classification of two class motor imagery tasks using hybrid ga-pso based k-means clustering

Research ArticleClassification of Two Class Motor Imagery Tasks Using HybridGA-PSO Based 119870-Means Clustering

Suraj1 Purnendu Tiwari2 Subhojit Ghosh3 and Rakesh Kumar Sinha4

1Electrical and Electronics Engineering Birla Institute of Technology Mesra Ranchi 835215 India2M Tech Computer Technology National Institute of Technology Raipur 492001 India3Electrical and Electronics Engineering National Institute of Technology Raipur 492001 India4Birla Institute of Technology Mesra Ranchi 835215 India

Correspondence should be addressed to Suraj surajboomgmailcom

Received 17 October 2014 Revised 20 March 2015 Accepted 21 March 2015

Academic Editor Steven L Bressler

Copyright copy 2015 Suraj et al This is an open access article distributed under the Creative Commons Attribution License whichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Transferring the brain computer interface (BCI) from laboratory condition to meet the real world application needs BCI to beapplied asynchronously without any time constraint High level of dynamism in the electroencephalogram (EEG) signal reasonsus to look toward evolutionary algorithm (EA) Motivated by these two facts in this work a hybrid GA-PSO based 119870-meansclustering technique has been used to distinguish two class motor imagery (MI) tasks The proposed hybrid GA-PSO based 119870-means clustering is found to outperform genetic algorithm (GA) and particle swarm optimization (PSO) based119870-means clusteringtechniques in terms of both accuracy and execution timeThe lesser execution time of hybrid GA-PSO technique makes it suitablefor real time BCI application Time frequency representation (TFR) techniques have been used to extract the feature of the signalunder investigation TFRs based features are extracted and relying on the concept of event related synchronization (ERD) anddesynchronization (ERD) feature vector is formed

1 Introduction

A brain-computer interface (BCI) is a system that translateshuman thoughts into an action via a control signal Electroen-cephalogram (EEG) is the most widely used biological signalused for operating a BCI system since it has relatively shorttime constants and offers relatively high temporal resolutionamong noninvasive methods [1ndash3] Numerous works overthe past few decades have revealed that EEG signal recordedfrom the scalp are basis for BCIs [4] Recently developmentof brain-controlled devices have received a great attentionbecause of their ability to bring mobility back to people withdevastating neuromuscular disorder and improve the qualityof their life [5] Rapid increase in the volume and pace ofBCI research is a result of concern and effort of the BCIgroups to provide better opportunity to people who are inldquolocked in staterdquo because of neuromuscular disorders Thishas further increased due to the advancement in technologyand allied fields of BCI [6 7] A BCI system is considered

the most advanced neurofeedback system available [8ndash10]Transferring the BCI from laboratory condition to real worldapplication needs BCI to be applied asynchronously withoutany time constraint [11] To achieve the same ongoing brainactivities are analyzed continuously irrespective of state thatis control or noncontrol An asynchronous BCI system isindependent of the cue based manner [12]

Asynchronous scheme of BCI system being independentof cue depends on clustering for classification problem Clus-tering techniques have been extensively used for classifyingEEG signals [13] Initial approach to develop asynchronoussystem was presented in late 1990s [8 14] High level ofdynamism in the EEG signal reasons us to look toward thetechnique which could address complexities such as dyna-mism and uncertainty Evolutionary algorithms (EA) whichare known to solve complex engineering problems have beenwidely applied in BCI research in different aspects such asfeature selection [15ndash17] and dimension reduction [16 18]Genetic algorithm (GA) particle swarm optimization (PSO)

Hindawi Publishing CorporationComputational Intelligence and NeuroscienceVolume 2015 Article ID 945729 11 pageshttpdxdoiorg1011552015945729

2 Computational Intelligence and Neuroscience

ant colony optimization (ACO) are EA methods which haveattracted wide attention because of self-learning and popula-tion based search capabilities [16] In this work we have usedGA PSO and hybrid GA-PSO based 119870-means clustering todistinguish two class motor imagery (MI) tasks These EAbased techniques have been used to select the initial clustercentres for119870-means clustering The hemispheric asymmetryreflected for the MI based task is exploited to differentiatethe specific tasks using 119870-means clustering based on GAPSO and hybrid GA-PSO optimization techniques Time-frequency representation (TFR) based features are extractedand relying on the concept of event related synchroniza-tion (ERS) and event related desynchronization (ERD) [19ndash24] feature vector is formed We have used asynchronousapproach of BCI to classify MI based two class tasks recordedin synchronous approach of BCIThe feature vector extractedusing TFR constitutes the data population which is classifiedusing hybrid GA-PSO based 119870-means clustering along withGA and PSO based 119870-means clustering To the best of theknowledge of the authorrsquos hybrid GA-PSO based techniquehas not been used in classification of two classes MI tasks

The next section deals with the methodology used withemphasis on experimental setup and various signal process-ing techniques used in this experiment Paradigm of theexperiment and extraction of spectral information from therecorded EEG signal are explained briefly Section 3 presentsthe results of different optimization based 119870-means in termsof accuracy and speed of convergence in addition to compa-rison with the conventional119870-means clustering Finally con-clusion and discussions are presented in Section 4

2 Materials and Methods

Nine right-handed subjects (20ndash28 years all males) are invo-lved in this experiment All the subjects were novice tothe BCI systems and had no previous experience withany biomedical signal recording Before the experiment it isconfirmed that all subjects are physically andmentally fit andhad no prior medical history of any neurological disorderAll the subjects are given briefing of the work and tasks thatthey are supposed to perform Subjects are informed that exp-eriment involves noninvasive method and it has no adverseeffect on the subjects in any form Prior written consent istaken for the recording adhering to the ethical guidelines forbiomedical research involving human research mentioned in[25] Figure 1 displays the recording setup for the undertakenexperiment

A paradigm is prepared for two types of motor imag-ination tasks along with rest state The duration of singleparadigm operation (trial) was twelve seconds in which thesubjects are asked to stay in the relaxed state for the firstsix seconds At the sixth second alerting beep sound is usedAfter one second of the beep sound a cue for 125 secondsis set during which subjects are instructed to imagine amovement of the left- or right-hands in accordance with thearrow direction which is set to appear on the screen kept ata distance of about two feet from the subject Subjects areinstructed to apply just enough effortforce so that they couldexperience an imagination of moving left- or right-hands

Figure 1 Recording setup for the BCI experiment

without any real movement Duration of cue is chosen as125 seconds in accordance with the protocol mentioned in[15 26 27] Particular cue appeared randomly in order toeliminate any chance of predicting the preceding cue Forevery subject two sessions of recordings are planned Everysession is planned for eight runs Each run comprised twentytrials with different task that is imagination of left and rightarms along with state of restThus each run consists of twentyleft and twenty right imagination trials

21 Signal Recording In the experiment six channels bipolarrecording are done that is three for motor cortex area(using three bipolar channels C3 C4 and Cz) and rest threechannels are used to detect unwanted movement and toeliminate portions of recordings so as to reduce the influenceof physiological noise Electrooculogram (EOG) is recordedbipolarly in referential manner using three electrodes oneplaced medially above and two placed laterally below theleft and right eyes respectively In addition to this EMG isrecorded from the 119898 extensor digitorum communis of theright and left arms as mentioned in [27] to detect unwantedmovements if there any during recording For recordingEMG and EOG gold plated electrodes are used and EEGsignal is recorded using EEG cap with 21 Ag electrodesAll signals including three channel EEG signals EOG andtwo channels of EMG are sampled at 500Hz To filter outmains hum from the 50Hz power line a notch filter is usedSignals recorded from all six channels are grouped intothree categories as right imagery left imagery and rest stateproducing eighteen sets of one-dimensional data

