CN109512390A - Sleep stage method and wearable device based on EEG time domain various dimensions feature and M-WSVM - Google Patents
Sleep stage method and wearable device based on EEG time domain various dimensions feature and M-WSVM Download PDFInfo
- Publication number
- CN109512390A CN109512390A CN201811259442.2A CN201811259442A CN109512390A CN 109512390 A CN109512390 A CN 109512390A CN 201811259442 A CN201811259442 A CN 201811259442A CN 109512390 A CN109512390 A CN 109512390A
- Authority
- CN
- China
- Prior art keywords
- signal
- feature
- eeg
- sleep
- time
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/48—Other medical applications
- A61B5/4806—Sleep evaluation
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/369—Electroencephalography [EEG]
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/68—Arrangements of detecting, measuring or recording means, e.g. sensors, in relation to patient
- A61B5/6801—Arrangements of detecting, measuring or recording means, e.g. sensors, in relation to patient specially adapted to be attached to or worn on the body surface
- A61B5/6802—Sensor mounted on worn items
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
Abstract
The invention discloses a kind of sleep stage method and wearable device based on EEG time domain various dimensions feature and M-WSVM, first acquisition EEG continuous time signal, the time domain various dimensions feature for extracting EEG continuous time signal is mapped by amplitude versus time;Then extraction signal characteristic is selected, to obtain the feature of optimal signal;Finally is analyzed and handled using sleep stage of the M-WSVM algorithm to different classifications level, the monitoring of real-time perfoming sleep stage.Device includes signal acquisition module, signal processing module and signal transmission module, it can be communicated in real time with the user terminal of smart machine, by carrying out the model learning of EEG training data at the end PC, the algorithm model of study is transplanted in running on smart machine, to carry out real-time sleep stage monitoring.The present invention simplifies sleep stage complexity using EEG signal feature extraction and classifying method and develops wearable sleep stage device using physiological signal measurements circuit, can obtain real-time, high-precision sleep mode automatically effect by stages.
Description
Technical field
The present invention relates to the sleep mode automatically of intelligent algorithm, field, in particular to one kind are special based on EEG time domain various dimensions by stages
The sleep stage method and wearable device of sign and M-WSVM.
Background technique
With the proposition of national big health strategy and the progress of artificial intelligence technology, intelligent health industry will obtain considerable
Development is of great significance to sleeping disorders diagnosis, treatment and prevention based on EEG Sleep Staging Research.Currently, in EEG signal
Sleep Staging Research in, the application of various signal converter techniques and machine learning method all achieves certain achievement, especially
It is the Sleep Staging Research of single channel EEG signal, it is easier to the monitoring and analysis of sleep, document are realized on wearable device
[1-9] all has studied the sleep stage using single channel EEG signal.But in existing technology, the letter that correlative study utilizes
Number transformation and sorting algorithm have certain complexity mostly, are not suitable for real-time wearable sleep stage system.
EEG is non-stationary signal, and the signal intensity under different sleep states has different, how to utilize effective letter
Number feature is extremely important to the classifying quality for improving intelligent algorithm.It, can by various transformation and the feature extracting method of coding
Relatively good sample of signal feature is obtained, but increases the complexity or feature quantity of feature extraction algorithm.Document [1] utilizes
EEG signal is decomposed into 24 grades of signal subspace band by the method for adjustable Q factor wavelet transformation (TQWT), and is extracted from signal subspace band
The feature of 4 statistical moments;Time domain EEG signal is converted to frequency domain using Fast Fourier Transform (FFT) (FFT) by document [2], and is extracted
The signal characteristic of different frequency band;Raw EEG signal is decomposited the signal wave of a variety of rhythm and pace of moving things by document [3-5], and to each signal
The signal characteristic of wave subband extraction sleep stage;Document [6] is that continuous signal is divided into multiple signal segments, on each signal segment
Extract identical statistical nature.By the analysis to open source sleep EEG signal, the waveform under different sleep states is in time and width
Spending has relatively apparent difference in both direction, show as the variation that different sleep states correspond to different frequency rhythm and pace of moving things wave.Directly
It connects and obtains characterization time domain, the signal difference alienation feature of frequency domain and time-frequency domain using continuous EEG signal, can promote wearable real-time
The exploitation of sleep analysis monitoring system.
According to the general standard of sleep stage, sleep stage is main are as follows: 6 states of Wake, S1, S2, S3, S4 and Rem,
Wherein S1 and S2 is sleeping state, and S3 and S4 are deep sleep states, with increasing sleep state high-subdividing number purpose, to classification
The requirement of algorithm is also higher.Document [7] estimates sleep state using Markov model;It is used in document [8] a kind of quick
The algorithm of the complex values convolutional neural networks of differentiation carries out dormant feature extraction and classifying;Document [9] applies respectively
Decision tree, support vector machines and neural network many algorithms carry out classification experiments.During actual classification algorithm training, by
In data nonbalance, such as certain people's deep sleeps are fewer, result in nicety of grading with biggish difference, simultaneously as portion
Divide sleep state similitude, so that the error in classification between two classes is larger.Utilize the multi-level sleep state point of different classes of requirement
Class model is weighted corresponding dormant mistake classification, to improve system sleep effect by stages.
Due to the faint property of EEG signal, vulnerable to interference, there is very big challenge to effective acquisition of signal, meanwhile, it is advanced
Artificial intelligence technology and method in lightweight mobile device application also have certain deficiency.By designing efficient bottom
Layer signal acquisition module obtains high-precision EEG sleep signal, and the model by carrying out intelligent algorithm at the end PC in real time
It practises, is then transplanted to the mode of lightweight equipment operation, establishes wearable sleep stage system.
Bibliography:
Document [1]: HASSAN, Ahnaf Rashik;SUBASI,Abdulhamit.A decision support
system for automated identification of sleep stages from single-channel EEG
signals.Knowledge-Based Systems,2017,128:115-124.
