International Conference on Issues and Challenges in Networking, Intelligence and Computing Technologies 2-3 September 2011KIET, Ghaziabad
work. The segment of each EEG trial between 3s and 8s(5s of an event-related signal which corresponds to 640samples) is divided into Gaussian windows dependingon the settings of two parameters: window length, length
and the amount of overlap, lap.The EEG features are calculated separately for eachtime segment. The first stage amounts to time averagingof spectral content in the specified frequency bands.Next, the square norm of these mean values iscalculated (i.e. the square root of the sum of thecomponents squared), which serves as a sub-element
of the feature vector R, where
is an index of arecording channel (i.e. either Cz, C3 or C4). The index
implies that a complete feature vector consists ofsegments of the normalized spectral EEG content foreach signal recorded:
R = (R1,.., R2*Nwin )
(1)The windowing technique allows for effective controlover the architecture of the feature space. Firstly, thenumber of windows dictates its dimension. Secondly, thewindow length along with the amount of overlapdetermines the character of information about the t-fevolution of the relevant EEG components exploited.
Among available FISs , ,
we chose the Chiu’s
FIS (CFIS) . Indeed, CFIS is robust to noise, whichis fundamental when dealing with such noisy data asEEG signals. Moreover, according to Chiu, the CFIS isgenerally more efficient than neural networks. Finally, itis a clustering-based FIS, making it suitable for dealingwith small training sets [1
5]. With the CFIS, fuzzy “if
then” rules can be automatically extracted
from data in three successive steps.1. Clustering of training data. Aclustering algorithm
known as “substractive
14] is applied to thetraining data of each class separately. This algorithmenables to find automatically the number of clusters andtheir positions.
Generation of the fuzzy rules.
A fuzzy “if
then” rule is
generated for each cluster found previously. For acluster j, belonging to class Cli, the generated fuzzy ruleis: if X1 is Aj1 and ; . . . ; and XN is AjN then class is Cliwhere N is the dimensionality of the data, Xk is the kthelement of a feature vector X and Ajk is a Gaussianfuzzy membership functionTo increase accuracy, the
membership functions can be “two
with a plateau and a different standard deviation oneach side .3.
Fuzzy rule optimization.
Each membership functionAjk is tuned according to gradient descent formulas thatuse a classification error measure E and a learning rate
Once trained, the CFIS can use its set of fuzzy rules toclassify any new feature vector X. The class assigned toX corresponds to the class associated with the rule j forwhich its degree of fulfilment is the highest.
Conception of “Hand
It is possible to add handmade fuzzy rules (HMFR) toa FIS as
knowledgeconcerning hand motor imagery EEG data concerns thepresence of contralateral ERD in the
bands .A human expert could formalize this knowledge usingsimple rules.Using schemes of trials and errors on the trainingsets, the optimal
value for λ was chosen to be 15.32 in
the four membership functions. It should be noted thatsuch rules cannot be learnt by the CFIS as theydescribe relationships between features and not theproperties of the features.
The results obtained with the fuzzy logic in classifyingeach task are on an average of
of correctclassifications, with a peak of 82% (table I).
SubjectPercentage of correctclassification
Right hand movement 82Left hand movement 79.1Right foot movement77Left foot movement 76
The error-reject curve is displayed on Fig. 4 fordifferent subjects. The error-reject curves suggest CFISis able to identify and reject efficiently the outliers, whichmakes its error rate drop dramatically.
Fig. 4. The error-reject curve for different subjects