wake}+n_{sleep}},g\left(c_{sleep}\right|w_{wake},c_{wake})=\left\{\begin{array}{ccc}1& if& w_{wake}{c}_{wake}^T(2c_{sleep}-c_{wake})>0\\ -1& if& w_{wake}{c}_{wake}^T(2c_{sleep}-c_{wake})>0\end{array}\right.$

In formula, w_wakeFor the weighted value of clear-headed type, w_sleepWeighted value for sleep pattern.

8. the mask method of sleep state sample data type according to claim 3 is it is characterised in that described weighted value The computing formula of w is as follows：

$w=\left\{\begin{array}{ccc}\frac{0.5n_{sleep}}{n_{wake}+n_{sleep}}& if& type=wake\\ \frac{0.5n_{wake}}{n_{wake}+n_{sleep}}& if& type=sleep\end{array}\right.$

9. sleep state sample data type according to claim 3 mask method it is characterised in that described by this sample The type of notebook data is labeled as the step consistent with the type of this atom and includes：

According to the atomic type of classifying dictionary, identify the sample data of clear-headed type and the sample data of sleep pattern, then Sample data is marked corresponding data type.

10. a kind of labeling system of sleep state sample data type is it is characterised in that include：

Sample collection module, for gathering the EEG signals that user produces in sleep state analysis, obtains sample data；

Dictionary builds module, and the characteristic vector of sample data and characteristic vector for building multiple sleep state types are assembled Cluster center, object function is set up according to described characteristic vector and its cluster center；Wherein, described object function characterizes minimum Change the sample data of same type and the distance of dictionary atom, and maximize the distance between different types of atom；

Dictionary training module, for selecting several characteristic vector conducts respectively from the sample data of multiple sleep state types The initial value of atom, each sample data is distributed to described atom and solves described object function, obtain classifying dictionary；

Sample labeling module, for sample data is inputted classifying dictionary, compare the type of the atom nearest with sample data with Distance, if distance, less than default threshold value, the type of this sample data is labeled as consistent with the type of this atom.