22 Signal Processing These data are further grouped as pertrial and run producing twelve sets of three-dimensional dataof size 625 times 10 times 8 corresponding to left and right imagerytasks for each channel and six sets of three-dimensional dataof size 3000 times 20 times 8 corresponding to rest state for eachchannel This size is as per trial duration of 125 secondsfor imagery tasks with sampling frequency of 500Hz thusresulting in 625 data points in one trial which is also chosenas epoch length In every session there are eight runs eachconsisting of ten right and ten left imagery tasks For therest state duration of six seconds with sampling frequencyof 500Hz agrees with the size for the data corresponding to

Computational Intelligence and Neuroscience 3

rest state Considering every rest state as a task twenty sets ofrest task are obtained for each session Eighteen sets of three-dimensional data (119897 times 119898 times 119899) where 119897 119898 and 119899 representnumbers of data points per epoch trial number and runrespectively constituting the time domain dataset Signal isfurther filtered using band pass filtered in the range of 05ndash30Hz For every task that is left imagery right imagery andstate of rest for C3 C4 and Cz two sets of data are preparedone having spectral content in range of 7ndash14Hz and the otherin the range of 17ndash26Hz Buttorworth filter with order six isselected to filter the signal vector of length 625 points at atime All eighteen sets of three-dimensional data mentionedabove are filtered producing thirty-six sets of filtered data ofappropriate size Filtered signal is feed to feature extractorwhich invokes eleven features extracting techniques for twodifferent bands of signal thus resulting in a total of twenty-twofeatures Features are extracted for 120583 and szlig bands of C3 for leftimagery tasks C3 for right imagery tasks C4 for left tasksand C4 for right imagery tasks These features are computedusing different algorithms of time frequency representationsDetails of these techniques are mentioned in [15 28]

The TFR features used in classification of MI based tasksin this experiment are listed in Table 1 After computingdifferent features in order to form feature vector absolutevalue of features is computed For extracting features a datavector of length 125 second is taken with sampling frequencyof 500Hz This is also chosen as epoch length and for everyprocessing a signal vector of 625 points is used at a time Wehave used data corresponding to C3 and C4 channels for taskdiscrimination as used in [29] Figure 2 shows Born Jordanfeature computed for arbitrary trials These features havebeen computed using time frequency toolbox of MATLAB[24] Maximum value of the absolute value of feature value(TFRs) obtained for every epoch of data for various featuresis determinedThenmean value is computed for every trial ofeach type of tasks Further concept of ERS and ERD is used informing feature vector as mentioned in [15 26] which statesthat there is ERS of the 120583 rhythm on the contralateral sideand a slight ERS in the central szlig rhythm on the ipsilateralhemisphere This hemispheric asymmetry reflected in theEEGs is exploited to differentiate the task desirably To exploitthe concept of hemispheric asymmetry with regard to 120583 and szligbands we combined feature matrix obtained with respect to120583 and szlig bands to form a pattern which is likely to inherit theproperty of ERS and ERD as mentioned in [15] This resultedin a feature matrix of dimension 160 times 22 where there aretwenty-two features using two bands of signal with elevendifferent features for eighty each of the two classes for everysession In our present workwe used two sessions at a time forpurpose of classificationsThus a composite feature matrix ofdimension 320 times 22 is obtained Different techniques of TFRare extensively used in the area of BCI [15 30ndash39]

23 Clustering Clustering of a data population involvesgrouping of abstract unlabelled patterns into groups calledclusters such that patterns in same cluster are similar to eachother and dissimilar to patterns from other clusters An unla-belled set of sample patterns comprising a population is givenas input to a clustering algorithm which partitions the input

Table 1 Features used for classification with their type

Type FeatureLinear time frequencyprocessing Short time Fourier transform

Bilinear time frequencyprocessing (Cohenrsquosclass)

Born JordanChoi-Williams distributionPseudo-Wigner-Ville distributionSmoothed pseudo-Wigner-VilledistributionWigner-Ville distributionZhaos-Atlas-Marks distribution

Bilinear time frequencyprocessing (Affine class)

Unitary Bertrand distributionD-Flandrin distributionScalogram for Morlet waveletSmoothed pseudo-Affine-Wignerdistribution

patterns into groups called clusters and labels the samplesaccordingly [13]

The most widely used clustering algorithm that is 119870-means algorithm takes 119870 as input and partitions dataset of119873 objects into K clusters where 119870 lt 119873 Cluster similarityis measured with respect to the dissimilarity between a datapoint and mean value of the patterns (centroid) in cluster Ingeneral square-error criterion is used as criterion functiongiven as

119863 =

119870

sum

119895=1

sum

119901isin119862119895

10038161003816100381610038161003816119898119895minus 119901

10038161003816100381610038161003816

2

(1)

where119863 is sum of square-error for all patterns in dataset119898119895

is mean of pattern in 119895th cluster 119862119895 and 119901 represents pattern

or point in cluster 119862119895

Starting with a set of randomly selected cluster centresthe centres are iteratively updated with aim of minimizingcriterion function 119863 In other words 119870-means can also beviewed as optimization strategy which searches for appro-priate cluster centre so as to minimize criterion function 119863given by (1) based on initial cluster centres given as input tothe 119870-means Conventional 119870-means is very sensitive to theinitial selected centers Several methods have been reportedin literature to solve the cluster centre initialization (CCI)problem The limitations of conventional 119870-means are asfollows

(i) Objective function of 119870-means as given by (1) isconvex [40] hence it may contain local minima Con-sequently there is possibility of convergence to thelocal minima

(ii) Using conventional 119870-means algorithm we are notable to determine global solution to the problem thatis clustering as output of 119870-means is highly depen-dent on initial cluster centres and hence gives localsolutions to clustering problem [40ndash42]

To avoid premature convergence to the local optimal pointin the present work119870-means algorithm has been formulated


0 2 40

01

02

03

04

PSD

Nor

mal

ized

freq

uenc

y (H

z)

SamplesN

orm

aliz

ed fr

eque

ncy

(Hz)

Time frequency representation

0 200 400 6000

01

02

03

04

0 200 400 600

0001

Sign

al

minus001

(a)

Samples

Nor

mal

ized

freq

uenc

y (H

z)


0 100 200 300 400 500 6000

00501

01502

02503

03504

045

0 100 200 300 400 500 600

0001002

Sign

al

minus001

minus002

0 2 40

005

01

015

02

025

03

035

04

045

PSD

Nor

mal

ized

freq

uenc

y (H

z)

(b)

Samples

Nor

mal

ized

freq

uenc

y (H

z)


0 100 200 300 400 500 6000

00501

01502

02503

03504

045

0 100 200 300 400 500 600

00050

0010015

Sign

al

minus0005

minus001

minus0015

0 2 40

005

01

015

02

025

03

035

04

045

PSD

Nor

mal

ized

freq

uenc

y (H

z)

(c)

Figure 2 Continued


Time (s)

Nor

mal

ized

freq

uenc

y (H

z)


0 100 200 300 400 500 6000

00501

01502

02503

03504

045

0 100 200 300 400 500 600

00005

001

Sign

al

minus0005

minus001

0 21 30

005

01

015

02

025

03

035

04

045

PSD

Nor

mal

ized

freq

uenc

y (H

z)

(d)

