CN103735262B - Dual-tree complex wavelet and common spatial pattern combined electroencephalogram characteristic extraction method - Google Patents

Abstract

The present invention relates to a kind of dual-tree complex wavelet feature extracting methods that spatial model combines together. The present invention chooses the EEG signals in appropriate channel first, then according to dual-tree complex wavelet frequency segmentation the characteristics of, to original frequency carry out up-sampling or down-sampling, then utilize dual-tree complex wavelet multi-resolution decomposition, to obtain , , With The corresponding frequency range of frequency range of four species rhythm waves, and signal reconstruction is carried out under the scale, obtain multiple reconstruction signals under corresponding band, then identical decomposition and reconstruction is carried out to each appropriate channel, again the reconstruction signal of each frequency range in each channel is joined together to be input in spatial filter, the feature vector of 6 dimensions is obtained, Mental imagery classification of task is finally completed using support vector machines. Method proposed by the present invention not only carries out frequency information analysis to Mental imagery EEG signals, moreover it is possible to it is insufficient effectively to overcome the problems, such as that electrode is chosen.

Description

A kind of dual-tree complex wavelet together spatial model combine brain electrical feature extracting method

Technical field

The invention belongs to EEG Processing field, relate to a kind of dual-tree complex wavelet together spatial model combine brain electrical feature extracting method.

Background technology

For the patient of those neuromuscular system function major injuries, be badly in need of a kind of newly carry out with the external world method that exchanges.And a kind of method that brain-computer interface is so just, it does not rely on brain outside nervous system and muscular tissue, is that one sets up a kind of directly communication and control passage between human brain and computer or other electronic equipments.At present, the EEG signals type being applicable to brain-computer interface has a variety of, and wherein Mental imagery EEG signals applies one of maximum type, because Mental imagery EEG signals has without the need to environmental stimuli, can the advantages such as asynchronous communication be realized, meet the demand for development of BCI technology.

Up to the present, people still know little about it to the forming process of human thinking, and the various thinking activitiess being read people by brain electricity are also unrealistic.But EEG signals is used for Mental imagery identification aspect, has achieved certain progress.More typical application example is that the Niels Birbaurmer of Tübingen, Germany university utilizes the EEG signals of paralytic successfully to control computer cursor in nineteen ninety to move.The researchs such as the Pfurtscheller of Graz University of Science and Technology of Austria show, when human body makes monolateral limb motion or imagery motion, brain offside produces Event-related desynchronization current potential (ERD), and brain homonymy produces event-related design current potential (ERS).The experiment of Pfurtscheller can dope the motion of left hand or the right hand simultaneously, and classification accuracy rate reaches 85%.2000, Graz research group carried out the research of right-hand man's two kinds of Mental imagery classification.At present, the research group that professor Pfurtscheller leads successfully establishes GrazI and GrazII system, first, experimenter by the imagination right-hand man or forefinger, right crus of diaphragm motion produce EEG signals system is controlled, part experimenter is by after training, the Mental imagery discrimination of two classes obtain up to 85% at linear resolution, but the discrimination of three classes only has 77%.

The research that U.S. Wadsworth carries out at center is the one dimension or the two dimensional motion that are carried out steering needle by the EEG signals of record brain motorsensory cortex.In its experiment, subjects first to be trained to learn freely to control the amplitude of μ rhythm and beta response, thus realize that cursor moves by the change of these rhythm and pace of moving things, the function such as letter spelling and prosthesis control.Germany Berlin polytechnical university Muller etc. have developed Hex-O-Spell typewriter, and experimenter is by imagining that the spelling of typewriting is carried out in the motion of trick.Japanese scholars Yamawaki sets up time-frequency model and carries out sort research for Mental imagery EEG signals, also achieves good classifying quality.Subsequently, based on the feature extracting method of public space pattern (CSP), substantially increase the signal to noise ratio of EEG signals.The people such as the Kai Keng Ang of Singapore improve CSP method, propose spatial filter group (FBCSP), the method such as space filtering (RFBCSP), Time-Frequency Domain Filtering network (OSSFN) based on robustness bank of filters respectively, progressively improve the degree of accuracy of pattern classification, and in the BCI contest of 2008, achieve excellent achievement.Recently, Minnesota university of U.S. He etc. utilizes multitask imagery motion EEG signals to control helicopter in virtual environment to complete the motions such as forward/backward, rise/fall, left/right rotation, and three experimenters obtain about 85% accuracy generally.