Document [2]: LIU, Zhiyong, et al.Sleep staging from the EEG signal using
multi-domain feature extraction.Biomedical Signal Processing and Control,
2016,30:86-97.
Document [3]: MEMAR, Pejman;FARADJI,Farhad.A novel multi-class EEG-based
sleep stage classification system.IEEE Transactions on Neural Systems and
Rehabilitation Engineering,2018,26.1:84-95.
Document [4]: ABOALAYON, Khald AI;ALMUHAMMADI,Wafaa S.;FAEZIPOUR,Miad.A
comparison of different machine learning algorithms using single channel EEG
signal for classifying human sleep stages.In:Systems,Applications and
Technology Conference(LISAT),2015IEEE Long Island.IEEE,2015.p.1-6.
Document [5]: ALMUHAMMADI, Wafaa S.;ABOALAYON,Khald AI;FAEZIPOUR,
Miad.Efficient obstructive sleep apnea classification based on EEG
signals.In:Systems,Applications and Technology Conference(LISAT),2015IEEE
Long Island.IEEE,2015.p.1-6.
Document [6]: DIYKH, Mohammed;LI,Yan.Complex networks approach for EEG
signal sleep stages classification.Expert Systems with Applications,2016,63:
241-248.
Document [7]: KANG, Dae Y., et al.A state space and density estimation
framework for sleep staging in obstructive sleep apnea.IEEE Transactions on
Biomedical Engineering,2018,65.6:1201-1212.
Document [8]: ZHANG, Junming;WU,Yan.A new method for automatic sleep stage
classification.IEEE transactions on biomedical circuits and systems,2017,
11.5:1097-1110.
Document [9]:Baha,et al.A comparative study on classification of sleep
stage based on EEG signals using feature selection and classification
algorithms.Journal of medical systems,2014,38.3:18.
Summary of the invention
In order to solve the above-mentioned technical problems, the present invention provides a kind of based on EEG time domain various dimensions feature and M-WSVM
Sleep stage method and wearable device.
Technical solution used by method of the invention is: a kind of sleeping based on EEG time domain various dimensions feature and M-WSVM
It sleeps method by stages, which comprises the following steps:
Step 1: obtaining EEG continuous time signal;
Step 2: extracting the time domain various dimensions feature of EEG continuous time signal;
Step 3: the signal characteristic of extraction being selected, to obtain the feature of optimal signal;
Step 4: being analyzed and handled using sleep stage of the M-WSVM algorithm to different classifications level, real-time perfoming is slept
The monitoring slept by stages.
Preferably, the specific implementation of step 2 includes following sub-step:
Step 2.1: dividing a data sample according to predetermined time period, the EEG continuous time signal that will acquire carries out
Amplitude versus time mapping, mapped by amplitude axis obtain numerical point, numerical point number, numerical point interval characteristic information, and pass through
Time shaft mapping obtains the characteristic information of original signal, peak point and modal point;
Step 2.2: to amplitude axis map information extract the average number of numerical point, numerical point, numerical point equispaced and
Average and standard deviation the sample of signal feature of standard deviation, each numerical point number and average time interval;
To time shaft map information extract the maximum value of original signal, minimum value, mean value, range, activity, mobility,
The sample characteristics of complexity;Extract number, range, mean value, standard deviation, average, interval the standard deviation at interval of peak signal
Sample characteristics;Extract the sample characteristics of the number of mode signal, the average of mode interval, mode separation standard difference;
In the feature of extraction, maximum value, minimum value, range, the characteristics of mean of original signal characterize EEG signal variation
Time domain specification, peak point number, equispaced, standard deviation characteristic characterize the frequency domain characteristic of EEG signal variation, and signal is unique
The time-frequency domain characteristic of numerical point number, mean value, equispaced characteristic present EEG signal variation;
Step 2.3: the signal characteristic of the reflection time domain of extraction, frequency domain and time-frequency domain different characteristics being combined, is passed through
In the way of different time domain various dimensions feature mixing calculating, such as the characterization temporal signatures that amplitude axis maps are reflected with time shaft
The characterization frequency domain character penetrated ask the operation of poor, ratio etc., so that the sample characteristics of characterization time-frequency domain overall characteristic are extracted,
In include that signal numerical point number and the ratio of range of signal, the mean value of non-peak point signal, signal numerical value mode and signal are put down
The difference of mean value.
Preferably, the specific implementation of step 3 includes following sub-step:
Step 3.1: according to extracting and selecting Optimal Signals feature, establishing the data set of EEG signal sample characteristics, and calculate
The core of signal characteristic;
Step 3.2: judging whether there is reduction feature core;
If it is not, then directly executing following step 3.3;
If so, the sample of signal feature of output reduction, establishes new signal characteristic sample data set;
Step 3.3: calculate sample characteristics class spacing and;
Step 3.4: calculate sample characteristics inter- object distance standard deviation and;
Step 3.5: judging whether to traverse all features;
If so, executing following step 3.6;
Step 3.3 is executed if it is not, then turning round;
Step 3.6: calculating Pawlak Attribute Significance;
Step 3.7: in conjunction with the evaluation criterion A of the Pawlak Attribute Significance construction feature classification capacity of each featurem;
Wherein, sigmFor the Pawlak Attribute Significance of sample characteristics m, reflected sample feature divides conceptual data sample
Class ability;N is the sleep state number of classification, μiFor the center of the i-th class sample, ωiFor the dormant set of data samples of the i-th class,
The fraction part of evaluation criterion indicate m-th of feature to data sample between class distance and with all kinds of interior centre distance standard deviation summations
Ratio;
Step 3.8: according to AmValue size is ranked up to feature is extracted, using the method testing classification for being stepped up feature
Ability, and select optimal signal characteristic.