Figure 2 (a) Born Jordan feature for szlig band in C3 channel for arbitrary trial corresponding to right task (b) Born Jordan feature for szlig bandin C4 channel for arbitrary trial corresponding to right task (c) Born Jordan feature for 120583 band in C4 channel for arbitrary trial correspondingto left task (d) Born Jordan feature for 120583 band in C4 channel for arbitrary trial corresponding to right task

as a global optimization problem in the present workApproaches for solving the optimization problem can bebroadly classified into two groups that is gradient based andpopulation based evolutionary approach Though gradientbased methods quickly converge to an optimal solution itfails for nondifferentiable or discontinuous problems It is notapplicable even if the objective function is not completelyknown due to limited knowledge which is very likely forreal time applications In this context EA based optimizationtechnique has been found to be quite effective formultimodalobjective functions High level of dynamism in EEG signalrequires techniques that could address complexities such asdynamism and uncertainty In this work we have used threeEA based approaches that is PSO GA and hybrid GA-PSOto select the initial centres Both PSO and GA have theirown advantages and limitations PSO and GA are initializedwith a group of a randomly generated population and havefitness values to evaluate the population They update thepopulation and search for the optimum solutions In contraryto GA PSO does not have genetic operators like crossoverand mutation In PSO particles update themselves with theirinternal velocity and they also have memory (all particlesremember their previous best position) The informationsharingmechanism in PSO andGA is also significantly differ-ent [43] In GAs chromosomes share information with eachother So the whole population moves towards an optimalarea as a group In PSO only global best particle gives outthe information to others and hence it follows one wayinformation sharingmechanism In PSO all the particles tendto converge to the best solution faster in comparison withGA Angeline compared betweenGAs and PSOs and has sug-gested in [44] that a hybrid of the standard GA and PSOmodels could furnish better resultMotivated by this a hybrid

GA-PSO based clustering algorithm has been proposedwhich combines standard position and update rules of PSOwith operators of genetic algorithm selection crossover andmutation as mentioned in [45] The tuning parameters ofboth the optimization and the clustering technique are selec-ted based on a series of pilot runs to determine the bestpossible values for higher classification accuracy

24 Swarm Representation In order to use PSO to solve clus-tering problem each individual particle position represents119870 cluster centres consisting of 119873 lowast 119870 real numbers in a119870 times 119873 matrix where first row represents first cluster centreand second represents second cluster centre and so on

25 Swarm Initialization In this step each particle swarmencoded with cluster centre is initialized by randomly choos-ing 119870 data points Each particle is represented by 119870 times 119873

matrix This process is repeated 119875 times where 119875 is numberof particles

26 Fitness Computation Each particle in consideration isgiven as input to 119870-means and each point 119901

119894in the data set

is assigned to its nearest cluster 119862119895with 119898

119894as centre Then a

new cluster mean is obtained to get new cluster centre whichthen replaces the particle in consideration

Using the above newly obtained cluster centre fitness ofthe particle is calculated using (1)

27 Update 119875119887119890119904119905

and 119866119887119890119904119905

In this step each particle updatesits personal best position denoted by 119875best and the associatedfitness value 119875best contains each individual particlersquos best


position seen up to the current generation It is updated asfollows

(i) For each particle

(a) compare particlersquos current fitness value with fit-ness value of 119875best

(b) if current fitness value is better than fitness valueof 119875best then set 119875best value equal to the currentposition of particle and the update the fitnessvalue associated with it equal to particle currentfitness value

(c) otherwise do nothing

Updation119866best involves positioning of particle with best posi-tion seen up to current generation In the first generation ofPSO119866best and associated fitness are equal to the position andfitness of particle with best fitness value In the subsequentgenerations 119866best is updated as follows

(i) For each particle in swarm

(a) compare the fitness value of119875best with the fitnessvalue 119866best

(b) if fitness value of 119875best is better than the fitnessof 119866best then set 119866best to 119875best and fitness valueof 119866best is equal to fitness value of 119875best


28 Position and Velocity Update Half of the worst perform-ing particles (let them be called 119875

2) are eliminated on basis of

their fitness Remaining half of best performing particles (letthem be called 119875

1) undergo position and velocity update as in

standard PSO given by (2)

V (119905 + 1) = 119908 times V (119905) + 1198881times (119901 (119905) minus 119909 (119905)) + 119888

2times (119892 (119905) minus 119909 (119905))

119909 (119905 + 1) = 119909 (119905) + V (119905 + 1)

(2)

where 119909(119905) 119909(119905 + 1) is the position of the particle at iteration119905 and 119905 + 1 respectively 119901(119905) is the personal best (119875best)position of the particle 119909(119905) found so far 119892(119905) is the globalbest (119866best) position of the any of the particle found so far V(119905)is the velocity of the particle 119909(119905) to update its position 119908(119905)

represents confidence in its ownmovement or inertial weight1198881is a constant called cognitive parameter 119888

2is a constant

called social parameter 119908(119905) varies with iterations accordingto equation given below

119908 (119905) =max iterartions minus 119905

max iterartions (3)

29 Selection Remaining half that is 1198752 particles are gen-erated fromparticles that have undergone position and veloc-ity updating using appropriate selection techniques In thiswork we have used tournament selection strategy for repro-duction or forming the mating pool In tournament selectionldquo119898rdquo individuals are randomly selected from the populationwhere ldquo119898rdquo is called tournament size The individual with

best fitness among ldquo119898rdquo individuals is selected into matingpool Advantage of tournament selection over roulette wheelselection is that it does not depend on negative fitness valuesand whether a problem is a maximization or minimizationproblem unlike roulette wheel selection Also it is not fullybiased towards fittest individual in population as in the caseof Roulette wheel selection

210 Crossover and Mutation The particles generated inabove step using tournament selection go through velocitypropelled averaged crossover (VPAC) [45] and mutationIn VPAC two child particles are produced such that theirposition is between their parents but is accelerated away fromtheir current direction (by adding negative velocity) VPACcrossover is able to create necessary diversity in particlesto make searching process more effective New children areobtained based on

1198881 (119909) =

1199011 (119909) + 119901

2 (119909)

2minus 1205791times 1199011 (V)

1198882 (119909) =

1199011 (119909) + 119901

2 (119909)

2minus 1205792times 1199012 (V)

(4)

where 1198881(119909) and 119888

2(119909) are positions of child 1 and child 2

respectively 1199011(119909) and 119901

1(119909) are positions of parent 1 and

parent 2 respectively V1(119909) and V

2(119909) are velocities of parent 1

and parent 2 respectively 1205791and 1205792are uniformly distributed

random numbers between 0 and 1 and child particle retainstheir parents velocity that is 119888

1(V) = 119901

1(V) and 119888

2(V) = 119901

2(V)

Then particles after going through velocity propelledaveraged crossover undergo mutation as described in [46]In this step each chromosome undergoes mutation with fixedsmall probability 120583

119898 For mutating a chromosome whose

clustering metric is119872 a number 120597 in the range (minus119877 to +119877) isgenerated with uniform distribution where 119877 is calculated asfollows

119877 =

119872 minus 119872max119872max minus 119872min

if 119872max gt 119872

1 if 119872max = 119872min(5)

where 119872min and 119872max are the minimum and maximumvalues of the clustering metric respectively in the currentpopulation If theminimum andmaximum values of the dataset along the 119894th dimension (119894 = 1 2 119899) are 119909119894min and 119909

119894

maxrespectively and the position to be mutated is 119894th dimensionof a cluster with value 119909119894 then after mutation value becomes

119909119894+ 120597 times (119909

119894

max minus 119909119894) if 120597 gt 0

119909119894+ 120597 times (119909

119894minus 119909119894

min) otherwise(6)

