Research on Algorithm for Feature Extraction and Classification of Motor Imagery EEG Signals

Transcription

1 BIO Web of Conferences 8, 3 (7) DOI:.5/ boconf/783 ICMSB6 Research on Algorthm for Feature Extracton and Classfcaton of Motor Imagery EEG Sgnals uan Tan, a and Zhaochen Zhang College of Medcal Informaton and Engneerng, Tashan Medcal Unversty, Taan 76, Chna Abstract.Ths paper made a research on the feature extracton and pattern recognton of left and rght hands motor magery EEG sgnals. In combnaton wth the data from BCI Competton III, denosng preprocessng s carred out for EEG sgnals frstly; and then, the relatve wavelet energy s extracted as a feature vector from the Channels C 3 and C 4 by use of the algorthm for relatve wavelet energy, and pattern recognton s carred out by use of the radal bass functon neural network (RBFNN). Smulaton results show that the proposed method acheves good classfcaton results. Introducton Bran-Computer Interface (BCI) s a short form of the nterface between bran and computer. It reflects the conscousness of people through the EEG sgnals but not the normal physologcal output pathways consttuted by perpheral nerves and muscles, so t s a new way of communcaton and control []. There are many knds of EEG sgnals generally used n BCI. As the modes of generaton are dfferent, EEG sgnals are manly dvded nto two types,.e., nduced EEG sgnals and spontaneous EEG sgnals []. Motor magery EEG sgnal s a knd of spontaneous EEG sgnal whch s the most wdely used. The research on BCI proects based on motor magery EEG sgnals was frstly carred out by Pfurtscheller et al. The result showed that, durng unlateral lmb movement or motor magery, Event related desynchronzaton(erd) was produced n the contralateral bran areas and event related synchronzaton (ERS) [] was produced n the pslateral bran area. Specfcally, durng the experment, the cerebral cortex where channels C 3 and C 4 were located represented the area of rght and left hands motor magery n the bran; durng the motor magery, ERD/ERS n the areas on both sdes of the bran was symmetrcal,.e., durng the rght hand motor magery, ERD occurred on the cerebral cortex for motor magery n the left area of the bran whle ERS occurred on the cerebral cortex n the rght area of the bran; and vce versa. Further research showed that durng the motor magery, obvous ERD/ERS phenomenon occurred manly wthn the frequences of Rhythm μ (8-Hz) and Rhythm β (8-4Hz) of the correspondng cerebral cortex for motor magery n the contralateral bran areas, and at these tme ponts, the energy n the frequences of Rhythm μ and Rhythm β decreased (or ncreased) [3-6]. A complete BCI system manly ncludes acquston, preprocessng, feature extracton, pattern recognton, control equpment and perpheral equpment for EEG sgnals [7], among whch feature extracton and pattern recognton are the most crtcal parts. The man contents of ths paper nclude feature extracton and pattern recognton of EEG sgnals by use of wavelet transformaton and fuzzy a Correspondng author: tanuan@63.com The Authors, publshed by EDP Scences. Ths s an open access artcle dstrbuted under the terms of the Creatve Commons Attrbuton Lcense 4. (

