Feature extraction and classification of electroencephalographic (EEG) signals towards the use of brain-computer interface in cognitive applications

In this thesis, the brain computer interface system is developed for the cursor movement through the cognitive signals. The EEG data is handled from the Emotiv Neuroheadset device via our developed program written in c# language. We have worked on classification of five cognitive tasks; up, down, left, right and no movement. Before the movement of the cursor, the participants need to be trained to control the brain signals. In the training phase, the training screen is designed to consist of the visual stimuli. The participants have trained the program three times in different days and each session includes 24 trainings. EEG signals are very complex and the extracting information is difficult. Also, EEG signals has many artifacts occurred by the eye movement, muscle movement and the noise in environment. Therefore, median filtering and the normalization method are used in the preprocessing phase. Then the specific features for all cognitive tasks are extracted by the multifractal detrended fluctuation analysis and the fast Fourier transform. The Ph values and beta signals calculated from the MFDFA and FFT methods respectively, are used as features. Finally these features are classified by the nearest neighbor algorithms. Nearest neighbor (NN) algorithms are simple but effective methods for performing pattern classification. The CxK nearest neighbor algorithm is firstly used for cognitive EEG signal classification in this thesis and this method has given acceptable results when compared with the other studies in literature. Keywords: EEG, BCI, feature extraction, classification, k-nearest neighbor algorithm, CxK-nearest neighbor algorithm.