This scheme of mutation provides perturbation in a maxi-mum range to strings either when they have the largest valueof 119872 in the population (ie 119872 = 119872max) or when all thestrings have the same value of the clustering metric (ie119872 =

119872min minus 119872max) On the other hand the best string(s) in thepopulation (ie the one with 119872 = 119872min) is not perturbedat all in the current generation Moreover the perturbation issuch that themutated centres still lie within the bounds of thedata points


Table 2 Classification performance against different K-means based clustering

Subject 119870-means GA based 119870-means PSO based119870-means GA-PSO based119870-means

Subject 1Average () 5719 5832 5948 6042Standard deviation 3271 0229 1484 0137





Subject 6Average () 5719 5850 6028 5737Std deviation 5552 0868 5148 1782




211 Termination Criteria The process of fitness computa-tion updating of personal best global best velocity and posi-tion and crossover and mutation repeated for maximumnumber of iterations Position of global best particle at thelast iteration gives solution to the clustering problem

3 Results

The appropriateness of the proposed hybrid evolutionarytechnique in classifying two classMI tasks has been evaluatedin this sectionThe objective functions of GA based119870-meansclustering algorithm PSO based 119870-means clustering andGA-PSO based119870-means clustering algorithm were modifiedto give misclassification as output for the optimization prob-lem The formulation of the objective function in terms ofclassification accuracy allows using the optimization based119870-means clustering for maximizing classification accuracyThe class information is not used for clustering but forobtaining the solution of the optimization problemThe classinformation is further used to test the performance of theclustering resultThedata sets for each subjectwere separately

evaluated using the evolutionary algorithms and conven-tional 119870-means clustering Keeping population size as 1000algorithms were executed for 100 iterations For GA based119870-means classifier algorithm crossover probability 120583

119888= 065

and mutation probability 120583119898= 008 were taken For PSO

based 119870-means clustering algorithm cognitive parameter1198881= 149 and social parameter 119888

2=149 and inertial weight

according to (3) were taken For GA-PSO based 119870-meansclassifier algorithm parameters crossover probability 120583

119888=

065 mutation probability 120583119898= 008 cognitive parameter

1198881= 149 social parameter 119888

2= 149 and inertial weight

according to (3) were taken Above parameters are selectedby hit and trial method

Table 2 shows the results for performance of GA based119870-means classifier algorithm PSO based119870-means classifier andGA-PSO based 119870-means classifier algorithm based on accu-racy obtained on test set for each of nine subjects Figure 3shows the variation of misclassification () with numberof iterations for an arbitrary subject It can be seen that thehybrid technique not only achieves a higher classificationbut also achieves the same with lesser iterations The lesser


0 20 40 60 80 10030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(a)

0 2010 4030 6050 8070 1009030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(b)

0 20 40 60 80 10032

34

36

38

40

42

44

46

48

Iterations

Misc

lass

ifica

tion

()

(c)

Figure 3 (a) Variation in misclassification () with iterations for GA-PSO based 119870-means clustering for different subjects (b) Variationin misclassification () with iterations for GA based 119870-means clustering for different subjects (c) Variation in misclassification () withiterations for PSO based 119870-means clustering for different subjects

execution time of the hybrid technique makes it suitable forreal time BCI application where imagery signal needs to beclassified at a very fast rate Further we used statistical test onthe results to test the significance of the result Table 3 indi-cates the average ranking of clustering algorithms based onthe Friedmanrsquos test and Table 4 shows various statisticalvalues from Friedman and Iman-Davenport tests indicatingrejection of null hypothesis The rejection of null hypothesisindicates similarity in the superiority of the proposed hybridGA-PSO method over other techniques across all the sub-jects

4 Conclusion and Discussions

It can be observed from Table 2 that except for subject 5and subject 6 GA-PSO based 119870-means classifier algorithm

is able to outperform or is comparable in terms of averageaccuracy values obtained for 100 executions of each algo-rithm The performance of proposed clustering algorithmmay further increase if it is executed for more number ofiterations and with greater size population The hybrid GA-PSO algorithm has been used to search cluster centres inthe search space such that criterion function 119863 given by(1) is minimized The knowledge that the mean of pointsbelonging to same cluster represents the cluster centre hasbeen used for improving the search capability of optimiza-tion based clustering method Floating point representationhas been adopted to represent candidate solutions (in GAcandidate solutions are known as chromosomes and inPSO they are known as particles) since it is found tobe more appropriate and natural for encoding the clustercentres It can be observed from results that GA-PSO based


Table 3 Average ranking of clustering algorithms based on Friedmanrsquos test with a critical value 119883(005 3) = 7814733

Algorithms 119870-means GA based 119870-means PSO based119870-means Hybrid GA-PSObased 119870-means

Ranking 4 289 178 133

Table 4 Friedman and Iman-Davenport statistical tests

Method Statistical value 119875 value HypothesisFriedman 2313333 379 times 10minus5 RejectedIman-Davenport 478621 lt000001 Rejected

119870-means clustering algorithm performs better or at leastgives comparable performance with GA based 119870-meansclustering algorithm and PSO based 119870-means clusteringalgorithm Figure 3 displays the convergence characteristicsof the three optimization based clustering algorithms forall the three cases It is clear that the hybrid technique(Figure 3(a)) is able to converge to the final solution quiteearly (lesser iterations and computational time) as comparedto the isolated GA (Figure 3(b)) and PSO (Figure 3(c)) basedtechniques Thus the hybrid technique not only achieves ahigher classification but also achieves the same with lesseriterations For validation of the GA-PSO based 119870-meansclustering statistical analysis is also performed along withthe experimental analysis A statistical analysis is performedto demonstrate the significant differences among the resultsobtained using the different algorithms To prove the sameFriedman and Iman-Davenport statistical tests are carriedout The nonparametric statistical test allows checking thesignificance and repeatability of the results In other wordsTable 2 indicates the average ranking of clustering algorithmsbased on Friedmanrsquos test using sum of intracluster distance(average) parameter and the critical value obtained forFriedman test 119883 (005 3) is 7814733 The 119875 values computedby the Friedman test and the Iman-Davenport test are givenin Table 4 both of these tests reject the null hypothesis andsupport the presence of significant differences among theperformance of all clustering algorithms used in this work

Basically MI based BCIs are composed of two stages fea-ture extraction and feature classification [47 48] The abilityto analyze EEG is limited by methods that require stationaryepochs of data Stationary time-series techniques such asFourier transform for feature extraction cannot be used forEEG signal hence joint time-frequency analysis is required[30] TFR is based on the principle of extracting energy distri-bution on time versus frequency In this way different freque-ncy components are localized at a good temporal resolution[31] TFRs can provide amplitude and phase spectra in bothfrequency and time domain [49] TFRs methods have beenproposed for EEG [30ndash32] electromyogram (EMG) [50] andEGG [51] signal analysis TFRs offer the ability to analyzerelatively long continuous segment of data even when dyna-mics are rapidly changing One approach to the accurateanalysis of nonstationary signal is to represent the signalas a sum of waveforms with well-defined time frequencyproperties Time frequency techniques are further classifiedinto two classes namely linear and bilinear time frequency

representations In our analysis we have used techniques fromboth classes Details of these techniques are mentioned in[28] For classifyingMI based task a hybridGA-PSObased119870-means clustering algorithm has been used in this work alongwith GA and PSO based119870-means clustering to compare andhence to check its suitability in classification problems of MIbased taskOur result shows that combination ofGAandPSOforming hybridGA-PSO based119870means clustering algorithmalmost outperforms both the GA and PSO based 119870-meansclustering algorithm Enhanced performance achieved incase of hybrid GA-PSO based 119870-means clustering can beexplained in terms of the benefit that PSO and GA facilitatesThe algorithm is designed such that GA facilitates a globalsearch and PSO facilitates a local searchThe hybrid GA-PSOapproach merges the standard velocity and position updaterules of PSO with that of GArsquos operations (ie selectioncrossover and mutation) [45] Data is likely to get affectedby level of degree of attention of the subject performing thetask and the changes in their concentration can affect theresult adversely Therefore few sets of pretraining prior toactual recording and ensuring that subject take sufficient restbefore the actual recording becomes essential In this analysiswe have used fixed frequency band (ie 120583 and szlig bands)for every subject for ERDERS discrimination to reduce thecomplexity of the system Selecting varying 120583 and szlig bands asused in [15] is likely to give better result Overall performancewill also depend on reliability and performance of othertechniques used such as data filtering feature extraction andnormalization