2 BIO Web of Conferences 8, 3 (7) DOI:.5/ boconf/783 ICMSB6 neural network, takng the ERD/ERS phenomena correspondng to left and rght hands motor magery as the dfferentatng crtera n combnaton wth the characterstcs. Feature Extracton The purpose of feature extracton s to transform the preprocessed EEG sgnals nto dfferent feature vectors representng dfferent conscousness tasks. Feature extracton s the most crtcal step n the dentfcaton of BCI sgnals, drectly mpactng on the desgn for the later classfer. In ths paper, the relatve energy extracted by use of the wavelet transformaton technology from the characterstc frequency band s manly used as the characterstc value.. Source of Expermental Data The expermental data used n ths paper comes from Dataset IIIb of BCI Competton III held n 5. It was provded by Graz Unversty of Technology n Austra and dvded nto two categores: left hand motor magery and rght hand motor magery. The EEG data was collected from three healthy subects, denoted as O3, S4 and X. Among them, O3 had 64 peces of data, and the expermental task was to adopt the method of vrtual realty to control the feedback tems on the screen by use of left and rght hands motor magery; S4 and X had,8 peces of data respectvely, and the expermental task was to control the vrtual ball through motor magery, movng the ball to the left or rght end of the screen. All the data were collected from the EEG sgnals recorded n two-lead mode at electrodes C 3, C Z and C 4, among whch the electrodes C 3 and C 4 were located n the functonal area of the prmary sensormotor cortex of the bran, and they could reflect the most effectve nformaton about the changes n bran status durng the left and rght hands motor magery of subects; as a reference electrode, the samplng frequency of C 4 s 5Hz after.5-3hz flterng. In order to reduce the amount of calculaton, the expermental data was selected from S4 for ths tme, and there was a total of 7677 x peces of data [8]. The complete expermental procedure s descrbed as follows: the duraton of each experment s 8s, and t s dvded nto three stages. Durng the frst s, the subects have a rest to keep a relaxed state. When T=s, they begn to enter the stage of magery; when T=3s, a short beep sounds and t suggests that the subects conscousness s beng concentrated and the motor magery s begnnng; meanwhle, the cursor on the screen moves left or rght, whch wll last for the whole process of the experment, whle the subects magne the movements as planned accordng to the movement of the cursor. When T=4s, the stage of nformaton feedback begns, and the duraton s 4s (wthn 4s-8s). The classfcaton results of the feedback stage are gven to the subects from the classfer. The expermental process s shown n Fgure. Fgure. Experment Process of Left and Rght Hands Motor Imagery

3 BIO Web of Conferences 8, 3 (7) DOI:.5/ boconf/783 ICMSB6. Preprocessng of Expermental Data The research shows that, durng the acquston process of EEG sgnals, false nformaton such as ECG, EMG, REM, blnk and power frequency nterference wll nterfere wth the observaton and analyss aganst the EEG sgnals by people. Therefore, before the feature extracton, n ths paper, denosng s carred out for the orgnal EEG data by use of the wavelet transformaton technology and collect the total a trgger 8 seconds n the data. The comparson of EEG sgnals before and after denosng s shown n Fgure Sample Ponts Sample Ponts (a) (b) Fgure. Comparson of EEG Sgnals Before and After Denosng (a) Before Denosng (b) After Denosng.3 Feature Extracton Based on Wavelet Transformaton.3. Basc Theory of Wavelet Transformaton Wavelet transformaton s a mult-scale sgnal analyss method. In both tme and frequency domans, t represents the local sgnal characterstcs and s used for mult-resoluton analyss, so t s applcable to analyze the transent and tme-varyng characterstcs of non-statonary sgnals. EEG sgnals are a knd of complex nonlnear and non-statonary sgnal. Therefore, n ths paper, wavelet transformaton s used for the analyss and feature extracton of EEG sgnals [9]. It s assumed that the wavelet bass functon s ψ ( ), then, ψ a, b t b ( t) = ψ ( ) a a a, b t where, a represents the contracton-expanson factor or scale factor, and b represents the shft factor. Dscrete bnary method s adopted for the calculaton of a and b, and the followng dscrete bnary wavelet functon can be obtaned: ψ ψ, k ( t) = ( t k) The dscrete wavelet transformaton coeffcent: DWTx (, k) = f ( t) ψ *( t k) dt (3) The nverse transformaton: ( t) = DWTx (, k) ( = k= f ψ t k) (4) It s assumed that the length of EEG data segment used for feature extracton s m x N ponts, then, Formula (4) s converted nto: N m m = k = t k) f ( t) = DWT (, k) ψ ( (5) x Accordng to the Mallat decomposton algorthm, the fnte layer decomposton for f (t) s () () 3

