Wavelet Based Feature Extraction Scheme of Electroencephalogram

Transcription

1 ISSN (Online) : ISSN (Print) : An ISO 397: 007 Certified Organization, Volue 3, Special Issue 1, February 014 Wavelet Based Feature Extraction Schee of Electroencephalogra Mr. C. E. Mohan Kuar 1, Mr. S. V. Dharani Kuar Assistant Professor, GRT Institute of Engineering & Technology, Tailnadu, India 1, Abstract- The Electroencephalograph (EEG) signals is one of the ost widely used in the bioinforatics field due to its rich inforation about huan tasks. The Electroencephalogra is a neuronal activity that represents the electrical activity of the brain. The uses of EEG signals in the field of Brain coputer Interface (BCI) have obtained a lot of interest with diverse applications ranging fro edicine to entertainent. BCI is designed using EEG signals where the subjects have to think of only a single ental task. The specific features of EEG are used as input to Visual Evoked Potential (VEP) based Braincoputer Interface or self paced BCIs (SBCI) for counication and control purposes. This work proposes schee to extract feature vectors using wavelet transfor as alternative to the coonly used Discrete Fourier Transfor (DFT). Brain Coputer Interface is a direct connection between the brain and a coputer, without using any of the brains natural output pathways. Visually-evoked Potentials extracted fro the electroencephalographic activity in the visual cortex recorded fro the overlying scalp. Wavelets are powerful candidates for decoposition, feature extraction, and classification of non- stationary EEG signals for BCI applications. Wavelet Transfor (WT) is superior to Discrete Fourier Transfor due to its high localization in tie and frequency doain. The ain objective of Wavelet Transfor usage is to localize the artifact coponent. Index Ters: Electro-Encephalogra (EEG), Brain- Coputer Interface (BCI), Wavelet Transfor (WT), Continuous Wavelet Transfor (CWT), Discrete Wavelet Transfor (DWT), Short tie Fourier transfor (STFT), visually Evoked Potential (VEP), Discrete Fourier Transfor (DFT), Multi-resolution Analysis (MRA). I. INTRODUCTION A Brain Coputer Interface (BCI) or Brain Machine Interface (BMI) has been proposed as an alternative counication pathway, bypassing the noral cortical-uscular pathway. BCI is a syste that provides a neural interface to substitute for the loss of noral neuronal-uscular outputs by enabling individuals to interact with their environent through brain signals rather than uscles. Most BCI research is aied towards developing tools for patients with severe otor disabilities and paralyzes. This group of potential users could particularly benefit fro BCI technology, since output pathways that are norally eployed by the brain can no longer be used. Brain Coputer Interface is also referred to as Enhanced for of neuro-prosthestic support syste was conceived as counication interface between achines (usually a coputer) and the brain of a user. They should perit the use to perfor a certain task, usually without ipleenting any otor action. This iplies that neural ipulses generated by the user s brain are detected, elaborated and utilized by the achines approxiately in real tie, to perfor definite tasks. Electroencephalograph (EEG) represents coplex irregular signals that ay provide inforation about underlying neural activities in the brain. Electroencephalogras are recordings of the tiny electrical potentials (generally less than 300µV) produced by the brain. The brain waves recorded fro the scalp have sall aplitudes of approxiately 100µV. The frequencies of these brain waves range fro 0.5 to 100Hz, and their characteristics are highly dependent on the degree of activity of the cerebral cortex. The EEG spectru contains soe characteristic Copyright to IJIRSET 908

2 ISSN (Online) : ISSN (Print) : An ISO 397: 007 Certified Organization, Volue 3, Special Issue 1, February 014 wavefors that fall priarily within four frequency bands: delta (1-4Hz), theta (4-8Hz), alpha (8-13Hz), and beta (13-30Hz). The first step in BCI systes is the data collection and filtering, the filters are designed in such a way not to introduce any change or distortion to the signals. High pass filter with a cut-off frequency of usually less than 0.5Hz are used to reove the disturbing very low frequency coponents such as those of breathing. On the other hand, high frequency noise is itigated by using low pass filter with a cut-off frequency of approxiately 40-70Hz. BCI is coposed of signal collection and processing, pattern identification and control systes. Signal Acquisi tion Pre- Processi ng Feature Extracti on Classific ation Fig.1. The Block diagra of the BCI work Fig.1. illustrates the part of the work. The EEG signals are collected and represented using special filters by the ethod of signal acquisition, then the acquired signal