Preferably, the specific implementation process of step 4 is: being basic sleep state point with Weighted Support Vector WSVM
Class algorithm, and multi-level sorting algorithm is constructed according to multiple sleep states of monitoring;
In 5 class sleep state Sleep and 1 class waking state Wake totally 6 classification tasks, it is designed as point of M=1~4 layer
Class algorithm, has been utilized respectively the similitude of sleeping state S1, S2 and rapid eye movement state Rem, and with deep sleep state S3,
The otherness of S4;
According to different dormant similitudes and otherness, first with the signal width of Wake state and Sleep state
The difference characteristic of degree and frequency variation, classifies to Wake state and Sleep state in first layer, wherein Sleep state packet
Entire sleep state and waking state are carried out two classification, with acquisition pair by 5 kinds of sleep states for including S1, S2, S3, S4 and Rem
The high-precision classification of Wake state;The second layer is the similitude according to signal, to the sleep state of S1, S2, Rem and S3, S4 into
Row two is classified, since the sleeping state and Rem state EEG signal of S1 and S2 have similitude, so combining two states
Whole sleeping state realizes the high-precision classification to shallow sleep and deep sleep state;Third layer further binding signal
Otherness and similitude, classify respectively to the sleep state of S1, S2 and Rem and S3 and S4, realize to Rem and deep sleep
The classification task of S3, S4 state;4th layer is finally to classify to the sleeping state of S1 and S2, corresponding shallow to detect
Sleep state.
Preferably, in step 4, in the sleep sorting algorithm of each layer, according to the injustice of each sleep state data sample
The demand of weighing apparatus and practical sleep monitor, is weighted processing to corresponding data sample, to adjust system entirety sleep stage essence
Degree.
Preferably, using genetic algorithm optimization parameter, establishing the WSVM algorithm for meeting nicety of grading requirement in step 4
Model;Specific implementation includes following sub-step:
Step 4.1: by acquiring dormant EEG continuous time signal, and according to the optimal time domain extracted and selected
Various dimensions signal characteristic establishes the experimental data set for being used for sleep state classification.
Step 4.2: penalty factor parameter c, nuclear parameter g and weighting processing parameter w in setting WSVM algorithm model, and it is right
The disaggregated model of M-WSVM under different levels is trained, wherein M=1~4;
Step 4.3: if the algorithm model of training meets the required precision of test, the training pattern of output algorithm is no
Parameter optimization then is carried out to w, c and g using genetic algorithm, until training pattern meets measuring accuracy requirement.
Technical solution used by the device of the invention is: a kind of sleeping based on EEG time domain various dimensions feature and M-WSVM
It sleeps wearable device by stages, it is characterised in that: including signal acquisition module, signal processing module and signal transmission module;
The signal acquisition module, for carrying out the acquisition of single channel EEG sleep signal;
The signal processing module, for the acquisition mode, frequency, data format of setting signal, and according to the letter of extraction
Number sample characteristics, the edge calculations and information optimized in terminal to acquisition signal are transmitted;
The signal transmission module utilizes BLE wireless transmission method, real-time Transmission EEG signal data;
The wearable device for sleep stage can in real time and smart device communication, by carrying out EEG at the end PC
The model learning of training data transplants learning model in running on smart machine, to carry out real-time sleep stage monitoring.
It is slept preferably, the acquisition module carries out single channel EEG using high-precision physiological signal measurements ADS1299 chip
Dormancy signal acquisition;By microcontroller design signal processing module, with the acquisition mode, frequency, data format etc. of setting signal,
And according to the sample of signal feature of extraction, in the edge calculations and information transmission that terminal optimizes acquisition signal;Transmit mould
Block is carried out data transmission with the communication of low-power consumption bluetooth (BLE), according to the information requirement and communication format of EEG signal
Requirement, the sleep stage device of flexible design and the communication protocol of user terminal real-time and efficiently transmit EEG signal characteristic
According to.
Signal processing module is analyzed and processed original EEG signal using the method for edge calculations, according to signal
Feature extraction and selection directly obtains corresponding signal characteristic data in signal processing module, selects to original signal and spy
Sign data are transmitted;Signal transmission module carries out the processing of data, including data-conversion, displacement, xor operation in bottom,
To improve the efficiency of user terminal model learning;The signal data of transmission is designed according to communication byte, data format simultaneously,
Guarantee the precision of data transmission.
Compared with prior art, innovation and advantage possessed by the present invention are as follows:
1, the present invention is by the analysis and research to single channel EEG signal, only using continuous time signal obtain characterization time domain,
Frequency domain and time-frequency domain and combined various dimensions feature are not required to any signal transformation and coding techniques, with feature quantity is few, is easy to
The advantages of extraction, while being suitable for the intelligent classification algorithm of wearable sleep stage;
2, the present invention carries out sleep stage monitoring using the method for multilayer subseries, according to the phase of sleep state EEG signal
Successively classified like property with otherness, every layer is identified the sleep state for being easy to classify accordingly, meanwhile, it is capable to according to respectively sleeping
The data sample imbalance of dormancy state is weighted processing to respective classes, using the small sample classification advantage of WSVM algorithm, if
Multi-level sleep stage algorithm is counted, whole higher nicety of grading can not only be obtained, also adjustable partial sleep state
Nicety of grading;
3, sleep stage device of the invention acquires EEG signal using high-precision physiological signal measurements chip ADS1299, and
According to the sample characteristics for extracting signal, the characteristics of high-performance calculation real-time using bottom MCU, edge calculations are carried out to acquisition data
Optimization processing, in combination with the communication mode of BLE, efficiently to transmit sleep EEG data, in smart phone, raspberry pie etc.
The verifying and interpretation of result of user terminal progress model.Sleep stage device design runs various intelligence learnings in user terminal
Algorithm can extend multiple sleep stage device connection user terminals, have good practical performance.