Domestic, the systematic studys such as the EEG signals research group of the Tsing-Hua University BCI cursor carried out based on Mental imagery moves, rehabilitation supplemental training, achieve the online average correct classification rate of 79.48% and the off-line average correct classification rate of 85% in the Study of recognition of three type games imagination tasks.Xi'an Communications University Zheng Chong merit etc. utilize product complexity theory to extract EEG signals in addition, realize two class mental tasks classifications, obtain good effect by the analysis of Mahalanobis distance discrimination formula.Shanghai Communications University Zhang Liqing etc. have studied common tensor Discrimination Analysis Algorithm to extract the characteristic vector of the EEG sample of single, and recycling support vector machine is classified, and result shows that this algorithm improves classification accuracy.Southeast China University Song to love homeland etc. and adopts wavelet transform and AR model to carry out feature extraction to the EEG signals of Mental imagery, two generic task Mental imagery discrimination average out to 89.5% respectively.

It is generally acknowledged, BCI systems attempt adopts the passage of few way, and the benefit done like this is apparent, and required electrode is few, not only shortens time, and low volume data needs little information processing cost.Corresponding with it, the scholars such as German Berlin polytechnical university Blankertz point out, a small amount of passage adopting nervous physiology priori to select might not produce the result better than full tunnel collection, and electrode chooses deficiency also can reduce classification accuracy rate.Subsequently, to several experimenter, Sannelli, Schroder, Barachant etc. imagine that the EEG data that right-hand man, foot etc. move is studied, especially, when optionally selection portion subchannel is studied, the nicety of grading that different combination of channels obtains differs greatly.Best classifying quality is obtained for target with minimum eeg data, the scholars such as Arvaneh, Gao study CHANNEL OPTIMIZATION select permeability further, result shows to adopt CSP and the method such as expansion algorithm, SVM recurrence passage exclusive method thereof can both find the channel position of the most applicable particular subject under certain selecting criterion, not only reduce number of electrodes, and raising classification performance, also point out that a large amount of passage capable of being provides the information of more horn of plenty simultaneously.

Need a large amount of electrodes for CSP algorithm, lack the shortcoming of frequency information analysis, the present invention mainly reduces electrodeplate, reduces information processing cost, meanwhile increases frequency information analysis, makes up the deficiency of CSP algorithm in EEG Processing.

Summary of the invention

Object of the present invention is widely used in EEG feature extraction for CSP algorithm exactly, but the method needs a large amount of electrodes, lacks frequency information analysis, and a kind of dual-tree complex wavelet the proposed feature extracting method that combines of spatial model together.Lacking the defect of frequency information in order to compensate for CSP space filtering, improve the discrimination of the eigenvalue corresponding to different motion imagination task, and then improve the recognition accuracy of grader.

In order to realize above object, the inventive method mainly comprises the following steps:

Step (1) obtains the Mental imagery EEG signals of suitable passage.Specifically, carry out different Mental imagery tasks and can activate zones of different on brain motor cortex, select corresponding electrode according to brain, thus obtain the Mental imagery EEG signals that this electrode gathers, its sample frequency is

Step (2) asks for suitable double sampling frequency f _s, the frequency range that guarantee step (3) reconstructs the EEG signals obtained is 0 ~ 30Hz.Specifically, according to the frequency range of the l layer wavelet coefficient that dual-tree complex wavelet multi-resolution decomposition obtains, suitable f is solved _sand l, namely

f _s/2 ^l≈30Hz (1)

Wherein, 30Hz is the maximum of EEG signals beta response ripple.Again according to f _sto original frequency carry out corresponding up-sampling or down-sampling, the sample frequency finally obtained is f _s.If concrete sampling selection principle is <f _stime, then carry out up-sampling, otherwise carry out down-sampling.

