Acquisition and implementation of new features for machine learning in EEG signals

In this thesis, the reception and analysis of signals in order to control devices or systems using the brain computer interface (BCA) is discussed. The analysis of signals that occur during brain functions is called electroencephalography (EEG). EEG signals are taken, analyzed and classified to determine which state they belong to. One of the studies carried out for this purpose is the shape prediction application using visual stimulus potential. With this application, the features of the recorded EEG signals for 4 different shapes are extracted and classified. In the study on movement, the features are extracted and classification is carried out in order to decide which state they belong to by taking the EEG signals in a coordinated way by moving the eyes and arms together. In our studies carried out using a ready-made alcoholic data set, statistical features are extracted first, and a hybrid feature vector is created by extracting deep features. By classifying this created feature vector, it is decided which group the person belongs to. In our study of another alcoholic data set, moment properties are extracted and higher order moments are calculated as well as the first 4 moments whose meanings are known, and their effects on classification are examined. In the study for shapes, the classification of EEG signals determined which shape was displayed with 99.99 percent accuracy. These results show that different signals are produced in the brain according to the structure of the image displayed. In the application using the alcoholic data set, the accuracy rate was 81.2 percent for only the classification of statistical features, and 95.71 percent for deep learning only, but 99.2 percent for hybrid features derived using the recommended Deep - Statistical Features Classification (DSFC). For the other alcoholic study, the highest classification accuracy is 99.60 percent when all features from Moment 1 to 120 are used, while it is 99.80 percent from Moment 1 to 20, and it seems to give better results. Accordingly, it was concluded that EEG signals can be represented by a 20th-order polynomial. As a result, it is seen that alcohol use disorder (AUD) causes different EEG signals in people exposed to visual stimuli compared to normal people. Keywords: Electroencephalography, Alcohol use disorder, Hybrid feature, Classification, Spectrogram, Momen