4 BIO Web of Conferences 8, 3 (7) DOI:.5/ boconf/783 ICMSB6 obtaned: f t) = A + D ( (6) = where, represents the decomposton layer, A represents the low frequency approxmate component coeffcent, and D represents the hgh frequency detal coeffcent at dfferent scales. Under the premse that the wavelet bass functon s orthogonal bass functon, accordng to the energy conservaton property of wavelet transformaton, the formula for wavelet energy s obtaned: E tot = t f ( t) In combnaton wth Formula (6), t can be further expressed as: Etot = A + D = = + = Then, the relatve wavelet energy[] s: p E D = = + Etot = D (7) D (8) (9).3. Analyss of Smulaton Results As the wavelet db possesses excellent local characterstcs n both tme and frequency domans [], n ths paper, wavelet db4 n the tme length of s s selected as the wavelet functon to carry out 4-layer decomposton for the preprocessed EEG sgnals; the obtaned wavelet coeffcents of the detaled part are D, D, D 3, and D 4, and that of the approxmate part s A 4. The correspondng frequency ranges are: 8-3.5Hz, 3.5-9Hz, 9-4.5Hz, and 4.5-3Hz, respectvely. And then, for the EEG sgnals of Channels C 3 and C 4, the relatve wavelet energes wthn 4-7s are selected as the characterstc values of left and rght hands motor magery. The smulaton results are shown n Fgures3 and Sample ponts Sample ponts (a) (b) Fgure 3. Relatve Energes n Channels C3 and C4 durng Left Hand Motor Imagery(a) Relatve Energy n Channel C3 (b) Relatve Energy n Channel C4 4

5 BIO Web of Conferences 8, 3 (7) DOI:.5/ boconf/783 ICMSB Sample Ponts (a) Sample Ponts (b) Fgure 4. Relatve Energes n Channels C3 and C4 durng Rght Hand Motor Imagery(a) Relatve Energy n Channel C3 (b) Relatve Energy n Channel C4 From Fgure3, t can be seen that durng left hand motor magnaton, obvous ERD/ERS occurs at the sample ponts of 6-7 n Channels C 3 and C 4 ; from Fgure 4, t can be seen that durng rght hand motor magnaton, obvous ERD/ERS occurs at the sample ponts of 7-8 n Channels C 3 and C 4. 3 Pattern Recognton Pattern recognton s a very mportant part of BCI system, and the core part of pattern recognton s desgn of classfer. The common classfcaton algorthms manly nclude lnear dscrmnant method, support vector machne-based method, and artfcal neural network method. The artfcal neural network method s wdely used because of ts smple applcaton, convenent parameter selecton, and hgh accuracy of classfcaton results. Accordng to the dfferent neural network models, artfcal neural network s dvded nto feed-forward neural network, feedback neural network, and self-organzng neural network, among whch feed-forward neural network ncludes perceptron network, BP network, lnear neural network, and radal bass functon neural network. For radal bass functon neural network, the nearest neghbor clusterng algorthm s adopted and t has such advantages as short learnng tme, small amount of calculaton and good performance, so n ths paper, RBF neural network s selected for the classfcaton of left and rght hands motor magery EEG sgnals. 3. Basc Theory of RBF Neural Network RBF neural network s composed of nput layer, hdden layer and output layer. For the hdden layer, radal bass functon s adopted as the exctaton functon. Gauss's functon s taken as the radal bass functon, and s recorded as: where x c g = exp( ) σ The output form s: m y = w + = wg w s the weght from the hdden layer to the output layer. () () c and σ are the center and 5