Step (3) carries out dual-tree complex wavelet decomposition and reconstruction to EEG signals.Specifically, first the EEG signals after double sampling is carried out dual-tree complex wavelet decomposition, then choose the frequency range that δ (0.5 ~ 4Hz), θ (4 ~ 8Hz), α (8 ~ 14Hz) and β (14 ~ 30Hz) 4 species rhythm wave frequency range are corresponding, and reconstructed.

Step (4) combines the reconstruction signal of each passage and carries out CSP space filtering.Specifically, combined by the reconstruction signal of each suitable passage, be then input in CSP spatial filter simultaneously and carry out filtering, filtered signal obtains the characteristic vector f needed for grader through formula (2) _p.

$f_p=\lg (\mathrm{var}\left(Z_p\right)/\underset{i=1}{\overset{2m}{\mathrm{\&Sigma;}}}\mathrm{var}\left(Z_i\right)),p=1,...,2m$

Wherein var represents the variance asking vector, and m represents the logarithm of CSP wave filter, and log represents natural logrithm computing, Z _p(p=1,2 ..., 2m) represent the filtered characteristic vector of CSP.

Compared with existing Mental imagery brain electrical feature extracting method, the method that the present invention proposes not only can remove the frequency band irrelevant with Mental imagery, make up the defect that CSP space filtering lacks frequency information, thus improve the discrimination of the eigenvalue corresponding to different motion imagination task; The requirement of CSP to port number can also be reduced, thus decrease redundancy and noise that some uncorrelated or sound pollution passages provide to a certain extent.

The inventive method can meet the requirements for extracting features in pattern recognition task preferably, has broad application prospects at brain-computer interface, disease of brain diagnostic field.

Accompanying drawing explanation

Fig. 1 is implementing procedure figure of the present invention, i.e. EEG feature extraction handling process.

Detailed description of the invention

Describe the brain electrical feature extracting method that the present invention is based on dual-tree complex wavelet spatial model together in detail below in conjunction with accompanying drawing, Fig. 1 is implementing procedure figure.

As Fig. 1, the enforcement of the inventive method mainly comprises following step:

(1) multichannel Mental imagery EEG signals eeg data is obtained;

(2) select corresponding electrode, obtain the EEG signals data of respective channel;

(3) calculate the sample frequency relevant to the multiple dimensioned frequency division of dual-tree complex wavelet, and up-sampling or down-sampling are carried out to original sample frequency;

(4) each passage EEG signals carries out dual-tree complex wavelet decomposition;

(5) choose relevant frequency range, and reconstructed;

(6) combine the reconstruction signal of each passage, be input in CSP wave filter;

(7) filtered signal is carried out change of taking the logarithm, be input to support vector machine classifier and carry out training and testing.

One by one each step is described in detail below.

Step one: obtain multichannel Mental imagery EEG signals sample data

In U.S. Neuro Scan company Scan4.3 collecting device 40 conduction polar caps are adopted to carry out Mental imagery process eeg signal acquisition.Experimenter has worn brain electricity cap recoil on request on wheelchair, keeps quite, naturally, watches the sight prompting set in experimental situation attentively.Adopt the following two kinds Mental imagery experimental paradigm: right hand manipulation wheelchair control bar forward, left hand manipulation wheelchair control bar backward respectively corresponding wheelchair advance, the controlled motion form of brake, in implementation process, also the design of experimental concrete condition to experiment model can do suitable correction.

Step 2: select corresponding electrode, obtains the EEG signals data of respective channel

Carry out different Mental imagery tasks according to brain and can activate zones of different on brain motor cortex, then experimentally normal form selects corresponding electrode, and obtain the EEG signals data of this passage, sample frequency is

Step 3: the selection of double sampling frequency

In order to ensure that the frequency range that step 4 reconstructs the EEG signals obtained is 0 ~ 30Hz, then the frequency range of the l layer wavelet coefficient obtained according to dual-tree complex wavelet multi-resolution decomposition, asks suitable f according to formula (1) _sand l.Again according to f _sto original frequency carry out corresponding up-sampling or down-sampling, the sample frequency finally obtained is f _s.If concrete sampling selection principle is time, then carry out up-sampling, otherwise carry out down-sampling.