Detailed description of the invention
Fig. 1 is that the amplitude axis of the mapping of EEG signal amplitude versus time described in the embodiment of the present invention maps schematic diagram;
Fig. 2 is that the time shaft of the mapping of EEG signal amplitude versus time described in the embodiment of the present invention maps schematic diagram;
Fig. 3 is the schematic diagram of EEG signal time domain various dimensions feature extraction described in the embodiment of the present invention;
Fig. 4 is that the embodiment of the present invention is selected using the dispersion of RST algorithm and signal characteristic signal characteristic is extracted
Flow chart;
Fig. 5 is M-WSVM of embodiment of the present invention sleep state classification schematic diagram;
Fig. 6 is the WSVM algorithm model learning process figure that the embodiment of the present invention utilizes genetic algorithm optimization parameter;
Fig. 7 is the structure and functional schematic that the EEG signal of sleep stage of embodiment of the present invention device is obtained and analyzed;
Fig. 8 is the overall structure diagram of sleep stage method and device integrated application described in the embodiment of the present invention.
Specific embodiment
Understand for the ease of those of ordinary skill in the art and implement the present invention, with reference to the accompanying drawings and embodiments to this hair
It is bright to be described in further detail, it should be understood that implementation example described herein is merely to illustrate and explain the present invention, not
For limiting the present invention.
The feature extraction of EEG signal time domain various dimensions carries out amplitude versus time mapping firstly the need of to signal in the present embodiment, with
Relevant characteristic information is obtained, specific embodiment is that the continuous EEG signal of acquisition is mapped to amplitude axis and time shaft, is such as schemed
Shown in 1, each numerical point in sample of signal is mapped to amplitude axis, and calculate unique numeric appeared in sample of signal
Point, the number of numerical point, numerical point equispaced, for example after the mapping of 3000 sample points, 1000 are obtained in amplitude axis
Unique numerical point, the equispaced 5 of 20,20 points of number of numerical point 12;The time shaft for being illustrated in figure 2 sample of signal reflects
It penetrates, wherein original signal to be mapped to the one-dimension array for obtaining sample of signal point, peak point, which is carried out mapping, to be obtained
Each peak point one-dimension array in sample of signal, further the modal point of sample of signal is mapped, is obtained between modal point
Every information.
Fig. 3 is that the present invention maps the implementation for extracting the time domain various dimensions feature of EEG continuous time signal by amplitude versus time
Example, can be with the continuous EEG signal that will acquire first carries out amplitude versus time mapping, is mapped by amplitude axis and obtains numerical value by diagram
Point, numerical point number, numerical point interval characteristic information, and pass through time shaft mapping and obtain original signal, peak point and mode
Then the characteristic information of point carries out maximum value, minimum value, range, number, mean value and standard deviation to each mappings characteristics information respectively
Feature extraction.Feature of present invention extracts the number for obtaining numerical point, and numerical point is average, numerical point equispaced and standard
The average and standard deviation of difference, each numerical point number and average time interval, the maximum value of original signal, minimum value, mean value, model
It encloses, activity, mobility, complexity, number, range, mean value, standard deviation, average, interval the standard at interval of peak point
Difference, the number of modal point, the average of mode interval, mode interval standard deviation equal samples feature.
In the feature of extraction, the EEG signal change of the characteristic presents such as maximum value, minimum value, range, mean value of original signal
The time domain specification of change, the frequency domain characteristic of the characteristic presents such as peak point number, equispaced, standard deviation EEG signal variation, signal
The time-frequency domain characteristic of the characteristic presents such as unique numeric point number, mean value, equispaced EEG signal variation.Further, with group
The mode for closing feature extracts the signal characteristic of characterization time-frequency domain characteristic, including the ratio of signal numerical point number and range of signal
The sample of signal feature such as value, the mean value of non-peak point signal, the difference of signal numerical value mode and signal averaging.
In order to remove the characteristic information of redundancy, the complexity of sorting algorithm model is reduced, the present invention is to extraction signal characteristic
It is selected, to obtain the feature of optimal signal.The present invention selects optimal use to the time domain various dimensions signal characteristic of extraction
In sleep stage sample characteristics embodiment as shown in figure 4, using RST algorithm carry out signal characteristic reduction, first calculate letter
The core of number feature will add new non-core feature according to Pawlak Attribute Significance and obtain to sample spy if there is feature core
One reduction of sign, if there is no feature core, by calculate each sample characteristics between class distance and with inter- object distance standard deviation
With, further in conjunction with each feature Pawlak Attribute Significance construction feature classification capacity evaluation criterion:
Wherein, sigmFor the Pawlak Attribute Significance of sample characteristics m, reflected sample feature divides conceptual data sample
Class ability;N is the sleep state number of classification, μiFor the center of the i-th class sample, ωiFor the dormant set of data samples of the i-th class,
The fraction part of evaluation criterion indicate m-th of feature to data sample between class distance and with all kinds of interior centre distance standard deviation summations
Ratio;
Between class distance in the present invention and be Different categories of samples mean value mutual distance and, rather than Different categories of samples mean value and complete
The distance of portion's sample average and, can be avoided partial category sample in this way far from influence caused by overall data mean value;The present invention
Inter- object distance standard deviation is utilized and to evaluate sample characteristics to the dispersion of Various types of data sample, the smaller then table of criterion distance difference
Be illustrated feature class in sample clustering effect it is better, further, between class distance and with inter- object distance standard deviation and ratio
The feature point that signal characteristic is reflected to overall sample classification ability, and combines the Attribute Significance building of sample characteristics comprehensive
Class merit rating standard, and then select optimal EEG sleep state classification feature.
In the present invention using M-WSVM algorithm establish each sleep state classification model in sleep stage implementation such as Fig. 5 and
Shown in Fig. 6, according to different dormant similitudes and otherness, first with the signal width of Wake state and Sleep state
The difference characteristic of degree and frequency variation, classifies to Wake state and Sleep state in first layer, wherein Sleep state packet
Entire sleep state and waking state are carried out two classification, with acquisition pair by 5 kinds of sleep states for including S1, S2, S3, S4 and Rem
The high-precision classification of Wake state;The second layer is the similitude according to signal, to the sleep state of S1, S2, Rem and S3, S4 into
Row two is classified, since the sleeping state and Rem state EEG signal of S1 and S2 have similitude, so combining two states
For whole sleeping state, the high-precision classification to shallow sleep and deep sleep state is realized;Third layer further combines letter
Number otherness and similitude, classify respectively to the sleep state of S1, S2 and Rem and S3 and S4, realize to Rem and sound sleep
The classification task of dormancy S3, S4 state;4th layer is finally to classify to the sleeping state of S1 and S2, corresponding to detect
Sleeping state.Multi-level sleep state classification of the invention fully considers EEG sleep state signals characteristic, is utilized respectively difference
Status signal otherness and similitude are successively classified, and are respectively completed corresponding classification task in each layer, more with routine
Classification method is compared, the algorithm complexity not only reduced, but also avoids influence of the similarity signal state to whole nicety of grading.