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] J R Wolpaw N Birbaumer D J McFarland G Pfurtschellerand T M Vaughan ldquoBrain-computer interfaces for communi-cation and controlrdquo Clinical Neurophysiology vol 113 no 6 pp767ndash791 2002

[2] A Nijholt and D Tan ldquoBrain-computer interfacing for intelli-gent systemsrdquo IEEE Intelligent Systems vol 23 no 3 pp 72ndash792008

[3] R K Sinha and G Prasad ldquoPsychophysiology and autonomicresponses perspective for advancing hybrid brain-computer


interface systemsrdquo Journal of Clinical Engineering vol 38 no4 pp 196ndash209 2013

[4] D J Krusienski D J McFarland and J R Wolpaw ldquoValue ofamplitude phase and coherence features for a sensorimotorrhythm-based brain-computer interfacerdquo Brain Research Bul-letin vol 87 no 1 pp 130ndash134 2012

[5] L Bi X-A Fan and Y Liu ldquoEEG-based brain-controlledmobile robots a surveyrdquo IEEE Transactions onHuman-MachineSystems vol 43 no 2 pp 161ndash176 2013

[6] F Popescu S Fazli Y Badower B Blankertz and K R MullerldquoSingle trial classification ofmotor imagination using 6 dry EEGelectrodesrdquo PLoS ONE vol 2 no 7 article e637 2007

[7] J R Wolpaw N Birbaumer D J McFarland G Pfurtschellerand TM Vaughan ldquoBrain-computer interface for communica-tion and controlrdquoClinical Neurophysiology vol 133 pp 767ndash7912002

[8] R Kus D Valbuena J Zygierewicz T Malechka A Graeserand P Durka ldquoAsynchronous BCI based on motor imagerywith automated calibration and neurofeedback trainingrdquo IEEETransactions on Neural Systems and Rehabilitation Engineeringvol 20 no 6 pp 823ndash835 2012

[9] N Birbaumer and P Sauseng ldquoBrain-computer interface inneurorehabilitationrdquo in Brain-Computer Inferfaces pp 65ndash782010

[10] G Townsend B Graimann and G Pfurtscheller ldquoContinuousEEG classification during motor imagery-simulation of anasynchronous BCIrdquo IEEE Transactions on Neural Systems andRehabilitation Engineering vol 12 no 2 pp 258ndash265 2004

[11] R Leeb D Friedman G R Muller-Putz R Scherer M Slaterand G Pfurtscheller ldquoSelf-paced (asynchronous) BCI controlof a wheelchair in virtual environments a case study with atetraplegicrdquo Computational Intelligence and Neuroscience vol2007 Article ID 79642 8 pages 2007

[12] A A Nooh J Yunus and S M Daud ldquoA review of asyn-chronous electroencephalogram-based brain computer inter-face systemsrdquo in Proceedings of the International Conference onBiomedical Engineering and Technology (IPCBEE rsquo11) vol 11 pp55ndash59 IACSIT Press 2011

[13] R O Duda P E Hart and D G Stork Pattern ClassificationWiley 2nd edition 2000

[14] A Schloegl C Neuper and G Pfurtscheller ldquoSubject specificEEG patterns during motor imaginaryrdquo in Proceedings of the19th Annual International Conference of the IEEE Engineeringin Medicine and Biology Society (EMBS rsquo97) pp 1530ndash1532November 1997

[15] P Herman G Prasad T M McGinnity and D Coyle ldquoCom-parative analysis of spectral approaches to feature extraction forEEG-based motor imagery classificationrdquo IEEE Transactions onNeural Systems and Rehabilitation Engineering vol 16 no 4 pp317ndash326 2008

[16] A Atyabi M Luerssen S P Fitzgibbon and D M W PowersldquoThe use of Evolutionary algorithm-based methods in EEGbased BCI systemsrdquo in Swarm Intelligence for Electric andElectronic Engineering IGI global Hershey Pa USA 2013

[17] A Atyabi M Luerssen S P Fitzgibbon and D M W PowersldquoEvolutionary feature selection and electrode reduction forEEG classificationrdquo in Proceedings of the IEEE Congress onEvolutionary Computation (CEC rsquo12) pp 1ndash8 June 2012

[18] A Atyabi M Luerssen S P Fitzgibbon and D M W PowersldquoDimension reduction in EEG data using particle swarm opti-mizationrdquo in Proceedings of the IEEE Congress on Evolutionary

Computation (CEC rsquo12) pp 1ndash8 IEEE Brisbane Australia June2012

[19] G Pfurtscheller ldquoFunctional brain imaging based on ERDERSrdquo Vision Research vol 41 no 10-11 pp 1257ndash1260 2001

[20] G Pfurtscheller and F H Lopes da Silva ldquoEvent-relatedEEGMEG synchronization and desynchronization basic prin-ciplesrdquo Clinical Neurophysiology vol 110 no 11 pp 1842ndash18571999

[21] Y Jeon C S Nam Y-J Kim and M C Whang ldquoEvent-related(De)synchronization (ERDERS) during motor imagery tasksimplications for brain-computer interfacesrdquo International Jour-nal of Industrial Ergonomics vol 41 no 5 pp 428ndash436 2011

[22] C S Nam Y Jeon Y J Kim I Lee and K Park ldquoMovementimagery-related lateralization of event-related (de)synchro-nization (ERDERS) motor-imagery duration effectsrdquo ClinicalNeurophysiology vol 122 no 3 pp 567ndash577 2011

[23] B Blankertz C Sannelli S Halder et al ldquoNeurophysiologicalpredictor of SMR-based BCI performancerdquoNeuroImage vol 51no 4 pp 1303ndash1309 2010

[24] VMorash O Bai S Furlani P Lin andM Hallett ldquoClassifyingEEG signals preceding right hand left hand tongue and rightfoot movements and motor imageriesrdquo Clinical Neurophysiol-ogy vol 119 no 11 pp 2570ndash2578 2008

[25] Council for International Organization of Medical SciencesInternational Ethical Guidelines for Biomedical Research Involv-ing Human Subjects Council for International Organizationof Medical Sciences Geneva Switzerland 2002 httpwwwwhointethicsresearchen

[26] G Pfurtscheller C Neuper A Schlogl and K Lugger ldquoSep-arability of EEG signals recorded during right and left motorimagery using adaptive autoregressive parametersrdquo IEEE Trans-actions on Rehabilitation Engineering vol 6 no 3 pp 316ndash3251998

[27] H Ramoser J Muller-Gerking and G Pfurtscheller ldquoOptimalspatial filtering of single trial EEG during imagined handmovementrdquo IEEE Transactions on Rehabilitation Engineeringvol 8 no 4 pp 441ndash446 2000

