CN102306303B - Electroencephalography signal characteristic extraction method based on small training samples - Google Patents

Abstract

The invention relates to a characteristic extraction method of imagining action potential in a BCI (Brain-Computer Interface) device and particularly relates to a characteristic extraction method combining a regularization method and a CSSD (Common Special Subspace Decomposition) algorithm. In the method provided by the invention, regularization parameters are led; a covariance matrix of training data of a target experimenter and a covariance matrix of training data of an auxiliary experimenter are combined to form a regularization covariance matrix under the action of the regularization parameters; a regularization space filter is constructed; and characteristic analysis is carried out on the test data of the target experimenter by utilizing the regularization space filter. By using the method in the invention, the problems that characteristic value is unstable, classification accuracy rate is low and the like in the CSSD algorithm are solved when a small-sample problem is processed.

Description

A kind of EEG feature extraction method based on little training sample

Technical field

The invention belongs to Pattern Recognition and Intelligent System and brain-computer interface field; Be particularly related to brain-computer interface (Brain-Computer Interface; BCI) extraction of Imaginary EEG signal characteristic vector in the system and device; Specifically be to spatial domain Subspace Decomposition algorithm (Common Special Subspace Decomposition altogether with regularization (Regularization) method; CSSD) improve last and k nearest neighbor (K-Nearest Neighbor, KNN) feature extraction and the sorting technique that combine of sorting algorithm.

Background technology

At present, there is multiple disease can damage the neuromuscular path that brain exchanges and controls with external environment condition, like brain paralysis multiple sclerosis and cerebral apoplexy etc.These diseases can make the people partly or entirely lose autonomous muscle control.Along with development of science and technology, modern life support technology can make patient's long-term surviving, but patient's quality of life is low, causes very white elephant for family and society.

Deepening continuously of the progress of Along with computer technology and brain function research; People begin to attempt to set up a kind of brand-new, do not rely on neururgic interchange of muscle and control path; Transmission information and order between brain and external environment condition; (Brain-Computer Interface BCI), is called for short brain-computer interface for Here it is so-called brain-computer interface.The brain-computer interface technology has important use to be worth in fields such as rehabilitation project, military affairs, especially receive much concern especially at the robot rehabilitation field.

The BCI structural representation is as shown in Figure 1, and the brain wave acquisition device is gathered EEG signals from cerebral cortex, partly carries out filtering through Signal Pretreatment, passes through the feature extraction and the Classification and Identification of EEG signals again, realizes the control to peripheral hardware thereby be converted into control signal.

BCI is the direct communication channel of between human brain and external unit, setting up.Through this passage, the people can directly give an order to external unit through brain, and does not need the support of language or action, can effectively strengthen health handicap patient and the extraneous ability that exchanges or control external environment condition, thereby improve patient's quality of life.Is to EEG signals (electroencephalography, online treatment EEG) with the BCI system applies in the prerequisite of recovering aid opertaing device.In the online experiment of brain-computer interface, the experimenter will carry out some uninteresting training experiments usually, sorter is imported the training data of a group echo classification.But the repetition training meeting makes nervous system produce tired and influences experimental result for a long time, and data volume is big, processing speed is slow.So people hope to reduce the number of times of required experiment as far as possible.Therefore, when previous important target is the time of reducing the initialization training, promptly to be issued to high recognition in the prerequisite of small sample training.Yet the core of BCI is the feature extraction to the EEG signal.

Feature extraction is exactly through transform method abundant in the signal Processing field some key characters of the relevant mental awareness of reflection brain in the EEG signal to be displayed in transform domain, thereby removes the nonsensical information of classifying.In the research of EEG signal, seeking the effective feature extraction method is one of gordian technique that improves recognition accuracy.

The triumphant utilization CSSD of group algorithm carries out feature extraction to the EEG signal on Tsing-Hua University's height in 2003, has obtained good identification effect.CSSD is a kind of airspace filter algorithm to the multichannel brain electric data, and its effect is the relevant component of signal of extraction task and incoherent component of inhibition task and noise.(event-related synchronization, ERS) phenomenon is very effective to the incident related synchronizationization of handling the EEG signal for CSSD.The eigenwert stability of the proper vector that its deficiency is to make up is low, and discrimination is poor, and is especially more obvious when small sample.