is preprocessed where the artifact coponent (undesired coponent) is reoved fro the desired coponents, then the EEG feature are extracted fro the feature vector using several ethods and finally those features are classified depending on the frequencies for ental task they represent. There are few new concepts in the design of EEG easureent systes like iniaturized, battery-powered front-end close to patient, with fiber optic data transfer to the signal processing PC. A group of ost iportant authors in the field of non-invasive BCIs gave the list of goals iportant for future progresses of these systes. Future progress will depend on: Identification of those signals, whether evoked potential, spontaneous rhyths, or neuronal firing rates, that user are best able to control. Developent of training ethods for helping users to gain and aintain that control. Delineation of the best algoriths for translating these signals into device coands. Attention to eliinate if artifacts as electro-y graphic and electro-oculographic activity. Adoption of precise and objective procedures for evaluating BCI perforance. II. FEATURE EXTRACTION Feature extraction ethodologies analyze signals to extract the ost proinent features that are representative of the various classes of signals. The ain ai of feature extraction is to obtain the further inforation s fro the raw signal. The neuron is the basic structural and functional unit of nervous syste. The brain's electrical charge is aintained by billions of neurons. Neurons are electrically charged (or "polarized") by ebrane transport proteins that pup ions across their ebranes. Neurons are constantly exchanging ions with the extracellular ilieu, for exaple to aintain resting potential and to propagate action potentials. Ions of like charge repel each other, and when any ions are pushed out of any neurons at the sae tie, they can push their neighbors, who push their neighbors, and so on, in a wave. This process is known as volue conduction. When the wave of ions reaches the electrodes on the scalp, they can push or pull electrons on the etal on the electrodes. Since etal conducts the push and pull of electrons easily, the difference in push or voltage between any two electrodes can be easured by a volteter. Recording these voltages over tie gives us the EEG. Fro the extracted feature the artifacts coponents are reoved by preprocessing stage then this feature is given as input to the classifier and the classified signal is given to the soe of the applications like wheel chair, coputer cursor oveents etc,. Since the features have very sall aplitude range special aplifier and booster circuits are used to enhance the strength of signals. A Special for of diensionality reduction in iage processing is transforing the external stiulus or sensations into the set of features in the surface of brain. To enable brain-coputer interface construction an efficient ethod of feature extraction fro EEG signal is needed. This paper proposes a feature extraction ethod based on higher order statistics calculated for the details of discrete wavelet transfor (DWT) of EEG signal. Liiting the nuber of electrodes is supposed to siplify the use of the interface and reduce the cost of the EEG signal aplifier. This would also facilitate the analysis, processing and classification of signals. The ethods reinforce the non-stationary EEG concept and Copyright to IJIRSET 909

3 ISSN (Online) : ISSN (Print) : An ISO 397: 007 Certified Organization, Volue 3, Special Issue 1, February 014 call for the necessity of extracting ore inforation to understand the brain signals and its dynaics. III. METHODOLOGY Tie-doain wavelets are siple oscillating aplitude functions of tie. So are the sine and cosine waves of Fourier analysis. However, unlike sine and cosine waves which are precisely localized in frequency but extend infinitely in tie (sines and cosines have definite single frequencies, e.g., 40Hz, constant for all tie); wavelets are relatively localized in both tie and frequency. They have large fluctuating aplitudes during a restricted tie period and are very low aplitude or zero aplitude outside of that tie range. That is, wavelets are said to be supported over a restricted doain of tie if the bulk of their energy is restricted to that tie period and are said to be copactly supported if all of their energy is restricted to a specific doain of tie. In this paper, the non-paraetric ethod of feature extraction based on ulti-resolution analysis of Wavelet Transfor (WT) is introduced. The EEG signal is non-stationary, tie doain signal and the signal energy distribution is scattered. The signal features are buried away in the noise. In order extract the features, the EEG signal is analyzed to give a description of the EEG energy as the function of tie or/and frequency. The joint tie-frequency resolution