[28] F Auger P Flandrin P Gonealves and O Lemoine Time-Frequency Toolbox Tutorial CNRS Clermont-Ferrand France1997 httptftbnongnuorgrefguidepdf

[29] S Lemm C Schafer and G Curio ldquoBCI Competition 2003mdashdata set III modeling of sensorimotor 120583 rhythms for classifi-cation of imaginary hand movementsrdquo IEEE Transactions onBiomedical Engineering vol 51 no 6 pp 1077ndash1080 2004

[30] P Celka B Boashash and P Colditz ldquoPreprocessing and time-frequency analysis of newborn EEG seizuresrdquo IEEE EngineeringinMedicine andBiologyMagazine vol 20 no 5 pp 30ndash39 2001

[31] S Tong Z Li Y Zhu and N V Thakor ldquoDescribing thenonstationarity level of neurological signals based on quantifi-cations of time-frequency representationrdquo IEEETransactions onBiomedical Engineering vol 54 no 10 pp 1780ndash1785 2007

[32] P J Franaszczuk G K Bergey P J Durka andHM EisenbergldquoTime-frequency analysis using thematching pursuit algorithmapplied to seizures originating from the mesial temporal loberdquoElectroencephalography and Clinical Neurophysiology vol 106no 6 pp 513ndash521 1998

[33] Y Xu S Haykin and R J Racine ldquoMultiple window time-frequency distribution and coherence of EEG using Slepiansequences and Hermite functionsrdquo IEEE Transactions onBiomedical Engineering vol 46 no 7 pp 861ndash866 1999


[34] N Robinson A P Vinod K K Ang K P Tee and C T GuanldquoEEG-based classification of fast and slow hand movementsusing wavelet-CSP algorithmrdquo IEEE Transactions on BiomedicalEngineering vol 60 no 8 pp 2123ndash2132 2013

[35] M Roessgen M Deriche and B Boashash ldquoA comparativestudy of spectral estimation techniques for noisy non-stationarysignals with application to EEG datardquo in Proceedings of the 27thAsilomar Conference on Signals Systems and Computers vol 2pp 1157ndash1161 IEEE Pacific Grove Calif USA November 1993

[36] M B Shamsollahi L Senhadji and R Le Bouquin-JeannesldquoTime-frequency analysis a comparison of approaches forcomplex EEG patterns in epileptic seizuresrdquo in Proceedings ofthe IEEE Engineering in Medicine and Biology 17th Annual Con-ference and 21st Canadian Medical and Biological EngineeringConference pp 1069ndash1070 September 1995

[37] Z Shao-bai and H Dan-dan ldquoElectroencephalography fea-ture extraction using high time-frequency resolution analysisrdquoTELKOMNIKA Indonesian Journal of Electrical Engineeringvol 10 no 6 pp 1415ndash1421 2012

[38] N Stevenson M Mesbah and B Boashash ldquoA feature set forEEG seizure detection in the newborn based on seizure andbackground charactersiticsrdquo in Proceedings of the 29th AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (EMBS rsquo07) pp 7ndash10 Lyon France August2007

[39] M Sun S Qian X Yan and et al ldquoLocalizing functional activityin the brain through time-frequency analysis and synthesis ofthe EEGrdquo Proceedings of the IEEE vol 85 no 9 pp 1302ndash13111996

[40] D W van der Merwe and A P Engelbrecht ldquoData cluster-ing using particle swarm optimizationrdquo in Proceedings of theCongress on Evolutionary Computation (CEC rsquo03) vol 1 pp 215ndash220 December 2003

[41] S Kalyani andK S Swarup ldquoParticle swarmoptimization basedK-means clustering approach for security assessment in powersystemsrdquo Expert Systems with Applications vol 38 no 9 pp10839ndash10846 2011

[42] Z S Shokri and M A Ismail ldquoK-means-type algorithms ageneralized convergence theorem and characterization of localoptimalityrdquo IEEE Transactions on Pattern Analysis andMachineIntelligence vol PAMI-6 no 1 pp 81ndash87 1984

[43] S N Sivanandam and S N Deepa Introduction to GeneticAlgorithms Springer 2007

[44] P J Angeline ldquoEvolutionary optimization versus particle swarmoptimization philosophy and performance differencesrdquo inEvolutionary Programming VII vol 1447 of Lecture Notes inComputer Science pp 601ndash610 Springer Berlin Germany 1998

[45] M Settles and T Soule ldquoBreeding swarms a GAPSO hybridrdquoin Proceedings of the 7th Annual Conference on Genetic andEvolutionary Computation (GECCO rsquo05) pp 161ndash168 ACMNew York NY USA June 2005

[46] S Bandyopadhay and U Maulik ldquoAn evolutionary techniquebased on K-means algorithm for optimal clustering in R119873rdquoInformation Sciences vol 146 no 1ndash4 pp 221ndash237 2002

[47] S Siuly and Y Li ldquoImproving the separability of motor imageryEEG signals using a cross correlation-based least square supportvector machine for brain-computer interfacerdquo IEEE Transac-tions on Neural Systems and Rehabilitation Engineering vol 20no 4 pp 526ndash538 2012

[48] W Wu X Gao B Hong and S Gao ldquoClassifying single trialEEG during motor imagery by iterative spatio-spectral pattern

learning (ISSPL)rdquo IEEETransactions onBiomedical Engineeringvol 55 no 6 pp 1733ndash1743 2008

[49] C R Pinnegar H Khosravani and P Federico ldquoTime fre-quency phase analysis of ictal eeg recordings with the s-transformrdquo IEEE Transactions on Biomedical Engineering vol56 no 11 pp 2583ndash2593 2009

[50] J Duchene D Devedeux S Mansour and C Marque ldquoAnalyz-ing uterine EMG tracking instantaneous burst frequencyrdquo IEEEEngineering inMedicine and BiologyMagazine vol 14 no 2 pp125ndash132 1995

[51] J Chen J Vandewalle W Sansen G Vantrappen and JJanssens ldquoAdaptive spectral analysis of cutaneous electrogastricsignals using autoregressive moving average modellingrdquo Med-ical amp Biological Engineering amp Computing vol 28 no 6 pp531ndash536 1990

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Distributed Sensor Networks


Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014


ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014


Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications


Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia


Biomedical Imaging


ArtificialNeural Systems

Advances in


RoboticsJournal of



Computational Intelligence and Neuroscience

Industrial EngineeringJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in



ant colony optimization (ACO) are EA methods which haveattracted wide attention because of self-learning and popula-tion based search capabilities [16] In this work we have usedGA PSO and hybrid GA-PSO based 119870-means clustering todistinguish two class motor imagery (MI) tasks These EAbased techniques have been used to select the initial clustercentres for119870-means clustering The hemispheric asymmetryreflected for the MI based task is exploited to differentiatethe specific tasks using 119870-means clustering based on GAPSO and hybrid GA-PSO optimization techniques Time-frequency representation (TFR) based features are extractedand relying on the concept of event related synchroniza-tion (ERS) and event related desynchronization (ERD) [19ndash24] feature vector is formed We have used asynchronousapproach of BCI to classify MI based two class tasks recordedin synchronous approach of BCIThe feature vector extractedusing TFR constitutes the data population which is classifiedusing hybrid GA-PSO based 119870-means clustering along withGA and PSO based 119870-means clustering To the best of theknowledge of the authorrsquos hybrid GA-PSO based techniquehas not been used in classification of two classes MI tasks