In multi-level disaggregated model, the present invention carries out dormant two classification of each level with WSVM algorithm, further
Ground, since there are imbalance problems for the dormant data sample of EEG, the deep sleep of especially certain people is seldom, even
The sample of signal of deep sleep may be lacked, in order to overcome data nonbalance that error in classification is caused to be partial to Small Sample Database
And the problem that overall precision is low, the present invention by utilize WSVM algorithm, in conjunction with actual data sample size to classification task into
Row weighting processing, to improve whole nicety of grading.Meanwhile the present invention is excellent to model parameter w, c and g progress using genetic algorithm
Change, establishes the WSVM algorithm model for meeting nicety of grading requirement.
See Fig. 6, the present embodiment utilizes genetic algorithm optimization parameter, establishes the WSVM algorithm mould for meeting nicety of grading requirement
Type;Specific implementation includes following sub-step:
Step 4.1: by acquiring dormant EEG continuous time signal, and according to the optimal time domain extracted and selected
Various dimensions signal characteristic establishes the experimental data set for being used for sleep state classification.
Step 4.2: penalty factor parameter c, nuclear parameter g and weighting processing parameter w in setting WSVM algorithm model, and it is right
The disaggregated model of M-WSVM under different levels is trained, wherein M=1~4;
Step 4.3: if the algorithm model of training meets the required precision of test, the training pattern of output algorithm is no
Parameter optimization then is carried out to w, c and g using genetic algorithm, until training pattern meets measuring accuracy requirement.
Wearable device embodiment provided by the present invention for sleep stage includes that sleep EEG signal obtains and sleeps
EEG signal analyzes two parts.As shown in fig. 7, wherein signal acquisition is the wearable EEG signal acquisition device of user, including
Acquisition module, processing module and the transmission module of sleep EEG signal;The analysis of signal includes data processing and mould at the end PC
Type study and the real-time sleep stage monitoring of user terminal.The signal acquisition process of specific embodiment is using MCU controller, by setting
The operating mode, acquisition channel and frequency for setting ADS1299 realize the behaviour such as the amplification of signal, filtering, biasing feedback and A/D conversion
Make, while the feature high using the real-time calculated performance of MCU controller bottom, carries out the signal-data processing of edge calculations;This hair
A bright considerable advantage is that time-domain signal data are only utilized, and can carry out real-time amplitude versus time mapping to signal acquisition,
The data sample of signal characteristic is directly extracted in selection according to the end PC to data characteristics using the edge data processing of microcontroller
This, the data of such as one sample point of every acquisition store array in the way of magnitude map, pass through the analysis and calculating to array
And then each signal characteristic data are obtained, which is highly suitable for the real-time application of wearable device.Further, BLE is utilized
Mode carry out the transmission of characteristic, and carry out algorithm model in the user terminal of the smart machines such as smart phone or raspberry pie
Real time monitoring and analyzing.
EEG sleep stage method and wearable device of the present invention based on time domain various dimensions feature and M-WSVM
Integrated application is as shown in Figure 8.Different from other sleep stage methods, sleep stage method provided by the present invention is from practical application
From the point of view of, in conjunction with specific operable sleep stage device, EEG signal edge data is described in detail to the acquisition of signal
Processing, can acquire, handle in real time and transmission sleep state feature data sample;To the mapping of the amplitude versus time of signal and
It extracts time domain various dimensions feature to be summarized, and uses RST algorithm and carry out comprehensive feature selecting with feature dispersion, wherein
It is utilized respectively the mutual distance of Different categories of samples mean value and with inter- object distance standard deviation and to characterize between class and within-cluster variance;According to
Actual sleep stage mission requirements, using the algorithm model of multi-level sleep state classification, and according to the difference of sleep EEG signal
The sleep state classification model of anisotropic and similitude design different levels.Further, consider sleep state data nonbalance
Problem carries out dormant two classification of each level using the algorithm model of WSVM, meanwhile, genetic algorithm is utilized to algorithm
Model parameter and weighting weight parameter carry out optimizing, to improve the classification capacity of EEG sleep state entirety.
The time domain various dimensions feature extraction of application of the present invention towards wearable sleep stage, the sleep stage method is
In order to obtain optimal signal characteristic, by the parsing to original signal, characterization various dimensions letter directly in extraction time domain EEG signal
The sample characteristics of number characteristic.It is converted without signal or the feature extraction of the processing such as coding simplifies sleep stage method complexity,
And it can be improved the ability of real-time application.Single channel EEG sleep signal is subjected to amplitude versus time mapping first, in amplitude axis
Mapping Resolution obtains three groups of feature letters of average time interval of each numerical point of sample of signal, the number of each numerical point and each numerical point
Breath, and the characteristic information of original signal, peak signal and modal point signal is mapped on a timeline.Further, special to mapping
The signal characteristic of information extraction characterization time domain and frequency domain is levied, meanwhile, complex amplitude axis and time shaft characteristic extract characterization time-frequency domain
The feature of signal and relevant assemblage characteristic, the classifying quality of algorithm is advanced optimized by assemblage characteristic.It is right on this basis
The classification capacity for extracting EEG signal feature is analyzed, and the core and Attribute Significance of signal characteristic are calculated using RST algorithm, right
Each signal characteristic calculates centre distance and all kinds of summations of the interior center away from standard deviation in each sleep state class, and building signal is special
The evaluation criterion of sign classification capacity optimizes the signal characteristic of practical application to carry out the reduction and selection of feature.