obtained by WT akes it as a good candidate for the extraction of details as well as approxiations of the signal which cannot be obtained either by Fast Fourier Transfor (FFT) or by Short Tie Fourier Transfor (STFT).The nonstationary nature of EEG signals is to expand the onto basis functions created by expanding, contracting and shifting a single prototype function (, the other a, b wavelet), specifically selected for the signal under consideration. The other wavelet function (t) is a, b given as a, b ( t) 1 t b a a Where a, b R, a> 0, and R is the wavelet space. Paraeters a and b are the scaling factor and shifting factor respectively. The only liitation for choosing a prototype function as other wavelet is to satisfy the adissibility condition. Where ( ) ( ) Copyright to IJIRSET C d is the Fourier transfor of a, b (t).the wavelet function has been chosen due to their near optial tie-frequency localization properties. The CWT is not very efficient since localization of artifact coponents in tie and frequency is not possible and also it consues ore tie. The inforation it displays at closely spaced scales or at closely spaced tie points is highly correlated and therefore unnecessarily redundant for any analytic purposes. It is also tie consuing to copute directly. Although the CWT has soe advantages, there are any applications where a ore efficient and coputationally sipler wavelet analysis is desirable. Such an analysis, known as the Discrete Wavelet transfors. Wavelet transfor decoposes a signal into a set of basis functions. These basis functions are called wavelets. Wavelets are obtained fro a single prototype wavelet y (t) called other wavelet by dilations and shifting. a, b ( t) 1 t b a a Where a is the dyadic scaling paraeter and b is the dyadic shifting paraeter. A. Wavelet Analysis Wavelet analysis refers to a growing class of signal processing techniques and transfors that use wavelets and related functions called wavelet packets to efficiently easure and anipulate such non-stationary signals. These detail functions can isolate all scales of wavefor structure, fro the largest to the sallest pattern of variation in tie and space that is available in the neuroelectric data set. Consequently, wavelet analysis provides flexible control over the resolution with which neuroelectric coponents and events can be localized in tie, space, and scale. The set of wavelet functions is usually derived fro the initial (other) wavelet h (t) which is dilated by a value, translated by constant b = k noralized so that h, k ( t) 1 t b h a a 1 h and t k

4 ISSN (Online) : ISSN (Print) : An ISO 397: 007 Certified Organization, Volue 3, Special Issue 1, February 014 For integer values of, k and the initial wavelet defined either by the solution of a dilation equation or by an analytical expression. Both continuous and discrete signals can be then approxiated in the way siilar to Fourier series and discrete Fourier transfor. In case of a sequence n n 0 x having N g values it is possible to evaluate its expansion. x(n)= a s1 s1 a 1k h( n k) 0 s 0 k0 Wavelet transfor coefficients can be organized in a atrix T with its nonzero foring a triangle structure with each its row corresponding to a separate s dilation coefficient. The set of N decoposition coefficients a j j 0 of the wavelet transfor is defined in the way forally close to the Fourier transfor but owing to the general definition of wavelet functions they can carry different inforation using the orthogonal set of wavelet functions they are oreover closely related to the signal energy. The initial wavelet can be considered as a passband filter and in ost cases half-band filter covering the noralized frequency band (0.5, 0.5). Wavelet dilation by the factor a corresponds to a pass-band copression. This general property can be deonstrated for the haronic wavelet function and the corresponding scaling function by expressions. 1 j t j t h( t) ( e e ) j t 1 j t l ( t ) ( e 1) t As both these functions are odified by the scaling index = 0, 1 the wavelet is dilated and its spectru copressed resulting in tie and frequency doain representation presented in. Siilar approach can be also applied for other wavelet functions defined in either analytical or recurrent for. The set of wavelets define a special filter bank which can be used for signal coponent analysis and resulting wavelet transfor coefficients can be further applied as signal features for its classification. Signal decoposition prefored by a pyraidal algorith using the fast Fourier transfor. B. Decoposition The output of Wavelet transfored Signals are having two Co-Efficient called as Approxiate Co- Efficient and Detail Co-Efficient. The Approxiation Co-Efficient corresponds to Low frequency Coponent and the Detail Co-Efficient corresponds to High frequency Coponent of a Wavelet. This process is called Signal Decoposition Neuroelectric Signal Input Wavelet Coefficient(For detail function) Scaling Function Coefficients (For low resolution signal) Discrete Wavelet Decoposition High Pass Low Pass Fig.. Basic recursive Generate pyraidal Higher Resolution odel Signals schee for DWT At the beginning of a DWT coputation, a neuro-electric wavefor like EEG containing n saples is run through the high and low pass filters. The output each filter is a series of n wavelet coefficients. Every other coefficient is of discarded fro the series, leaving n/ coefficients for each filter output. This process of discarding alternate coefficients is known as down sapling and is indicated in the figure by the downward pointing arrow and adjacent sybol. The output of the high pass filter is the set of DWT wavelet coefficients associated with all of the discrete wavelets at the sallest single scale available for the particular digitized neuro-electric wavefor that Copyright to IJIRSET Generate More Signal Wavelet Coefficients(For detail function) Multi-resolution Wavelet Reconstruction High Pass Low Pass Scaling Function Coefficients (For low resolution signal) Reconstruc ted Neuroelectric Signal

5 ISSN (Online) : ISSN (Print) : An ISO 397: 007 Certified Organization, Volue 3, Special Issue 1, February 014 went into the filter. This output captures all of the high frequency energy in the wavefor. The output of the low pass filter is the set of DWT coefficients associated with a set of copanion functions called scaling factor. This output captures all of the low frequency energy of the wavefor. C. Wavelet vs Fourier Transfor FT provides a signal which is localized only in the frequency doain. It does not give any inforation of the signals of the signals in tie doain. Basis functions of the wavelet transfor are sall waves located in different ties. They are obtained using scaling and translation of a scaling function and wavelet function. Therefore wavelet is localized in both tie and Frequency doain. In addition wavelet transfor provides ultiresolution systes. If a signal has a discontinuity, FT produces any coefficients with large agnitude. But wavelet transfors generates a few significant co efficient around the discontinuity on linear approxiation is a ethod of bench ark the approxiation power of a transfor. In non linear approxiation we keep only a few signification coefficient of a signal and set the rest to zero. Then we reconstruct the signal using the signal coefficients. Wavelet produces a few signification coefficients for the signals with discontinuities. Thus we obtain better results for wavelet transfor non-linear approxiation when copared to Fourier transfors. D. Electrode Montage Schee Fig. 3. Electrode Montage Schee The details of the recording setup are given below for recording the electrical activity of the brain. This set up record conventional scalp EEG and the recording was obtained by placing silver chloride electrodes on the scalp with a conductive gel or paste, usually after preparing the scalp area by light abrasion to reduce ipedance due to dead skin cells. These systes use electrodes, each of which is attached to an individual wire. Electrode locations and naes are specified by the International 10 0 syste. This ontage schee ensures that the naing of electrodes is consistent. This setup eploys 19 recording electrodes (plus ground and syste reference). Each electrode is connected to one input of a differential aplifier (one aplifier per pair of electrodes); coon syste reference electrode is connected to the other input of each differential aplifier. These aplifiers aplify the voltage between the active electrode and the reference (typically 1, ,000 ties, or db of voltage gain). This EEG then filtered and the EEG signal is output as the signals were stored. IV. RESULT The Wavelet transfor decoposes the EEG signal to yield the approxiation co-efficients and detail co-efficients. These co-efficients were used as input to copute the energy of features. These values enable to extract the features associated with stiuli. V. CONCLUSION AND DISCUSSION While the electroencephalograph was invented nearly a century ago, it is only recently that researchers have begun to apply it to probles outside the edical and neuroscience doain such as the BCI systes. The role of signal processing is crucial in the developent of a BCI syste. We evaluated the ability of the frequency analysis (FFT), tie frequency analysis (STFT) and space tie frequency analysis over ultiple electrodes to discriinate between EEG signals. The Wavelet Transfor decoposes the EEG signal to yield the approxiation between coefficients, detail coefficients. These values enable to extract the features associated with stiuli. This paper presents the use of wavelet transfor for a given feature extraction associated with an electrode pair. Wavelet Transfor has ore advantage over the STFT and FT for the localization of artifact coponent in tie and frequency doain and also Copyright to IJIRSET 91

6 ISSN (Online) : ISSN (Print) : An ISO 397: 007 Certified Organization, Volue 3, Special Issue 1, February 014 wavelet transfor adopts the variable window length which can deal with the non-stationary signal with a variable frequency rate. Hence the wavelet based feature extraction is considered as the best when copared with any other based transfor based feature extraction. Matheatical basis of the wavelet transfor has also proved that EEG analysis based on wavelet transfor coefficients can be used very efficiently for the estiation of EEG features. VI. FUTURE WORK Further work suggestions include finding the best cobination of channels in the case of space-tiefrequency analysis for specific task. Also one of our suggestions for the future is building the hardware odel for the EEG feature extraction and classification syste using the Field Prograable Gate Array (FPGA) REFERENCES [1] Aleˇ sproch azk, Jaro ırkukal, Wavelet Transfor use for Feature Extraction and EEG Signal Segents Classification, Institute of Cheical and Control Engineering. [] Maan M. Shaker, EEG Waves Classifier using Wavelet Transfor and Fourier Transfor, International Journal of Biological and Life Sciences 1: 005. [3] G.Saravana Kuar, Dr. S. Ravi, Feature Extraction Schee for Brain-Coputer Interface using Wavelet Transfor, International Journal of Research and Reviews in Coputer Science (IJRRCS), Vol., No.1, March 011, pp [4] Abdul- Bary Raoquf Suleian, Toka Abdul- Haed Fathei, Features Extraction Techniques of EEG signal for BCI Applications, Coputer and Inforation Engineering Departent. [5] Luis Fernando Nicolas-Alonso, Jaie Goez-Gil, Brain Coputer Interfaces, a Review, Departent of Signal Theory, Counications and Teleetric Engineering, University of Valladolid, 31 January 01, pp [6] Febo Cincotti, Donatella Mattia, Fabio Aloise, Sion Bufalari, Laura Astolfi, Fabrizio De Vico Fallani, Andrea Tocci, Luigi Bianchi, Maria Grazia Marciani, Shangkai Gao, Jose Millan, Fabio Babiloni, High resolution EEG techniques for braincoputer interface applications". [7] Kenji Kanasaku, Brain Machine Interfaces for Persons with Disabilities, Systes Neuroscience Section, Departent of Rehabilitation for Brain Functions, Research Institute of National Rehabilitation Center for Persons with Disabilities (NCRD), pp [8] Nikhil Raachandran, A.K. Chellappa, Feature Extraction fro EEG using Wavelets: Spike Detection Algorith. [9] Mai S. Mabrouk, Non- Invasive EEG-based BCI syste for Left or Right Hand Moveent, Majlesi Journal of Electrical Engineering, Vol. 5, No. 3, Septeber 011, pp [10] Onder aydeir, Teel Kayikcioglu, Wavelet Transfor Based Classification of Invasive Brain Coputer Interface Data, Radio Engineering, Vol. 0, no.1, April 011, pp [11] Ayhan,T; Seker, S, BCI odeling using Discrete Wavelet Transfor, Innovations in intelligent Systes and Applications (INISTA), 011 International Syposiu on June 011, pp [1] Maatha, M.N.; Raachandran, S.; Chandrasekaran, M, BCI- A counication for physically challenged people, Counication Software and Networks (ICCSN), 011 IEEE 3rd International Conference on 7-9 May 011, pp [13] Murugappan, Nagarajan Raachandran, Yaacob Sazali, Classification of huan eotion fro EEG using discrete wavelet transfor, J. Bioedical Science and Engineering, 010, 3, pp [14] Ali S. Alejrad, Huan Eotions Detection using Brain Wave Signals: A Challenging, European Journal of Scientific Research, Vol.44 No.4 (010), pp [15] Azi, M.R; Ain, M.S.; Haque, S.A., Abia, M.N.; Shoeb, M.A, Feature extraction using EEG waves, Signal processing (ICSPS), 010 nd international conference on 5-7 July 010, V [16] Dietar Dietrich, Roland Lang,Dietar Bruckner, eorge Fodor, Brit Muller, Liitations, Possibilities and Iplications of Brain-coputer Interfaces, 010, pp [17] Haider Hussein Alwasiti, Ishak Aris and Adznan Jantan, Brain Coputer Interface Design and Applications: Challenges and Future, World applied science journal 11(7): pp , 010. [18] Mohaad Reza Nazari Kousarrizi, Abdolreza Asadi Ghanbari, Mohaad Teshnehlab, Mahdi Aliyari Shorehdeli, Ali Gharaviri, Feature Extraction and Classification of EEG Signals Using Wavelet Transfor, SVM and Artificial Neural Networks for Brain Coputer Interfaces, 009 International Joint Conference on Bioinforatics, Systes Biology and Intelligent Coputing. [19] Aihua Zhang, Bin Yang, Ling Huang, Feature Extraction of EEG Signals Using Power Spectral Entropy, 008 international conference on bioedical engineering and inforatics. [0] Oerhodzic, S. Avdakovic, A. Nuhanovic, K. Dizdarevic, energy Distribution of EEG Signals: EEG Signal Wavelet- Neural Network Classsifier, International Journal of Biological and Life Sciences 6: Copyright to IJIRSET 913