The next section deals with the methodology used withemphasis on experimental setup and various signal process-ing techniques used in this experiment Paradigm of theexperiment and extraction of spectral information from therecorded EEG signal are explained briefly Section 3 presentsthe results of different optimization based 119870-means in termsof accuracy and speed of convergence in addition to compa-rison with the conventional119870-means clustering Finally con-clusion and discussions are presented in Section 4

2 Materials and Methods

Nine right-handed subjects (20ndash28 years all males) are invo-lved in this experiment All the subjects were novice tothe BCI systems and had no previous experience withany biomedical signal recording Before the experiment it isconfirmed that all subjects are physically andmentally fit andhad no prior medical history of any neurological disorderAll the subjects are given briefing of the work and tasks thatthey are supposed to perform Subjects are informed that exp-eriment involves noninvasive method and it has no adverseeffect on the subjects in any form Prior written consent istaken for the recording adhering to the ethical guidelines forbiomedical research involving human research mentioned in[25] Figure 1 displays the recording setup for the undertakenexperiment

A paradigm is prepared for two types of motor imag-ination tasks along with rest state The duration of singleparadigm operation (trial) was twelve seconds in which thesubjects are asked to stay in the relaxed state for the firstsix seconds At the sixth second alerting beep sound is usedAfter one second of the beep sound a cue for 125 secondsis set during which subjects are instructed to imagine amovement of the left- or right-hands in accordance with thearrow direction which is set to appear on the screen kept ata distance of about two feet from the subject Subjects areinstructed to apply just enough effortforce so that they couldexperience an imagination of moving left- or right-hands

Figure 1 Recording setup for the BCI experiment

without any real movement Duration of cue is chosen as125 seconds in accordance with the protocol mentioned in[15 26 27] Particular cue appeared randomly in order toeliminate any chance of predicting the preceding cue Forevery subject two sessions of recordings are planned Everysession is planned for eight runs Each run comprised twentytrials with different task that is imagination of left and rightarms along with state of restThus each run consists of twentyleft and twenty right imagination trials

21 Signal Recording In the experiment six channels bipolarrecording are done that is three for motor cortex area(using three bipolar channels C3 C4 and Cz) and rest threechannels are used to detect unwanted movement and toeliminate portions of recordings so as to reduce the influenceof physiological noise Electrooculogram (EOG) is recordedbipolarly in referential manner using three electrodes oneplaced medially above and two placed laterally below theleft and right eyes respectively In addition to this EMG isrecorded from the 119898 extensor digitorum communis of theright and left arms as mentioned in [27] to detect unwantedmovements if there any during recording For recordingEMG and EOG gold plated electrodes are used and EEGsignal is recorded using EEG cap with 21 Ag electrodesAll signals including three channel EEG signals EOG andtwo channels of EMG are sampled at 500Hz To filter outmains hum from the 50Hz power line a notch filter is usedSignals recorded from all six channels are grouped intothree categories as right imagery left imagery and rest stateproducing eighteen sets of one-dimensional data

22 Signal Processing These data are further grouped as pertrial and run producing twelve sets of three-dimensional dataof size 625 times 10 times 8 corresponding to left and right imagerytasks for each channel and six sets of three-dimensional dataof size 3000 times 20 times 8 corresponding to rest state for eachchannel This size is as per trial duration of 125 secondsfor imagery tasks with sampling frequency of 500Hz thusresulting in 625 data points in one trial which is also chosenas epoch length In every session there are eight runs eachconsisting of ten right and ten left imagery tasks For therest state duration of six seconds with sampling frequencyof 500Hz agrees with the size for the data corresponding to













119863 =

119870

sum

119895=1

sum

119901isin119862119895

10038161003816100381610038161003816119898119895minus 119901

10038161003816100381610038161003816

2

(1)









0 2 40

01

02

03

04

PSD

Nor

mal

ized

freq

uenc

y (H

z)

SamplesN

orm

aliz

ed fr

eque

ncy

(Hz)


0 200 400 6000

01

02

03

04

0 200 400 600

0001

Sign

al

minus001

(a)

Samples

Nor

mal

ized

freq

uenc

y (H

z)


0 100 200 300 400 500 6000

00501

01502

02503

03504

045

0 100 200 300 400 500 600

0001002

Sign

al

minus001

minus002

0 2 40

005

01

015

02

025

03

035

04

045

PSD

Nor

mal

ized

freq

uenc

y (H

z)

(b)

Samples

Nor

mal

ized

freq

uenc

y (H

z)


0 100 200 300 400 500 6000

00501

01502

02503

03504

045

0 100 200 300 400 500 600

00050

0010015

Sign

al

minus0005

minus001

minus0015

0 2 40

005

01

015

02

025

03

035

04

045

PSD

Nor

mal

ized

freq

uenc

y (H

z)

(c)

Figure 2 Continued


Time (s)

Nor

mal

ized

freq

uenc

y (H

z)


0 100 200 300 400 500 6000

00501

01502

02503

03504

045

0 100 200 300 400 500 600

00005

001

Sign

al

minus0005

minus001

0 21 30

005

01

015

02

025

03

035

04

045

PSD

Nor

mal

ized

freq

uenc

y (H

z)

(d)













27 Update 119875119887119890119904119905

and 119866119887119890119904119905



















2times (119892 (119905) minus 119909 (119905))

119909 (119905 + 1) = 119909 (119905) + V (119905 + 1)

(2)



2is a constant



max iterartions (3)




1198881 (119909) =

1199011 (119909) + 119901

2 (119909)

2minus 1205791times 1199011 (V)

1198882 (119909) =

1199011 (119909) + 119901

2 (119909)

2minus 1205792times 1199012 (V)

(4)

where 1198881(119909) and 119888








1(V) = 119901

1(V) and 119888

2(V) = 119901

2(V)




119877 =


if 119872max gt 119872

1 if 119872max = 119872min(5)


119894


119909119894+ 120597 times (119909

119894

max minus 119909119894) if 120597 gt 0

119909119894+ 120597 times (119909

119894minus 119909119894

min) otherwise(6)
















3 Results



119888= 065





119888=







0 20 40 60 80 10030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(a)

0 2010 4030 6050 8070 1009030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(b)

0 20 40 60 80 10032

34

36

38

40

42

44

46

48

Iterations

Misc

lass

ifica

tion

()

(c)









Ranking 4 289 178 133








References
































































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in















119863 =

119870

sum

119895=1

sum

119901isin119862119895

10038161003816100381610038161003816119898119895minus 119901

10038161003816100381610038161003816

2

(1)









0 2 40

01

02

03

04

PSD

Nor

mal

ized

freq

uenc

y (H

z)

SamplesN

orm

aliz

ed fr

eque

ncy

(Hz)


0 200 400 6000

01

02

03

04

0 200 400 600

0001

Sign

al

minus001

(a)

Samples

Nor

mal

ized

freq

uenc

y (H

z)


0 100 200 300 400 500 6000

00501

01502

02503

03504

045

0 100 200 300 400 500 600

0001002

Sign

al

minus001

minus002

0 2 40

005

01

015

02

025

03

035

04

045

PSD

Nor

mal

ized

freq

uenc

y (H

z)

(b)

Samples

Nor

mal

ized

freq

uenc

y (H

z)


0 100 200 300 400 500 6000

00501

01502

02503

03504

045

0 100 200 300 400 500 600

00050

0010015

Sign

al

minus0005

minus001

minus0015

0 2 40

005

01

015

02

025

03

035

04

045

PSD

Nor

mal

ized

freq

uenc

y (H

z)

(c)

Figure 2 Continued


Time (s)

Nor

mal

ized

freq

uenc

y (H

z)