Summary of the invention

In motion imagination BCI system, there are problems such as eigenwert is unstable, classification accuracy is low in the CSSD feature extracting method under the less situation of training sample, proposes the new feature extracting method based on CSSD.The present invention gathers the healthy experimenter's in n position EEG signal, and SF is 128Hz.Choosing an experimenter is the target experimenter, and then remaining is the assistant experiment person.Require the experimenter to carry out imagination left hand task and imagination right hand task.Introduce regularization parameter ρ and σ then, under the effect of regularization parameter, target experimenter's training sample and assistant experiment person's training sample are combined the training sample as the target experimenter.Regularization method and CSSD method are combined, be regularization subspace, spatial domain (Regularized Common Special Subspace Decomposition, R-CSSD) decomposition algorithm altogether.Utilize the regularization spatial filter of R-CSSD structure that target experimenter's test sample book is carried out spatial filtering then, extract proper vector.At last, utilize the KNN sorter to accomplish classification to target experimenter test sample book.

The present invention is based on the hardware platform that is made up of electroencephalograph and PC; Gather the EEG signal through electrode cap; Then the EEG signal is sent to PC and handles, utilize PC to realize the R-CSSD algorithm, and combine with the KNN sorting technique and to accomplish classification the EEG signal.Passage is according to international standard path 10-20 system rest, and is as shown in Figure 2, and dash area is that selected passage distributes among the figure.Gather the EEG data of 3 passages, be respectively ' C ₃', ' C _Z', ' C ₄' passage.

The EEG signal processing method that the present invention utilized mainly may further comprise the steps:

(1) take n position experimenter to imagine left and right chirokinesthetic EEG signal, n be natural number and n greater than 4, SF is 128Hz.Require experimenter's peace and quiet to loosen and be sitting on the chair attonity; The initial moment that can show this time period on the computer screen, require the experimenter according to left, the prompting of arrow to the right carries out the imagery motion of left hand and right hand; Can complete documentation brain electricity change procedure in this process.When every experimenter imagined the left hand motion, its EEG signal corresponding class was labeled as category-A, and during the motion of the imagination right hand, its EEG signal corresponding class is labeled as category-B, X _A, X _BThe left and right chirokinesthetic EEG signal of the expression imagination respectively.

According to the achievement in research of clinical electrophysiology, the cerebral cortex zone that different limbs position motions is activated also has nothing in common with each other.Monolateral limb motion or imagery motion can activate main sensorimotor cortex; The relevant current potential ERD (event-related desynchronization) that desynchronizes of brain offside generation incident, the brain homonymy produces incident related synchronization current potential ERS (event-related synchronization).ERD is meant that the periodic activity of CF shows the reduction of amplitude when enliven in a certain cortex zone, and ERS is meant that working as a certain activity does not make relevant cortex zone enliven significantly constantly certain, and CF just shows amplitude and raises.According to this physiology imagination, utilize 8～30Hz BPF. that the EEG signal data that collects is carried out filtering, to obtain tangible ERD/ERS physiological phenomenon.

When Fig. 3 moves for imagination right-hand man, C ₃And C ₄ERD/ERS phenomenon on the passage.

As can beappreciated from fig. 3 imagine left hand when motion, C as the experimenter ₃The power of passage is higher than C ₄The power of passage.And imagination right hand when motion C ₄The power of passage is higher than C ₃Power.

(2) n position experimenter is carried out serial number, selected experimenter is the target experimenter, and other remaining n-1 position experimenters are the assistant experiment person;

(3) the EEG signal of gathering is carried out the bandpass filtering of 8-30Hz.

Electrophysiologic studies shows that the motion imagination can produce the variable power of the detectable motion-sensing rhythm and pace of moving things.Specifically, it is that the u rhythm and pace of moving things and the frequency of 8～12Hz be that the amplitude of the beta response of 13～28Hz suppresses is the relevant desynchronization ERD of incident that the motion imagination causes frequency, or the amplitude increase is incident related synchronization ERS.These physiological rhythm and pace of moving things signals can be used for differentiating the chirokinesthetic EEG signal in the imagination left and right sides.Therefore the EEG signal of gathering is carried out the bandpass filtering of 8～30Hz;