Step 4: dual-tree complex wavelet decomposes

Carry out L layer dual-tree complex wavelet to each passage EEG signals data selected to decompose, obtain the signal frequency range of L+1 reconstruct.If the sample frequency of EEG signals is fs, carry out the decomposition of L layer, obtain reconstruction signal, wherein the frequency range of l layer wavelet coefficient and scale coefficient is as follows:

f(d _l)∈[f _s/2 ^l+1,f _s/2 ^l]Hz，l∈1,2,...,L

f(c _L)∈[0,f _s/2 ^L+1]

Step 5: choose relevant frequency range, and reconstructed

First the EEG signals after sampling is carried out dual-tree complex wavelet decomposition, then choose the frequency range that δ, θ, α and β tetra-species rhythm wave frequency scopes are corresponding, and reconstructed.

Step 6: CSP filtering

Combined by the reconstruction signal of each frequency range of each passage, be then input in CSP spatial filter simultaneously and carry out filtering, its median filter logarithm is m.

Step 7: based on the Mental imagery classification of task of support vector machine

The characteristic vector needed for grader will be obtained through the filtered signal of step 6 through equation (2), and it can be used as the input of support vector machine classifier, carry out training and testing, complete the classification of multi-motion imagination task.

Claims (1)

1. dual-tree complex wavelet together spatial model combine a brain electrical feature extracting method, it is characterized in that the method comprises the steps:

Step (1) obtains the Mental imagery EEG signals of suitable passage; Specifically, carry out the zones of different on different Mental imagery mission-enabling brain motor cortexs, select corresponding electrode according to brain, thus obtain the Mental imagery EEG signals of this electrode collection, its sample frequency is

Step (2) asks for suitable double sampling frequency f _s, the frequency range that guarantee step (3) reconstructs the EEG signals obtained is 0 ~ 30Hz; According to the frequency range of the l layer wavelet coefficient that dual-tree complex wavelet multi-resolution decomposition obtains, solve suitable f _sand l, namely

f _s/2 ^l≈30Hz (1)

Wherein, 30Hz is the maximum of EEG signals beta response ripple; Again according to f _sto original frequency carry out corresponding up-sampling or down-sampling, the sample frequency finally obtained is f _s; If concrete sampling selection principle is time, then carry out up-sampling, otherwise carry out down-sampling;

If the sample frequency of EEG signals is f _s, carry out the decomposition of L layer, obtain reconstruction signal, wherein l layer wavelet coefficient f (d _l) and scale coefficient f (c _l) frequency range as follows:

f(d _l)∈[f _s/2 ^l+1,f _s/2 ^l]Hz，l∈1,2,...,L；f(c _L)∈[0,f _s/2 ^L+1]；

Step (3) carries out dual-tree complex wavelet decomposition and reconstruction to EEG signals; Specifically, first the EEG signals after double sampling is carried out dual-tree complex wavelet decomposition, then the frequency range that δ, θ, α and β tetra-species rhythm wave frequency range are corresponding is chosen, and reconstructed, wherein δ corresponds to 0.5Hz ~ 4Hz, θ and corresponds to 4Hz ~ 8Hz, α corresponding to 8Hz ~ 14Hz, β corresponding to 14Hz ~ 30Hz;

Step (4) combines the reconstruction signal of each passage and carries out CSP space filtering; Specifically, combined by the reconstruction signal of each suitable passage, be then input in CSP spatial filter simultaneously and carry out filtering, filtered signal obtains the characteristic vector f needed for grader through formula (2) _p;

$f_p=\lg (\mathrm{var}\left(Z_p\right)/\underset{i=1}{\overset{2m}{\mathrm{\&Sigma;}}}\mathrm{var}\left(Z_i\right)),p=1,...,2m---\left(2\right)$

Wherein, var represents the variance asking vector, and m represents the logarithm of CSP wave filter, and lg represents natural logrithm computing, Z _prepresent the filtered characteristic vector of CSP.