0 100 200 300 400 500 6000

00501

01502

02503

03504

045

0 100 200 300 400 500 600

00005

001

Sign

al

minus0005

minus001

0 21 30

005

01

015

02

025

03

035

04

045

PSD

Nor

mal

ized

freq

uenc

y (H

z)

(d)













27 Update 119875119887119890119904119905

and 119866119887119890119904119905



















2times (119892 (119905) minus 119909 (119905))

119909 (119905 + 1) = 119909 (119905) + V (119905 + 1)

(2)



2is a constant



max iterartions (3)




1198881 (119909) =

1199011 (119909) + 119901

2 (119909)

2minus 1205791times 1199011 (V)

1198882 (119909) =

1199011 (119909) + 119901

2 (119909)

2minus 1205792times 1199012 (V)

(4)

where 1198881(119909) and 119888








1(V) = 119901

1(V) and 119888

2(V) = 119901

2(V)




119877 =


if 119872max gt 119872

1 if 119872max = 119872min(5)


119894


119909119894+ 120597 times (119909

119894

max minus 119909119894) if 120597 gt 0

119909119894+ 120597 times (119909

119894minus 119909119894

min) otherwise(6)
















3 Results



119888= 065





119888=







0 20 40 60 80 10030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(a)

0 2010 4030 6050 8070 1009030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(b)

0 20 40 60 80 10032

34

36

38

40

42

44

46

48

Iterations

Misc

lass

ifica

tion

()

(c)









Ranking 4 289 178 133








References
































































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




0 2 40

01

02

03

04

PSD

Nor

mal

ized

freq

uenc

y (H

z)

SamplesN

orm

aliz

ed fr

eque

ncy

(Hz)


0 200 400 6000

01

02

03

04

0 200 400 600

0001

Sign

al

minus001

(a)

Samples

Nor

mal

ized

freq

uenc

y (H

z)


0 100 200 300 400 500 6000

00501

01502

02503

03504

045

0 100 200 300 400 500 600

0001002

Sign

al

minus001

minus002

0 2 40

005

01

015

02

025

03

035

04

045

PSD

Nor

mal

ized

freq

uenc

y (H

z)

(b)

Samples

Nor

mal

ized

freq

uenc

y (H

z)


0 100 200 300 400 500 6000

00501

01502

02503

03504

045

0 100 200 300 400 500 600

00050

0010015

Sign

al

minus0005

minus001

minus0015

0 2 40

005

01

015

02

025

03

035

04

045

PSD

Nor

mal

ized

freq

uenc

y (H

z)

(c)

Figure 2 Continued


Time (s)

Nor

mal

ized

freq

uenc

y (H

z)


0 100 200 300 400 500 6000

00501

01502

02503

03504

045

0 100 200 300 400 500 600

00005

001

Sign

al

minus0005

minus001

0 21 30

005

01

015

02

025

03

035

04

045

PSD

Nor

mal

ized

freq

uenc

y (H

z)

(d)













27 Update 119875119887119890119904119905

and 119866119887119890119904119905



















2times (119892 (119905) minus 119909 (119905))

119909 (119905 + 1) = 119909 (119905) + V (119905 + 1)

(2)



2is a constant



max iterartions (3)




1198881 (119909) =

1199011 (119909) + 119901

2 (119909)

2minus 1205791times 1199011 (V)

1198882 (119909) =

1199011 (119909) + 119901

2 (119909)

2minus 1205792times 1199012 (V)

(4)

where 1198881(119909) and 119888








1(V) = 119901

1(V) and 119888

2(V) = 119901

2(V)




119877 =


if 119872max gt 119872

1 if 119872max = 119872min(5)


119894


119909119894+ 120597 times (119909

119894

max minus 119909119894) if 120597 gt 0

119909119894+ 120597 times (119909

119894minus 119909119894

min) otherwise(6)
















3 Results



119888= 065





119888=







0 20 40 60 80 10030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(a)

0 2010 4030 6050 8070 1009030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(b)

0 20 40 60 80 10032

34

36

38

40

42

44

46

48

Iterations

Misc

lass

ifica

tion

()

(c)









Ranking 4 289 178 133








References
































































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




Time (s)

Nor

mal

ized

freq

uenc

y (H

z)


0 100 200 300 400 500 6000

00501

01502

02503

03504

045

0 100 200 300 400 500 600

00005

001

Sign

al

minus0005

minus001

0 21 30

005

01

015

02

025

03

035

04

045

PSD

Nor

mal

ized

freq

uenc

y (H

z)

(d)













27 Update 119875119887119890119904119905

and 119866119887119890119904119905



















2times (119892 (119905) minus 119909 (119905))

119909 (119905 + 1) = 119909 (119905) + V (119905 + 1)

(2)



2is a constant



max iterartions (3)




1198881 (119909) =

1199011 (119909) + 119901

2 (119909)

2minus 1205791times 1199011 (V)

1198882 (119909) =

1199011 (119909) + 119901

2 (119909)

2minus 1205792times 1199012 (V)

(4)

where 1198881(119909) and 119888








1(V) = 119901

1(V) and 119888

2(V) = 119901

2(V)




119877 =


if 119872max gt 119872

1 if 119872max = 119872min(5)


119894


119909119894+ 120597 times (119909

119894

max minus 119909119894) if 120597 gt 0

119909119894+ 120597 times (119909

119894minus 119909119894

min) otherwise(6)
















3 Results



119888= 065





119888=







0 20 40 60 80 10030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(a)

0 2010 4030 6050 8070 1009030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(b)

0 20 40 60 80 10032

34

36

38

40

42

44

46

48

Iterations

Misc

lass

ifica

tion

()

(c)









Ranking 4 289 178 133








References
































































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




















2times (119892 (119905) minus 119909 (119905))

119909 (119905 + 1) = 119909 (119905) + V (119905 + 1)

(2)



2is a constant



max iterartions (3)




1198881 (119909) =

1199011 (119909) + 119901

2 (119909)

2minus 1205791times 1199011 (V)

1198882 (119909) =

1199011 (119909) + 119901

2 (119909)

2minus 1205792times 1199012 (V)

(4)

where 1198881(119909) and 119888








1(V) = 119901

1(V) and 119888

2(V) = 119901

2(V)




119877 =


if 119872max gt 119872

1 if 119872max = 119872min(5)


119894


119909119894+ 120597 times (119909

119894

max minus 119909119894) if 120597 gt 0

119909119894+ 120597 times (119909

119894minus 119909119894

min) otherwise(6)
















3 Results



119888= 065





119888=







0 20 40 60 80 10030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(a)

0 2010 4030 6050 8070 1009030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(b)

0 20 40 60 80 10032

34

36

38

40

42

44

46

48

Iterations

Misc

lass

ifica

tion

()

(c)









Ranking 4 289 178 133








References
































































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in
















3 Results



119888= 065





119888=







0 20 40 60 80 10030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(a)

0 2010 4030 6050 8070 1009030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(b)

0 20 40 60 80 10032

34

36

38

40

42

44

46

48

Iterations

Misc

lass

ifica

tion

()

(c)









Ranking 4 289 178 133








References
































































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




0 20 40 60 80 10030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(a)

0 2010 4030 6050 8070 1009030

35

40

45

50

55

Iterations

Misc

lass

ifica

tion

()

(b)

0 20 40 60 80 10032

34

36

38

40

42

44

46

48

Iterations

Misc

lass

ifica

tion

()

(c)









Ranking 4 289 178 133








References
































































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






Ranking 4 289 178 133








References
































































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in































































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in



classification of two class motor imagery tasks using hybrid ga-pso based k-means clustering

Documents