Sleep stage method of the invention is designed as carrying out a variety of dormant monitorings and classification;The present invention is constructed from 2
Class to 6 classes multi-level sleep state classification, using WSVM algorithm as two basic classification methods, to every level according to signal phase
Like the imbalance problem of property and data, processing is weighted to corresponding classification, comprehensive utilization M-WSVM sorting algorithm is realized more
Efficient sleep stage overcomes some algorithms to only focus on total nicety of grading and make the nicety of grading of partial sleep state not high
The problem of, while can be according to the weighting in practical application to sleep state detection, the weighting coefficient of adjustment algorithm.Further
Ground carries out the parameter optimization of M-WSVM using self-adapted genetic algorithm according to the demand of practical sleep stage target, including basic
Algorithm model parameter and different sleep stages weight parameter.
Wearable device provided by the present invention for sleep stage measures chip using the EEG signal of 24 bit resolutions
ADS1299 modelled signal acquisition module;By utilizing MCU modelled signal processing module, to carry out the control of data acquisition and transmission
System;BLE wireless transmission method, real-time Transmission EEG signal data are utilized in signal transmission module.The sleep stage device can
It is communicated in real time with lightweights equipment such as smart phone, raspberry pies, it, will by carrying out the model learning of EEG training data at the end PC
Learning model transplanting is run in lightweight equipment, to carry out real-time sleep stage monitoring.The present invention is counted in real time using bottom
The high feature of performance is calculated, EEG data is handled according to feature is extracted, the advantage that jointing edge calculates, in signal processing mould
The operations such as block stored, parsed to acquisition signal, feature extraction, optimization calculate, and it is real-time transmitted to user terminal, to carry out
The classification and correlation analysis of sleep stage.
It should be understood that the part that this specification does not elaborate belongs to the prior art.
It should be understood that the above-mentioned description for preferred embodiment is more detailed, can not therefore be considered to this
The limitation of invention patent protection range, those skilled in the art under the inspiration of the present invention, are not departing from power of the present invention
Benefit requires to make replacement or deformation under protected ambit, fall within the scope of protection of the present invention, this hair
It is bright range is claimed to be determined by the appended claims.
Claims (8)
1. a kind of sleep stage method based on EEG time domain various dimensions feature and M-WSVM, which is characterized in that including following step
It is rapid:
Step 1: obtaining EEG continuous time signal;
Step 2: extracting the time domain various dimensions feature of EEG continuous time signal;
Step 3: selecting signal characteristic is extracted, to obtain the feature of optimal signal;
Step 4: it is analyzed and is handled using sleep stage of the M-WSVM algorithm to different classifications level, real-time perfoming sleep point
The monitoring of phase.
2. the sleep stage method according to claim 1 based on EEG time domain various dimensions feature and M-WSVM, feature exist
In the specific implementation of step 2 includes following sub-step:
Step 2.1: dividing a data sample according to predetermined time period, the EEG continuous time signal that will acquire carries out amplitude-
Time map, mapped by amplitude axis obtain numerical point, numerical point number, numerical point interval characteristic information, and pass through the time
Axis mapping obtains the characteristic information of original signal, peak point and modal point;
Step 2.2: the average number of numerical point, numerical point, numerical point equispaced and standard are extracted to amplitude axis map information
Average and standard deviation the sample of signal feature of poor, each numerical point number and average time interval;
Maximum value, minimum value, mean value, range, activity, mobility, the complexity of original signal are extracted to time shaft map information
The sample characteristics of property;Extract the number of peak signal, the sample of range, mean value, standard deviation, average, interval the standard deviation at interval
Eigen;Extract the sample characteristics of the number of mode signal, the average of mode interval, mode separation standard difference;
In the feature of extraction, maximum value, minimum value, range, the characteristics of mean of original signal characterize EEG signal variation when
Domain characteristic, peak point number, equispaced, standard deviation characteristic characterize the frequency domain characteristic that EEG signal changes, signal unique numeric
Put the time-frequency domain characteristic of number, mean value, equispaced characteristic present EEG signal variation;
Step 2.3: the signal characteristic of the characterization time domain of extraction, frequency domain and time-frequency domain different characteristics being combined, utilization is passed through
The mode that different time domain various dimensions feature mixing calculates, so that the sample characteristics of characterization time-frequency domain overall characteristic are extracted, wherein
Mean value, signal numerical value mode and the signal averaging of ratio, non-peak point signal including signal numerical point number and range of signal
The difference of value.
3. the sleep stage method according to claim 1 based on EEG time domain various dimensions feature and M-WSVM, feature exist
In the specific implementation of step 3 includes following sub-step:
Step 3.1: according to the Optimal Signals feature extracted and selected, establishing the data set of EEG signal sample characteristics, and calculate letter
The core of number feature;
Step 3.2: judging whether there is reduction feature core;
If it is not, then directly executing following step 3.3;
If so, the sample of signal feature of output reduction, establishes new signal characteristic sample data set;
Step 3.3: calculate sample characteristics class spacing and;
Step 3.4: calculate sample characteristics inter- object distance standard deviation and;
Step 3.5: judging whether to traverse all features;
If so, executing following step 3.6;
Step 3.3 is executed if it is not, then turning round;
Step 3.6: calculating Pawlak Attribute Significance;
Step 3.7: in conjunction with the evaluation criterion A of the Pawlak Attribute Significance construction feature classification capacity of each featurem;
Wherein, sigmFor the Pawlak Attribute Significance of sample characteristics m, classification energy of the reflected sample feature to conceptual data sample
Power;N is the sleep state number of classification, μiFor the center of the i-th class sample, ωiFor the dormant set of data samples of the i-th class, evaluation
The fraction part of standard indicate m-th of feature to data sample between class distance and and all kinds of interior centre distance standard deviation summations ratio
Value;
Step 3.8: according to AmValue size is ranked up to feature is extracted, using the method testing classification ability for being stepped up feature,
And select optimal signal characteristic.