6 BIO Web of Conferences 8, 3 (7) DOI:.5/ boconf/783 ICMSB6 the varance of the radal bass functon, respectvely. m s the number of the centers. Accordng to the emprcal formula, c can be, σ can be, and w can be -.5. In order to smplfy the calculaton, n ths paper, c s taken as, σ s taken as, and w s taken as Analyss of Classfcaton Result Analyss In order to reduce the amount of calculaton, n ths paper, 3 sets of characterstc data from Channels C 3 and C 4, among whch sets of data s taken as tranng samples and the other sets of data s taken as test samples. The features of the tranng samples are used as the nput data of RBF neural network, and the thnkng task categores of the tranng sample are used as the output data of RBF neural network, wth the category number recorded as "" and "". The rate of correct dentfcaton of the test samples (.e., the rato of the number of correctly dentfed samples to the total sample number) s used as the evaluaton ndex. Accordng to the above algorthm, ths paper uses Matlab to comple the correspondng program to realze recognton of left and rght hands motor. Test for tmes and the overall recognton rate approaches 9%. 4 Concluson Feature extracton and feature classfcaton are one of the key problems of BCI system. In ths paper, through the research on left and rght hands motor magery EEG sgnals, ERD/ERS phenomenon s taken as the dfferentatng crtera, relatve wavelet energy n the specal frequences of Channels C 3 and C 4 s used as feature vectors of left and rght hands motor magery EEG sgnals, RBF neural network s used for features classfcaton, and smulaton s carred out n combnaton wth the data from BCI Competton III held n 5. The results show that good classfcaton effects are acheved. The research on BCI, as a new human-computer nterface (HCI) mode, needs multdscplnary cooperaton. At present, the BCI system s not perfect and the correspondng theory and algorthm are not mature, so further research s requred. Wth the contnuous mprovement and perfecton of technology, the BCI system wll eventually come out of the laboratory and be wdely used n real lfe []. Acknowledgments The work s supported by Proect of Shandong Provnce Hgher Educatonal Scence and Technology Program, Chna(Grant No. 3LN4) References... R. Wolpaw, N. Brbaumer, W.. Heetderks,et al. Bran-computer nterface technology: a revew of the frst nternatonal meetng[],ieee Trans.Rehab.Eng.,,8(): P.Mssonner, G.Gold, F.R.Herrmann, et al.decreased theta event-related synchronzaton durng workng memory actvaton s assocated wth progressve mld cogntve mparment.[]. Dementa & Geratrc Cogntve Dsorders, 6, (): B.Aknc,N.G.Gencer,Onlne cue-based dscrmnaton of left/rght hand movement magnaton[],bomedcal Engneerng Meetng,,-4 4. S. Shahd, R. K. Snha, A. G. Prasad, Mu and beta rhythm modulatons n motor magery related post-stroke EEG: a study under BCI framework for post-stroke rehabltaton[],bmc Neuroscence,,():- 5. E. Thomas,, Frutet, M, Clerc, Investgatng bref motor magery for an ERD/ERS based 6

7 BIO Web of Conferences 8, 3 (7) DOI:.5/ boconf/783 ICMSB6 BCI[],Internatonal Conference of the IEEE Engneerng Medcne,,(4): E. Thomas,, Frutet, M. Clerc, Combnng ERD and ERS features to create a system-paced BCI[],ournal of Neuroscence Methods,3,6(): R. Wolpaw, N. Brbaumer, D.. McFarland, et al. Bran computer nterfaces for communcaton and control[]. Clncal neurophysology,, 3(6): BCI Competton III:5, URL< > 9. asmn Kevrc, Abdulhamt Subas, Comparson of sgnal decomposton methods n classfcaton of EEG sgnals for motor-magery BCI system[],bomedcal Sgnal Processng and Control,7,3: A.S.Sankar, S. S. Nar, V. S. Dharan, P. Sankaran, Wavelet Sub Band Entropy Based Feature Extracton Method for BCI[], Proceda Computer Scence,5,46: M.H.Alomar, E. A. Awada, O. Youns, Subect-Independent EEG-Based Dscrmnaton Between Imagned and Executed, Rght and Left Fsts Movements[], European ournal of Scentfc Research, 4,8(3): A. Ortzrosaro, A. Adel, Bran-computer nterface technologes: from sgnal to acton[], Revews n the Neuroscence, 3, 4(5):