(4) to after the signal filtering; From target experimenter category-A and category-B EEG signal, choose m EEG signal respectively as training sample; M is that natural number and m are less than 20; Then other EEG signals of target experimenter are test sample book, from every assistant experiment person's EEG signal, extract the training sample with target experimenter similar number then;

(5) ask the covariance matrix sum R of target experimenter category-A and category-B training sample respectively _AWith R _BThe covariance matrix sum of all assistant experiment person category-As and category-B training sample

With

(6) introduce regularization parameter ρ and σ, under the effect of regularization parameter, covariance matrix sum that will (5) middle target experimenter combines with assistant experiment person's covariance sum, constructs two types of average regularized covariance matrixes, and is as follows:

$Z_A(\mathrm{\ρ},\mathrm{\σ})=(1-\mathrm{\σ})\frac{(1-\mathrm{\ρ})\·R_A+\mathrm{\ρ}\·{\hat{R}}_A}{(1-\mathrm{\ρ})\·m+\mathrm{\ρ}\·(n-1)\·m}+\frac{\mathrm{\σ}}{3}\mathrm{tr}\left[\frac{(1-\mathrm{\ρ})\·R_A+\mathrm{\ρ}\·{\hat{R}}_A}{(1-\mathrm{\ρ})\·m+\mathrm{\ρ}\·(n-1)\·m}\right]\·I$

$Z_B=(\mathrm{\ρ},\mathrm{\σ})=(1-\mathrm{\σ})\frac{(1-\mathrm{\ρ})\·R_B+\mathrm{\ρ}\·{\hat{R}}_B}{(1-\mathrm{\ρ})\·m+\mathrm{\ρ}\·(n-1)\·m}+\frac{\mathrm{\σ}}{3}\mathrm{tr}\left[\frac{(1-\mathrm{\ρ})\·R_B+\mathrm{\ρ}\·{\hat{R}}_B}{(1-\mathrm{\ρ})\·m+\mathrm{\ρ}\·(n-1)\·m}\right]\·I$

Wherein, The mark of

expression

, I does

3 * 3 unit matrix.

(7) with two types in (6) average regularized covariance matrixes summations and carry out characteristic value decomposition, find the solution canonical albefaction matrix, as follows:

$Z(\mathrm{\ρ},\mathrm{\σ})=Z_A(\mathrm{\ρ},\mathrm{\σ})+Z_B(\mathrm{\ρ},\mathrm{\σ})$

$=\widehat{U\·}\widehat{\mathrm{\Λ}}\·{\hat{U}}^T$

Wherein, is the eigenwert diagonal matrix; is the characteristic of correspondence vector matrix, and then canonical albefaction matrix is:

$P(\mathrm{\ρ},\mathrm{\σ})={\widehat{\mathrm{\Λ}}}^{(-1/2)}\·\hat{U}\·{\widehat{\mathrm{\Λ}}}^{(-1/2)}$

(8) to the Z of gained in (6) _A(ρ, σ) and Z _B(ρ, σ) change as follows:

${\hat{Z}}_A(\mathrm{\ρ},\mathrm{\σ})=P(\mathrm{\ρ},\mathrm{\σ})\·Z_A(\mathrm{\ρ},\mathrm{\σ})\·P{(\mathrm{\ρ},\mathrm{\σ})}^T$

$=U_{0A}\·{\mathrm{\Λ}}_A\·U_{0A}^T$

${\hat{Z}}_B(\mathrm{\ρ},\mathrm{\σ})=P(\mathrm{\ρ},\mathrm{\σ})\·Z_B(\mathrm{\ρ},\mathrm{\σ})\·P{(\mathrm{\ρ},\mathrm{\σ})}^T$

${=U}_{0B}\·{\mathrm{\Λ}}_B\·U_{0B}^T$

Wherein, Λ _AAnd Λ _BBe eigenwert diagonal matrix, U _0AAnd U _0BBe the characteristic of correspondence vector matrix, utilize principal component analytical method U _0AAnd U _0BIn proper vector choose, choose diagonal matrix Λ _AMiddle eigenvalue of maximum characteristic of correspondence vector does

Λ _BMiddle eigenvalue of maximum characteristic of correspondence vector does

(9) two types of regularization spatial filters of structure are following:

${\mathrm{SF}}_A(\mathrm{\ρ},\mathrm{\σ})=U_{0A}^A\·P(\mathrm{\ρ},\mathrm{\σ})$

${\mathrm{SF}}_B(\mathrm{\ρ},\mathrm{\σ}){=U}_{0B}^B\·P(\mathrm{\ρ},\mathrm{\σ})$

(10) training sample and the A in the test sample book, the category-B EEG signal with the target experimenter passes through corresponding spatial filter SF respectively _A(ρ, σ) and SF _B(ρ σ) carries out spatial filtering, extracts the proper vector f of EEG signal _A, f _B, as follows:

f _A＝SF _A(ρ，σ)·X _A

f _B＝SF _B(ρ，σ)·X _B

Wherein, f _AThe proper vector of corresponding imagination left hand signal, f _BThe proper vector of corresponding imagination right hand signal.

(11) utilize the KNN sorting technique that the proper vector of target experimenter's test sample book in (10) is classified.

With the proper vector of the target experimenter's that obtains in (10) training sample as normal data; A given test sample book is calculated each training sample and its distance, according to these distances; Utilize the KNN algorithm computation to go out K nearest neighbours in the test set; Vote according to these neighbours' categorical attribute then, the predicted value that draws is composed to by the categorical attribute of object of classification, use the classification of the classification of this training sample as this test sample book.

Beneficial effect

The improved CSSD method of the present invention can make the eigenvector integration objective experimenter's of extraction training sample and assistant experiment person's training sample; Increase training sample; Problems such as the unstable and Classification and Identification rate of eigenwert is low have been avoided under the little training sample situation; Help the use of spread sorting device, in the brain-computer interface technical field, improve EEG signal classification accuracy, shown certain advantage.

Description of drawings

Fig. 1 midbrain of the present invention-machine interface structure synoptic diagram

International standard channel 10-20 system rest figure among Fig. 2 the present invention, dash area for gather among the present invention experimenter EEG signal used the passage distribution plan.

C when imagination right-hand man moves among Fig. 3 the present invention ₃And C ₄The ERD/ERS phenomenon that shows on the passage.

In Fig. 4 embodiment of the invention based on the EEG signal classification process figure of little training sample.

The variation diagram of EEG signal in Fig. 5 embodiment of the invention processing procedure.

Utilization traditional C SD method is carried out feature extraction with the R-CSSD method to identical EEG signal among table 1 the present invention, utilizes the KNN sorter to classify, the comparison of two kinds of method classification accuracies.

Embodiment

Below in conjunction with accompanying drawing specific embodiment of the present invention is further specified.Fig. 4 is based on the EEG signal classification process figure of little training sample in the embodiment of the invention.

G.tec electroencephalograph and PC that the present invention adopts Austrian g.tec company to produce are hardware platform.The g.tec electroencephalograph comprises electrode cap, eeg amplifier, A/D converter etc.The present invention gathers the EEG signal through electrode cap, and the EEG signal amplifies through eeg amplifier and the A/D conversion, transports to PC by the USB mouth, and stores in storer with signal voltage amplitude form.The EEG data of electroencephalograph being sent through MATLAB2009a application written program receive, handle.Concrete operations are following:

(1) gathers the EEG signal

Gather 5 experimenters' EEG signal in the present embodiment, test perdurability is 9s at every turn, carries out 180 experiments altogether, and wherein left hand and right hand thought experiment is each 90 times.In gatherer process, the classification under each experiment is marked.The sampling channel number of EEG signals is 3, and t is the sampling number of each passage.When every experimenter imagined the left hand motion, its EEG signal corresponding class was labeled as category-A, and during the motion of the imagination right hand, its EEG signal corresponding class is labeled as category-B, X _A, X _BThe left and right chirokinesthetic EEG signal of the expression imagination respectively, size is 3 * t.The original EEG signal of gathering is shown in Fig. 5 (a);

(2) experimenter is numbered and chooses target experimenter and assistant experiment person

5 experimenters are numbered No. 1, No. 2, No. 3, No. 4, No. 5, and the corresponding EEG signal of then being gathered by (1) of 5 experimenters is designated as data10, data20, data30, data40, data50 successively.Selected No. 1 experimenter is the target experimenter before the EEG signal Processing, and remaining 4 experimenter is the assistant experiment person;

(3) frequency domain filtering

Design 48 rank through MATLAB2009a, the FIR wave filter of 512 sampled points carries out the 8-30Hz bandpass filtering to EEG signal data in (2), and the EEG signal after the filtering is data11, data21, data31, data41, data51.EEG signal after the filtering is shown in Fig. 5 (b);

(4) choose training sample

After signal filtering; From the category-A of target experimenter data11 and category-B EEG signal, choose 10 EEG signals respectively and be designated as data11 ' as training sample; Then other 160 EEG signals of target experimenter are that test sample book is designated as data11 "; from every assistant experiment person category-A and category-B EEG signal, all extract 10 training samples then and be designated as data21 ', data31 ', data41 ', data51 ', then all assistant experiment persons' category-A and category-B training sample sum are 40;

(5) ask the covariance matrix sum R of target experimenter category-A and category-B training sample respectively _AWith R _B, the covariance matrix sum of all assistant experiment person category-As and category-B training sample

With

The covariance matrix sum of the middle category-A EEG of target experimenter's training sample data11 ' signal, the covariance matrix sum of category-B EEG signal are as follows:

$R_A=\underset{i=1}{\overset{10}{\mathrm{\Σ}}}\frac{X_{A_i}{X}_{A_i}^T}{\mathrm{tr}\left(X_{A_i}{X}_{A_i}^T\right)}$

$R_B=\underset{i=1}{\overset{10}{\mathrm{\Σ}}}\frac{X_{B_i}{X}_{B_i}^T}{\mathrm{tr}\left(X_{B_i}{X}_{B_i}^T\right)}$

Wherein,

(i=1,2...10) EEG signal of the i time imagination of expression target experimenter left hand motion.

(i=1,2...10) EEG signal of the i time imagination of expression target experimenter right hand motion.

Expression X _{(i, A)}Transposition,

Representing matrix

Mark.

The covariance matrix sum of the category-A EEG signal of assistant experiment person training sample data21 ', data31 ', data41 ', data51 ', the covariance matrix sum of category-B EEG signal are as follows:

${\hat{R}}_A=\underset{i=1}{\overset{40}{\mathrm{\Σ}}}\frac{{\hat{X}}_{A_i}{\hat{X}}_{A_i}^T}{\mathrm{tr}\left({\hat{X}}_{A_i}{\hat{X}}_{A_i}^T\right)}$

${\hat{R}}_B=\underset{i=1}{\overset{40}{\mathrm{\Σ}}}\frac{{\hat{X}}_{B_i}{\hat{X}}_{B_i}^T}{\mathrm{tr}\left({\hat{X}}_{B_i}{\hat{X}}_{B_i}^T\right)}$

Wherein,

(i=1; 2..., 40) and represent that the assistant experiment person imagines the EEG signal of left hand motion for the i time.

(i=1; 2..., 40) and represent that the assistant experiment person imagines the EEG signal of right hand motion for the i time;

(6) ask the regularized covariance matrix

Introduce regularization parameter ρ and σ, its span is ρ ∈ [0,1] and σ ∈ [0,1], and the step-length of getting ρ and σ in the present embodiment is 0.1, and the variation of regularization parameter will influence classification accuracy.Regularization parameter combines target experimenter's in (5) covariance matrix sum with assistant experiment person's covariance sum, construct two types of average regularized covariance matrixes, and is as follows:

$Z_A(\mathrm{\ρ},\mathrm{\σ})=(1-\mathrm{\σ})\frac{(1-\mathrm{\ρ})\·R_A+\mathrm{\ρ}\·{\hat{R}}_A}{(1-\mathrm{\ρ})\·10+\mathrm{\ρ}\·40}+\frac{\mathrm{\σ}}{3}\mathrm{tr}\left[\frac{(1-\mathrm{\ρ})\·R_A+\mathrm{\ρ}\·{\hat{R}}_A}{(1-\mathrm{\ρ})\·10+\mathrm{\ρ}\·40}\right]\·I$

$Z_B(\mathrm{\ρ},\mathrm{\σ})=(1-\mathrm{\σ})\frac{(1-\mathrm{\ρ})\·R_B+\mathrm{\ρ}\·{\hat{R}}_B}{(1-\mathrm{\ρ})\·10+\mathrm{\ρ}\·40}+\frac{\mathrm{\σ}}{3}\mathrm{tr}\left[\frac{(1-\mathrm{\ρ})\·R_B+\mathrm{\ρ}\·{\hat{R}}_B}{(1-\mathrm{\ρ})\·10+\mathrm{\ρ}\·40}\right]\·I$

Wherein, The mark of

expression

, I representes 3 * 3 unit matrix;

(7) find the solution canonical albefaction matrix P (ρ, σ)

With two types in (6) average regularized covariance matrixes summations and carry out characteristic value decomposition, find the solution canonical albefaction matrix, as follows:

$Z(\mathrm{\ρ},\mathrm{\σ})=Z_A(\mathrm{\ρ},\mathrm{\σ})+Z_B(\mathrm{\ρ},\mathrm{\σ})$

$=\hat{U}\·\widehat{\mathrm{\Λ}}\·{\hat{U}}^T$

Wherein,

is the eigenwert diagonal matrix;

is the characteristic of correspondence vector matrix, and then canonical albefaction matrix is:

$P(\mathrm{\ρ},\mathrm{\σ})={\widehat{\mathrm{\Λ}}}^{(-1/2)}\·\hat{U}\·{\widehat{\mathrm{\Λ}}}^{(-1/2)}$

(8) to the Z of gained in (6) _A(ρ, σ) and Z _B(ρ, σ) change as follows:

${\hat{Z}}_A(\mathrm{\ρ},\mathrm{\σ})=P(\mathrm{\ρ},\mathrm{\σ})\·Z_A(\mathrm{\ρ},\mathrm{\σ})\·P{(\mathrm{\ρ},\mathrm{\σ})}^T$

${=U}_{0A}\·{\mathrm{\Λ}}_A\·U_{0A}^T$

${\hat{Z}}_B(\mathrm{\ρ},\mathrm{\σ})=P(\mathrm{\ρ},\mathrm{\σ})\·Z_B(\mathrm{\ρ},\mathrm{\σ})\·P{(\mathrm{\ρ},\mathrm{\σ})}^T$

$=U_{0B}\·{\mathrm{\Λ}}_B\·U_{0B}^T$

Wherein, Λ _AAnd Λ _BBe eigenwert diagonal matrix, U _0AAnd U _0BBe the characteristic of correspondence vector matrix, utilize principal component analytical method (referring to " Digital Signal Analysis and Processing ") U _0AAnd U _0BIn proper vector choose, choose diagonal matrix Λ _AMiddle eigenvalue of maximum characteristic of correspondence vector does

Λ _BMiddle eigenvalue of maximum characteristic of correspondence vector does

(9) structure regularization spatial filter

Two types of regularization spatial filters are:

${\mathrm{SF}}_A(\mathrm{\ρ},\mathrm{\σ})=U_{0A}^A\·P(\mathrm{\ρ},\mathrm{\σ})$

${\mathrm{SF}}_B(\mathrm{\ρ},\mathrm{\σ}){=U}_{0B}^B\·P(\mathrm{\ρ},\mathrm{\σ})$

Wherein, SF _A(ρ, σ) corresponding left hand action potential, SF _B(ρ, σ) corresponding right hand action potential.So far, the structure of imagining the regularization spatial filter of two generic tasks motion is accomplished.

(10) extract the EEG signal characteristic vector

Target experimenter's training sample and the A in the test sample book, category-B EEG signal are passed through corresponding spatial filter SF respectively _A(ρ, σ) and SF _B(ρ σ) carries out spatial filtering, extracts the proper vector f of EEG signal _A, f _B, as follows:

f _A＝SF _A(ρ，σ)·X _A

f _B＝SF _B(ρ，σ)·X _B

Wherein, f _AThe proper vector of corresponding imagination left hand signal, f _BThe proper vector of corresponding imagination right hand signal.Shown in Fig. 5 (c);

(11) based on the sorting algorithm of KNN

The EEG signal that the present invention is extracted after with the spatial filter filtering in target experimenter's the training sample data11 ' process (9) is designated as data12 ', target experimenter's test data data11 " the EEG signal through being extracted after the spatial filter filtering in (9) is designated as data12 ".Data12 ' as master sample, and is divided into two groups of a group and b group with it, training sample be category-A all be placed on a group, training sample is that the b that all is placed on of category-B organizes.Then; A given data12 " in test sample book; utilize the KNN algorithm computation to go out K the nearest neighbours in the test set; to vote according to these neighbours' categorical attribute then, the predicted value that draws composed to by the categorical attribute of object of classification, with the classification of this training sample classification as this test sample book.Simultaneously, the span of regularization parameter ρ and σ is ρ ∈ [0,1] and σ ∈ [0,1] in (6), and the step-length that both change is 0.1, all corresponding classification accuracy rate of each group regularization parameter, and what then the highest sorted accuracy was corresponding is optimum regularization parameter.

The algorithm that the present invention uses is to utilize the Space Time analysis to discern the EEG signals of imagery motion.At first according to the characteristics of the ERD/ERS of EEG signals, design R-CSSD spatial filter, the proper vector method for distilling that proposes with this paper extracts the characteristic of signal.Utilize the KNN sorting technique that data are classified at last, the more traditional CSSD feature extracting method of classification accuracy is significantly improved, and is as shown in table 1.

Claims (1)

1. the EEG feature extraction method based on little training sample based on the hardware platform that is made up of electroencephalograph and PC, is gathered the EEG signal through electrode cap, then the EEG signal is sent to PC and handles, and it is characterized in that may further comprise the steps:

(1) take n position experimenter to imagine left and right chirokinesthetic EEG signal, n be natural number and n greater than 4, when every experimenter imagined the left hand motion, its EEG signal corresponding class was labeled as category-A, during the motion of the imagination right hand, its EEG signal corresponding class is labeled as category-B, X _A, X _BThe left and right chirokinesthetic EEG signal of the expression imagination respectively;

(2) n position experimenter is carried out serial number, selected experimenter is the target experimenter, and other remaining n-1 position experimenters are the assistant experiment person;

(3) the EEG signal of gathering is carried out the bandpass filtering of 8-30Hz;

(4) to after the signal filtering; From target experimenter category-A and category-B EEG signal, choose m EEG signal respectively as training sample; M is that natural number and m are less than 20; Then other EEG signals of target experimenter are test sample book, from every assistant experiment person's EEG signal, extract the training sample with target experimenter similar number then;

(5) ask the covariance matrix sum R of target experimenter category-A and category-B training sample respectively _AWith R _BThe covariance matrix sum of all assistant experiment person category-As and category-B training sample With

(6) introduce regularization parameter ρ and σ, under the effect of regularization parameter, covariance matrix sum that will (5) middle target experimenter combines with assistant experiment person's covariance sum, constructs two types of average regularized covariance matrixes, and is as follows:

Wherein, The mark of

expression

, I is 3 * 3 unit matrix;

(7) with two types in (6) average regularized covariance matrixes summations and carry out characteristic value decomposition, find the solution canonical albefaction matrix, as follows:

Wherein,

is the eigenwert diagonal matrix; is the characteristic of correspondence vector matrix, and then canonical albefaction matrix is:

(8) to the Z of gained in (6) _A(ρ, σ) and Z _B(ρ, σ) change as follows:

Wherein, Λ _AAnd Λ _BBe eigenwert diagonal matrix, U _0AAnd U _0BBe the characteristic of correspondence vector matrix, utilize principal component analytical method U _0AAnd U _0BIn proper vector choose, choose diagonal matrix Λ _AMiddle eigenvalue of maximum characteristic of correspondence vector does

Λ _BMiddle eigenvalue of maximum characteristic of correspondence vector does

(9) two types of regularization spatial filters of structure are following:

Wherein, SF _A(ρ, σ) corresponding left hand action potential, SF _B(ρ, σ) corresponding right hand action potential;

(10) training sample and the A in the test sample book, the category-B EEG signal with the target experimenter passes through corresponding spatial filter SF respectively _A(ρ, σ) and SF _B(ρ σ) carries out spatial filtering, extracts the proper vector f of EEG signal _A, f _B, as follows:

f _A=SF _A(ρ,σ)·X _A

f _B=SF _B(ρ,σ)·X _B

Wherein, f _AThe proper vector of corresponding imagination left hand signal, f _BThe proper vector of corresponding imagination right hand signal; (11) with the proper vector of the target experimenter's who obtains in (10) training sample as master sample; Utilize K arest neighbors (k-Nearest Neighbor; KNN) sorting technique is classified to the proper vector of target experimenter's test sample book in (10), and classification results is category-A and category-B.

Images