4. the sleep stage method according to claim 1 based on EEG time domain various dimensions feature and M-WSVM, feature exist
In the specific implementation process of step 4 is: with Weighted Support Vector WSVM for basic sleep state classification algorithm, and according to
Multiple sleep states of monitoring construct multi-level sorting algorithm;
In 5 class sleep state Sleep and 1 class waking state Wake totally 6 classification tasks, the classification for being designed as M=1~4 layer is calculated
Method, has been utilized respectively the similitude of sleeping state S1, S2 and rapid eye movement state Rem, and with deep sleep state S3, S4's
Otherness;
According to different dormant similitudes and otherness, first with the signal amplitude of Wake state and Sleep state with
Frequency variation difference characteristic, classify in first layer to Wake state and Sleep state, wherein Sleep state include S1,
Entire sleep state and waking state are carried out two classification, to obtain to Wake shape by 5 kinds of sleep states of S2, S3, S4 and Rem
The high-precision classification of state;The second layer is the similitude according to signal, carries out two points to the sleep state of S1, S2, Rem and S3, S4
Class, since the sleeping state and Rem state EEG signal of S1 and S2 have similitude, so by two states composite entity
Sleeping state realizes the high-precision classification to shallow sleep and deep sleep state;The difference of third layer further binding signal
Property and similitude, classify respectively to the sleep state of S1, S2 and Rem and S3 and S4, realize to Rem and deep sleep S3, S4
The classification task of state;4th layer is finally to classify to the sleeping state of S1 and S2, to detect corresponding shallow sleep
State.
5. the sleep stage method according to claim 1 based on EEG time domain various dimensions feature and M-WSVM, feature exist
In: in step 4, in the sleep sorting algorithm of each layer, according to the imbalance of each sleep state data sample and practical sleep prison
The demand of survey is weighted processing to corresponding data sample, to adjust system entirety sleep stage precision.
6. the sleep stage side described in -5 any one based on EEG time domain various dimensions feature and M-WSVM according to claim 1
Method, it is characterised in that: in step 4, using genetic algorithm optimization parameter, establish each WSVM algorithm mould for meeting nicety of grading requirement
Type;Specific implementation includes following sub-step:
Step 4.1: by acquiring dormant EEG continuous time signal, and according to the optimal time domain multidimensional extracted and selected
Signal characteristic is spent, the experimental data set for being used for sleep state classification is established;
Step 4.2: penalty factor parameter c, nuclear parameter g and weighting processing parameter w in setting WSVM algorithm model, and to difference
The disaggregated model of M-WSVM under level is trained, wherein M=1~4;
Step 4.3: if the algorithm model of training meets the required precision of test, the training pattern of output algorithm is otherwise sharp
Parameter optimization is carried out to w, c and g with genetic algorithm, until training pattern meets measuring accuracy requirement.
7. a kind of sleep stage wearable device based on EEG time domain various dimensions feature and M-WSVM, it is characterised in that: including letter
Number acquisition module, signal processing module and signal transmission module;
The signal acquisition module, for carrying out the acquisition of single channel EEG sleep signal;
The signal processing module, for the acquisition mode, frequency, data format of setting signal, and according to the signal sample of extraction
Eigen, in the edge calculations and information transmission that terminal optimizes acquisition signal;
The signal transmission module utilizes BLE wireless transmission method, real-time Transmission EEG signal data;
The wearable device for sleep stage can in real time and smart device communication, by carrying out EEG training at the end PC
The model learning of data transplants learning model in running on smart machine, to carry out real-time sleep stage monitoring.
8. the sleep stage wearable device according to claim 7 based on EEG time domain various dimensions feature and M-WSVM,
Be characterized in that: the signal processing module is analyzed and processed original EEG signal using the method for edge calculations, according to letter
Number feature extraction and selection, directly obtain corresponding signal characteristic data in signal processing module, select to original signal
It is transmitted with characteristic;The signal transmission module carries out the processing of data in bottom, including data-conversion, displacement, different
Or operation, to improve the efficiency of user terminal model learning;Simultaneously to the signal data of transmission according to communication byte, data format into
Row design, guarantees the precision of data transmission.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811259442.2A CN109512390B (en) | 2018-10-26 | 2018-10-26 | Sleep staging method and wearable device based on EEG time domain multi-dimensional features and M-WSVM |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811259442.2A CN109512390B (en) | 2018-10-26 | 2018-10-26 | Sleep staging method and wearable device based on EEG time domain multi-dimensional features and M-WSVM |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109512390A true CN109512390A (en) | 2019-03-26 |
CN109512390B CN109512390B (en) | 2020-10-13 |
Family
ID=65772505
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811259442.2A Active CN109512390B (en) | 2018-10-26 | 2018-10-26 | Sleep staging method and wearable device based on EEG time domain multi-dimensional features and M-WSVM |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109512390B (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111012132A (en) * | 2019-12-17 | 2020-04-17 | 珠海格力电器股份有限公司 | Sleep state adjusting method and device based on pillow and intelligent pillow |
CN111059687A (en) * | 2019-12-17 | 2020-04-24 | 珠海格力电器股份有限公司 | Control method and device of air conditioner and storage medium |
CN112842279A (en) * | 2021-03-01 | 2021-05-28 | 中山大学 | Sleep quality evaluation method and device based on multi-dimensional characteristic parameters |
CN113080864A (en) * | 2021-04-07 | 2021-07-09 | 电子科技大学 | Common sleep disease detection method through automatic sleep staging results |
