Font Size: a A A

Feature Extraction And Classification Of Imaginary Movements In EEG

Posted on:2013-10-29Degree:MasterType:Thesis
Country:ChinaCandidate:L J LiFull Text:PDF
GTID:2248330374475373Subject:Biomedical engineering
Abstract/Summary:PDF Full Text Request
Electroencephalogram(EEG) is the recording of electrical produced by the firing ofneurons within the brain,therefore it contains plenty of brain state information.Analyedquickly and efficiently on cognitive function,which is great signigicance for the detection andtreatment of cognitive disorders. Event-related potentials(ERP) is widely used in cognitivescience. but it’s not accurate because of losing some cognitive information when we get ERPthrough average some EEG.Cognitive activity not only lead to ERP,but aslo occurred to theevent-related desynchronization/synchronization(ERD/ERS).The ERD/ERS principle is thephysiological basis of EEG signal processing in this paper. The imaginary movements whichnot require structured physical environment is a widely used mental tasks,this paper studiesEEG which base on two types and multi-class imaginary movements ferture extraction andclassification algorithms.(1) EEG features extraction based on wavelet package decomposition coefficients andsubspace energy.At first,Using the AR model to get the signal frequency spectrum,andanalyze the ERD/ERS phenomenon difference of imagining left hand moving and right handmoving;second, decompose and reconstruct signal frequency range of the ERD/ERSphenomenon by wavelet packet;finally,Calculate C3、C4channel signal which alreadyreconstruction energy respectively,and the signal energy difference as a characteristic value.(2) The EEG features extraction based on the CSP. For the multi-channels EEG, in orderto extract better features value, we need to take into account the airspace characteristicsinformation of signals among the channels.So feature extraction of multi-channels EEG basedon time,frequency and spatial information. CSP algorithm concern the difference of the spatialenergy distribution between the two types of movement pattern, that is to build a spatialfilter,and make the variance difference between the two types of tasks in different directionsprojection. For the four-class tasks,Extend CSP with”one-to-rest” strategy to build fourspatial filter for signal feature extraction.(3) The linear classifier and support vector machine classify respectively the two typesof movement, the best results are both reached87.86%. For the multi-class recognition, usingsupport vector machine classification based on decision tree method, the best classificationaccuracy reached92.78%,Which is better than the “one-to-one” SVM classifier.If theclassifer is not good because sample of subjects is not sufficient, we can combine with support vector machines and Bayesian classifier, and expand the training set that using the testsamples which is the classification results with a large probability, and finally re-use ofsupport vector machine classification. The results show that increasing the training samplecan effectively improve the classification accuracy of the test set.
Keywords/Search Tags:Event-related desynchronization/synchronization, Wavelet package, Commonspatial patterns, Support vector machine, Sample expansion
PDF Full Text Request
Related items