1992-8645 EEG SIGNAL IDENTIFICATION BASED ON ROOT MEAN .
Journal of Theoretical and Applied Information Technologyst31 August 2014. Vol. 66 No.3 2005 - 2014 JATIT & LLS. All rights reserved.ISSN: 1992-8645www.jatit.orgE-ISSN: 1817-3195EEG SIGNAL IDENTIFICATION BASED ON ROOT MEANSQUARE AND AVERAGE POWER SPECTRUM BY USINGBACKPROPAGATIONHINDARTO1, MOH. HARIADI2, MAURIDHI HERY PURNOMO3Department of Informatics, Universitas Muhammadiyah Sidoarjo, Sidoarjo, Indonesia1,2,3Department of Electrical Engineering, Institut Teknologi Sepuluh Nopember, Surabaya, IndonesiaEmail :1hindarto@umsida.ac.id, e development of user interface for game technology has currently employed human centered technologyresearches in which EEG signal that utilizes the brain function has become one of the trends. The presentresearch describes the identification of EEG Signal by segmenting it into 4 different classes. Thesegmentation of these classes is based on Root Mean Square (RMS) and Average Power Spectrum(AVG), employed in feature extraction. Both Root Mean Square (RMS) and Average Power Spectrum(AVG) are employed to extract features of EEG signal data and then used for identification, by which aBackPropagation method is employed. The experiment,done with 200 tested signal data file, demonstratesthat the identification of the signal is 91% accurate.Keywords : Root Mean Square, Average Power Spectrum, BackPropagation, EEG signal.1.INTRODUCTIONElectroEncephaloGraphy (EEG) is a recording ofelectrical activity along the scalp. EEG voltagefluctuation is resulted from ionic stream in theneuron of the brain [1]. The response of the brain toexternal and internal stimulus can be recognized inclassical EEG, Event Related Potential phaselocked (ERP), non-locked phase (induction)reactivity. ERP can be clearly seen in the averageresponse EEG time which is properly synchronized;however, non-locked phase activity is cancelledfrom the average. Then, the classical induced bydesynchronization (ERD) and Event RelatedSynchronized (ERS) were calculated throughout thefollowing procedures : the most reactive frequencybands are selected throughtrial-and-error procedure;furthermore, the signal and band-pass were filteredin the bands; and then, the results of the calculationwere squared then calculated to yield the ation system that translates the directaction of user’s brain activity into signal andcontrol command. BCI is able to spell, browse inthe internet, control robotic devices, and/or performother tasks by using thoughts [3][4][5][6][7][8][9].By using appropriate design, input combined withextractor feature and classifier is a suitableframework for motor imagery. Different attribute ofEEG signal has been used as BCI input, such asrhythm (8-12 Hz) and beta rhythm (18-25 Hz),while ERP, for instance P300, established visualevoked response and or Slow Cortical Potential(SCP) [10][11][12].The present research attempts to focus on thediscussion of two major points namely: first, FastFourier Transform (FFT) method used to measurethe strength level of each signal of EEG sampledata, and Signal estimated by using Root MeanSquare (RMS); second, Average Power Spectrum(AVG) that is employed, and signal pattern ofEEG identified as the subjects would be inquired toimagine the movement of left finger, left arm andright arm. The paper would be organized in thefollowing sequences, as the following; section 1 isthe introduction, section 2 describes the materialsand methods, Section 3 contains the result anddiscussion, and the last section would be theconclusion.2. MATERIALS AND METHODS2.1. Description of datasetThis present study was carried out by doing anexperiment involving ten subjects of the study whowere coming from General Laboratory forBiomedical Engineering, Experimental Departmentof Informatics, University of Kyushu. The selectedsubjects were those who really found to have goodhealth condition as they would be involved in an782
Journal of Theoretical and Applied Information Technologyst31 August 2014. Vol. 66 No.3 2005 - 2014 JATIT & LLS. All rights reserved.ISSN: 1992-8645www.jatit.orgexperiment that required them to imagine themovements of left finger (up cursor), right finger(down cursor), left arm (left cursor) and right arm(right cursor) in front of computer screen. Duringthe process of imagining the movement, the SlowCortical Potential (SCP) was recorded. The activityof the brain was also recorded from two differentchannels by applying a sampling frequency of 256Hz. Two EEG electrodes were placed accordinglybased on the international 10-20 system as shownin figure 1 and referred to the point of Cz electrodeas follow: Line 1: C3 (Central Lobe3), Line2: C4(Central Lobe4). Overall process would beundertaken within 9 seconds: however, the processof recording the sample of the material needed forthe experiment just took only 4 seconds.E-ISSN: 1817-31952.3 Feature Extraction2.3.1 Feature Extraction by Using FFTTo attain a feature extraction, the research usedacoustic analysis method to reduce EEG signalinto several sets of parameter and statistictechnique to take EEG signal. By quantifying thevalues of acoustic feature parameters of variousEEG signal, each of them has been taken from thefeature extraction and has been used asintroduction sample of EEG signal.The strength of EEG signal is taken from thesignal to measure the strength level of each EEGsignal sample data. The signal was calculated byapplying Root Mean Square (RMS). Inmathematics, RMS is known as the quadraticmean. It is the statistics to measure the magnitudeof varying quantity. RMS is useful tobe usedwhen there are positive and negative variations,for instance sinusoid. RMS is used in variousfields and most often used in the field of signal.RMS in this feature calculates RMS in frequencydomain/FFT as the following formula:RMSj , j 1 M (4)Figure 1: EEG Electrode Montage as The International10-20 System2.2. Fast Fourier Transform (FFT)FFT is an optimal computational algorithm thatimplements Discreet Fourier Transform (DFT) witha rapid calculation technique and utilizes theperiodical nature of Fourier transformation. FFT isa mathematical operation that aims to decompose atime domain signal to frequency domain signal.DFT was employed by applying a transformation inwhich the length of N vector was acounted usingthe following formula::F(u) 1/NThe EEG signals that have been selected based onsubjects imagination movements of their left finger,right finger, left arm and right arm. Then, theperformance of each EEG signal is processed byusing FFT. Each signal is tested to find the featureof EEG signal. The results of FFT process is shownby Figure 2.f(x) exp[-2πux/N] (1)(1)F(u) 1/Nf(x) (cos (2πx/N) – sin (2πux/N)) (2)(2The calculation of FFT employs multiple reflectiontransformation in order that DFT result is derivedfrom counting the half value of signal data, thus thecalculation process became faster. Then, the otherhalf values were counted through conjugate valuemethod calculated by DFT. To divide the datasignal, this study uses the following formula:b (N 1) div 2 (3)783Figure 2: Left Finger of FFT
Journal of Theoretical and Applied Information Technologyst31 August 2014. Vol. 66 No.3 2005 - 2014 JATIT & LLS. All rights reserved.ISSN: 1992-8645www.jatit.orgE-ISSN: 1817-3195Based on the result of FFT, then the strength levelof each EEG signal data sample has beenmeasured by RMS as shown by Figure 3.Result of EEGsignal after theprocess of FFTFeaturesearchingprocessResulted RMSfeatureY1Z1W1Y2Z2W2Y3Z3W3X1NFigure 3:The Chronology of EEG Signal Input after FFTProcessX2------Y8Z17W152.3.2 Feature Extraction by Using AVGAVG is a process to measure the average power ofa deterministic periodical signal. The type ofsignal is time domain signal, however, it hasresulted discrete power spectrum. A signalconsists of sinusoid, for instance electrical signalthat has unlimited energy, but the average poweris limited. To measure the average power ofspectrum, it uses periodogram spectrum object andwindow hamming method. The formula forwindow hamming method as follow:Note:X1, x2 Input (Result of RMS and resultof AVG)Y1, Y2, Y3 Y8 Neuron-neuron hiddenlayer 1Z1, Z2, Z3 Z17 Neuron-neuron hiddenlayer 2W1, W2, W3 W15 Neuron-neuronhidden layer 3N OutputW(n) 0.54 – 0.46 cos (2πn/N-1), 0 n N-1 (5)After the process of windowing, then the valueswill be converted into logarithmic value by usingthe following formula:Avg power 10 x log (W(n)/2) (6)Figure 5: Architecture of BackPropagation Network 3Hidden LayerResultedEEGSignalFeaturesearchingprocessThis research uses BackPropagation (2-8-17-15-1)method, i.e. 2 inputs are derived from thecharacteristic of EEG signal and 3 hidden layer inwhich each of them consist of 8 units, 17 units, and15 units and 1 target (the movements of left finger,right finger, left arm and right arm).Resulted AVGfeatureFigure 4: The Chronology of EEG Signal Input toResult AVG FeatureTable 1: Pattern of Input and Target Designed fromBackPropagation Method2.3.3 Back Propagation Neural NetworkBackpropagation is one of the developments of aSingle Layer Neural Network architecture. Thisarchitecture consists of input layer, hidden layerand output layer, and each layer is composed byone or more artificial neurons.784Input PatternInput data ( RMS x1 andAVG x2 )OutputTargetSignal of rightfingerimaginationCharacteristic of EEG signalfor imagination of right fingermovement0.2Signal of leftfingerimaginationCharacteristic of EEG signalfor imagination of left fingermovement0.4Signal of leftarmimaginationCharacteristic of EEG signalfor imagination of left armmovement0.6Signal of rightarmimaginationCharacteristic of EEG signalfor imagination of right armmovement0.8
Journal of Theoretical and Applied Information Technologyst31 August 2014. Vol. 66 No.3 2005 - 2014 JATIT & LLS. All rights reserved.ISSN: 1992-8645www.jatit.orgThe design of the system in this research is madethrough three processes, namely taking process ofEEG signal, feature searching process andidentification as shown in figure 6.StartTaking EEG signal dataFFTWindow HammingRESULT AND DISCUSSIONIn this research, Root Mean Square (RMS) andAverage Power Spectrum (AVG) were employedto extract features from EEG signal data and thenfor identification, a BackPropagation method wasemployed. The total data taken in this research werecomprised from 200 data file of EEG signal derivedfrom 10 subjects by using 1 channel (C3). One fileof EEG signal has 1409 data point. This researchdivides one EEG signal into 2 features. The firstfeature uses FFT to take the value of its MRS, andthen the second feature uses the value of AVG.The result values of MRS and AVG are shown inthe table 2 by taking 5 examples for the values ofMRS and AVG of each signal for each movementimagination.AVGRMS3.E-ISSN: 1817-3195Feature extractionInput Data of RMS and AVG values are used in theprocess of identification by using BackPropagationNeural Network method. There are two steps in theidentification process, namely learning process andmapping process. The learning process uses thelearning rate parameter of 0.1, yet the errors arefound to be 0.001. The initial weight values aredetermined randomly by the range of -1 to 1.IdentificationFinishFigure 6: Planning System for The Identification Processof EEG Signal DataTable2: Results of RMS and AVG from EEG Signal for Each Signal Channel C3Signal Data12345VALUEVALUEVALUEVALUEVALUEMRSA VGMRSAVGMRSA VGMRSA VGMRSA VGLeft 5Right 1Left ght Arm3742-5.703715-9.033685-8.845476-5.195477-8.63To find the optimal parameter that results the bestperformance from Neural Network, it has been doneaccording to the magnitude of Mean Squared Error(MSE) and the number of optimal hidden unitduring training process. The result of performancehas been shown in the table 3 and figure 7 – 9.As seen in figure7 by using one hidden layer, MSEvalue of 0.0366 was obtained 1000 times Iteration.The desired error for the identification process of0.001 level, then the identification is not expectedto meet the target of 100%.\Figure7: Prosess of Trainning of 1 Hidden Layer785
Journal of Theoretical and Applied Information Technologyst31 August 2014. Vol. 66 No.3 2005 - 2014 JATIT & LLS. All rights reserved.ISSN: 1992-8645www.jatit.orgE-ISSN: 1817-3195Table 3:The Performance of Neural Network onDifferent Number of Hidden LayerFigure 8: Process of Trainning of 2 Hidden LayerIn figure 8, the number of hidden layer is 2 and thevalue of MSE in process of identification has yet tomeet the target. Figure 8 with 2 hidden layers usingvalues has obtained MSE of 0.0211 to 1000 timesIteration. The desired error for the identificationprocess by 0.001, then the identification is notexpected to meet the target of 100%.Figure 9: Process Trainning of 3 Hidden LayerIn the figure 9, the number of hidden layer is 3 andthe value of MSE in process of identification hasyet to meet the target. Figure 9 with 3 hidden layerusing MSE values obtained for 0.003 to 1000 timesIteration. The desired error for the identificationprocess by 0.001, then the identification is notexpected to meet the target of 100%. By using 3hidden layer of its MSE by using less than 1 or 2hidden layers.MSE of 1HiddenLayerMSE of 2HiddenLayerMSE of 3HiddenLayerTime29 second35 second46 uracy56 %87.75 %91 %Table 3 shows that by using 3 hidden layer MSE isgetting smaller, so that the process of identifyingthe level of accuracy by using 3 hidden layer isbetter, as that is equal to 91%.4.CONCLUSIONIn this research, the researchers introduced FFT bytaking the values of RMS and AVG from EEGsignal to extract the feature and to identify ofBackPropagation neural Network. This researchuses 100 data file of EEG signal for training, thenin step of identification, it is classified into fourclasses. The data file of EEG has been added to 100data from testing data EEG signal. Thus, the totaldata in the research is 200 EEG data signal. Theaccuracy to identify BackPropagation is 91% toexamine the data using 3 hidden layer.This research has shown that the number of hiddenlayer at BackPropagation affects the magnitude ofMSE. In the future, the researcher should examinethe appropriate search technique to extract thefeature and to identify EEG signal, so that theaccuracy level would be better. The result obtainedwill be compared to the method that has beenstudied.5.ACKNOWLEDGEMENTSThis research was supported by the ntalDepartmentofInformatics,University of Kyushu. The authors are thankful tothe participants for their dedication to this researchproject.786
Journal of Theoretical and Applied Information Technologyst31 August 2014. Vol. 66 No.3 2005 - 2014 JATIT & LLS. All rights reserved.ISSN: 1992-8645www.jatit.orgtetraplegia.,” Nature, vol. 442, no. 7099, Jul.2006, pp. 164–71.REFERENCES[1] H. I. Hemorrhage, “Book reviews.,” Americanjournal of veterinary research, vol. 75, no.1,Jan. 2014, p. 4.[2] G. Pfurtscheller and F. H. Lopes da Silva,“Event-related EEG/MEG synchronization anddesynchronization: basic principles.,” Clinicalneurophysiology : official journal of y, vol. 110, no. 11, Nov. 1999,pp. 1842–57.[3] B. Z. Allison, C. Brunner, C. Altstätter, I. C.Wagner, S. Grissmann, and C. Neuper, “Ahybrid ERD/SSVEP BCI for continuoussimultaneous two dimensional cursor control.,”Journal of neuroscience methods, vol. 209, no.2, Aug. 2012, pp. 299–307.[4] L. J. Trejo, R. Rosipal, and B. Matthews,“Brain-computer interfaces for 1-D and 2-Dcursor control: designs using volitional controlof the EEG spectrum or steady-state visualevoked potentials.,” IEEE transactions onneural systems and rehabilitation engineering :a publication of the IEEE Engineering inMedicine and Biology Society, vol. 14, no. 2,Jun. 2006, pp. 225–.[5] J. S. Brumberg, A. Nieto-Castanon, P. R.Kennedy, and F. H. Guenther, ” Speech communication, vol.52, no. 4, Apr. 2010, pp. 367–379.E-ISSN: 1817-3195[9] P. B. C. Interface, E. Donchin, K. M. Spencer,and R. Wijesinghe, “The Mental Prosthesis:Assessing the Speed of a,” vol. 8, no. 2, 2000,pp. 174–179.[10] L. Defebvre, J. L. Bourriez, P. Derambure, A.Duhamel, J. D. Guieu, and A. Destee,“Influence of chronic administration of l DOPA on event-related desynchronization ofmu rhythm preceding voluntarymovement inParkinson ’ s disease,” vol. 109, 1998, pp.161–167.[11] B. D. Mensh, J. Werfel, and H. S. Seung, “BCICompetition 2003--Data set Ia: combininggamma-band power with slow corticalpotentials to improve single-trial classificationof electroencephalographic signals.,” IEEEtransactions on bio-medical engineering, vol.51, no. 6, Jun. 2004, pp. 1052–6.[12] B. Kotchoubey, D. Schneider, H. Schleichert,U. Strehl, C. Uhlmann, V. Blankenhorn, W.Fröscher, and N. Birbaumer, “Self-regulationof slow cortical potentials in epilepsy: a retrialwith analysis of influencing factors.,” Epilepsyresearch, vol. 25, no. 3, Nov. 1996, pp. 269–76.[6] J. Jin, B. Z. Allison, E. W. Sellers, C. Brunner,P. Horki, X. Wang, and C. Neuper, “Anadaptive P300-based control system.,” Journalof neural engineering, vol. 8, no. 3, Jun. 2011,p. 036006.[7] E. M. Mugler, C. a Ruf, S. Halder, M. Bensch,and A. Kubler, “Design and implementation ofa P300-based brain-computer interface forcontrolling an internet browser.,” IEEEtransactionsonneuralsystemsandrehabilitation engineering : a publication of theIEEE Engineering in Medicine and BiologySociety, vol. 18, no. 6, Dec. 2010, pp. 599–609.[8] L. R. Hochberg, M. D. Serruya, G. M. Friehs,J. a Mukand, M. Saleh, A. H. Caplan, A.Branner, D. Chen, R. D. Penn, and J. P.Donoghue, “Neuronal ensemble control ofprosthetic devices by a human with787
issn: 1992-8645 www.jatit.org e-issn: 1817-3195 782 eeg signal identification based on root mean square and average power spectrum by using backpropagation hindarto 1, moh. hariadi 2, mauridhi hery purnomo 3
4 R. EEG T. Study Guide II. Common Neurologic Disorders 25% Study Resources: A. CNS infections Online course: EEG 210: EEG in Neurological Disorders Publication: EEG Clinical Correlations: Infectious, Vascular, & Structural Disorders, 3rd Ed. Publication: Handbook of ICU EEG Monitoring.LaRoche B. Head injury Online course: EEG
May 18, 2020 · Normal EEG Part 1 Dr. Hope 9:00 – 10:30 Normal EEG Part 2 Conference Dr. Hope 9:00 – 10:30 Sleep and EEG Benign Variants Dr. Hope 9:00 – 12:00 Epilepsy Case Management 9:00 – 10:30 Activation methods Dr. Hope 11:00 – 12 EEG Discussion Unknowns and Routine EEG Dr. Hope 11:00 – 12 EEG Discussion Unk
EEG signal processing is spectral analysis. Spectral analysis, which converts the original time series to the frequency domain, is a natural choice for EEG signal processing because EEG signals are often described by a, b, u, and d waves, whose Figure 1. EEG signals from a healthy person a
Normal EEG patterns found in the waking and sleep states are identified. Assignments focus on descriptive EEG terms, waveform descriptions, and features that promote the visual analysis of EEG. Information related to medication effects on the EEG is also provided. Normal EEG variants are a ke
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 360 identification method to forecast the dynamic model . Neural Network is one of artificial methods which can be used for the nonlinear dynamic system identification. It is consists of a number of neurons arranged in numerous layers.
Sanei S, Chambers J. Introduction to EEG: EEG Signal Processing. John Wiley and Sons Ltd., 2007. EMG: . Introduction Slide I-18 64-Channel EEG Hand Muscles EMG EMG Electromyography (EMG) is a technique for evaluating and recording the activation signal of muscles. The electrical potential generated by
to design and develop a BCI controlled robotic arm using EEG signal. In this project an EPOC (a headset consists of 14 electrodes) is used to detect and collect EEG signal from human scalp. Every facial expression made by human beings generates a specific EEG signal, the BCI robotic arm is designed to be controlled by human expressions.
[Class XII : Accountancy] [110] CHAPTER 7 ACCOUNTING FOR SHARE CAPIT AL (Share and Share Capital : Nature and types) “A Company is an artificial person created by law, having separate entity with a