CN113907709A (en) * | 2021-09-28 | 2022-01-11 | 西安电子科技大学 | Portable sleep monitoring system based on ear EEG |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2009149126A2 (en) * | 2008-06-02 | 2009-12-10 | New York University | Method, system, and computer-accessible medium for classification of at least one ictal state |
JP2011083393A (en) * | 2009-10-14 | 2011-04-28 | Osaka Bioscience Institute | Apparatus and method for automatically identifying sleep stage, and computer program for the same |
CN102254227A (en) * | 2011-07-06 | 2011-11-23 | 无锡泛太科技有限公司 | Rough set-based Back Propagation (BP) neural network fusion method of multiple sensors of Internet of things |
US20130253362A1 (en) * | 2008-04-15 | 2013-09-26 | Christopher Scheib | Method and system for monitoring and displaying physiological conditions |
CN107440716A (en) * | 2017-07-26 | 2017-12-08 | 电子科技大学 | Human body lower limbs athletic performance classification discrimination method based on single channel electromyographic signal |
CN107788976A (en) * | 2017-09-22 | 2018-03-13 | 复旦大学 | Sleep monitor system based on Amplitude integrated electroencephalogram |
CN108143409A (en) * | 2016-12-06 | 2018-06-12 | 中国移动通信有限公司研究院 | Sleep stage method and device by stages |
-
2018
- 2018-10-26 CN CN201811259442.2A patent/CN109512390B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20130253362A1 (en) * | 2008-04-15 | 2013-09-26 | Christopher Scheib | Method and system for monitoring and displaying physiological conditions |
WO2009149126A2 (en) * | 2008-06-02 | 2009-12-10 | New York University | Method, system, and computer-accessible medium for classification of at least one ictal state |
JP2011083393A (en) * | 2009-10-14 | 2011-04-28 | Osaka Bioscience Institute | Apparatus and method for automatically identifying sleep stage, and computer program for the same |
CN102254227A (en) * | 2011-07-06 | 2011-11-23 | 无锡泛太科技有限公司 | Rough set-based Back Propagation (BP) neural network fusion method of multiple sensors of Internet of things |
CN108143409A (en) * | 2016-12-06 | 2018-06-12 | 中国移动通信有限公司研究院 | Sleep stage method and device by stages |
CN107440716A (en) * | 2017-07-26 | 2017-12-08 | 电子科技大学 | Human body lower limbs athletic performance classification discrimination method based on single channel electromyographic signal |
CN107788976A (en) * | 2017-09-22 | 2018-03-13 | 复旦大学 | Sleep monitor system based on Amplitude integrated electroencephalogram |
Non-Patent Citations (1)
Title |
---|
EVANGELIA I. ZACHARAKI ET AL: "Spike pattern recognition by supervised classification in low dimensional embedding space", 《BRAIN INFORMATICS》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111012132A (en) * | 2019-12-17 | 2020-04-17 | 珠海格力电器股份有限公司 | Sleep state adjusting method and device based on pillow and intelligent pillow |
CN111059687A (en) * | 2019-12-17 | 2020-04-24 | 珠海格力电器股份有限公司 | Control method and device of air conditioner and storage medium |
CN112842279A (en) * | 2021-03-01 | 2021-05-28 | 中山大学 | Sleep quality evaluation method and device based on multi-dimensional characteristic parameters |
CN112842279B (en) * | 2021-03-01 | 2022-03-08 | 中山大学 | Sleep quality evaluation method and device based on multi-dimensional characteristic parameters |
CN113080864A (en) * | 2021-04-07 | 2021-07-09 | 电子科技大学 | Common sleep disease detection method through automatic sleep staging results |
CN113907709A (en) * | 2021-09-28 | 2022-01-11 | 西安电子科技大学 | Portable sleep monitoring system based on ear EEG |
CN113907709B (en) * | 2021-09-28 | 2023-05-02 | 西安电子科技大学 | Portable sleep monitoring system based on ear EEG |
Also Published As
Publication number | Publication date |
---|---|
CN109512390B (en) | 2020-10-13 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109512390A (en) | Sleep stage method and wearable device based on EEG time domain various dimensions feature and M-WSVM | |
WO2021160189A1 (en) | Csi method for recognizing human fall in wi-fi interference environment | |
CN106407905B (en) | Wireless aware action identification method based on machine learning | |
CN107122790B (en) | Non-invasive load identification algorithm based on hybrid neural network and ensemble learning | |
CN108629380B (en) | Cross-scene wireless signal sensing method based on transfer learning | |
CN105841961A (en) | Bearing fault diagnosis method based on Morlet wavelet transformation and convolutional neural network | |
CN103294199B (en) | A kind of unvoiced information identifying system based on face's muscle signals | |
CN106658590A (en) | Design and implementation of multi-person indoor environment state monitoring system based on WiFi channel state information | |
CN111062250A (en) | Multi-subject motor imagery electroencephalogram signal identification method based on depth feature learning | |
CN111476158B (en) | Multi-channel physiological signal somatosensory gesture recognition method based on PSO-PCA-SVM | |
CN105786189B (en) | A kind of self contained function recognition methods of finger portion and system that signal is moved based on flesh | |
CN103092971B (en) | A kind of sorting technique for brain-computer interface | |
CN111597991A (en) | Rehabilitation detection method based on channel state information and BilSTM-Attention | |
CN108363493A (en) | User characteristics method for establishing model, system and storage medium based on brain-computer interface | |
CN108694953A (en) | A kind of chirping of birds automatic identifying method based on Mel sub-band parameter features | |
CN107273841A (en) | A kind of electric sensibility classification method of the brain based on EMD and gaussian kernel function SVM | |
CN107766898A (en) | The three classification mood probabilistic determination methods based on SVM | |
CN114781463A (en) | Cross-scene robust indoor tumble wireless detection method and related equipment | |
CN114973330A (en) | Cross-scene robust personnel fatigue state wireless detection method and related equipment | |
CN109766847A (en) | A kind of spike separation method | |
CN113076878B (en) | Constitution identification method based on attention mechanism convolution network structure | |
Yang et al. | A simple deep learning method for neuronal spike sorting | |
CN112001305B (en) | Feature optimization SSVEP asynchronous recognition method based on gradient lifting decision tree | |
CN109886097A (en) | A kind of method for detecting fatigue driving based on artificial fish school optimization H-ELM | |
Ramakrishnan et al. | Epileptic eeg signal classification using multi-class convolutional neural network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |