Comparison of Different Data Augmentation Methods With an Experimental EEG Dataset

Electroencephalogram (EEG) signals are small amplitude, complex and high noise components. For the correct interpretation of these signals, powerful noise removal, advanced signal processing, data multiplexing and feature extraction methods are needed. Within the scope of this study, spectral features of EEG signals recorded while 30 volunteers were sniffing four different essential odors (lavender, rosemary, rose and peppermint) within the framework of an experimental procedure were analyzed separately, without data multiplexing and using four different data multiplexing methods such as Amplyfing all time data, Time shifting, Time Stretching and Gaussian Noise Addition were tried to be classified and success rates were compared. As a result of the data obtained, the highest performance was found with the amplifying all-time data augmentation method, and four different odor data could be classified with 88.9% accuracy with CART and 97.0% accuracy with KNN.