Nowadays, the development of brain-computer interfaces, as a bridge between the brain and the world outside, has accelerated. These systems mainly assist the activities of people with severe disabilities. The motor imagery (MI) task is a widespread paradigm used in the BCI framework. In MI-based BCIs, a subject imagines the movement of a body part, and the computer detects which part is imagined based on the received signal. Electroencephalography (EEG) is a standard method of brain activity recording. In recent recording devices, the number of electrodes has increased to improve the EEG's spatial resolution. However, the increase of channels causes computational and hardware costs and also causes setup difficulties. Hence, in commercial applications, we cannot use a high number of electrodes. The recent expansion of deep neural networks in pattern recognition and classification has brought these networks into the processing and classifying of EEG signals. A diverse range of network architectures is designed for different tasks. Despite improving the performance of DNN based BCI systems, the generality and accuracy of these systems have remained a challenge. Besides, despite classic classification methods that have a wide range of comparative studies, there is a lack of such studies in deep learning methods. In this thesis, we first consider a convolutional network as the base classifier to compare the effect of pre-processing methods in the deep learning-based classification of EEG signals for MI tasks. Then, the influence of different normalization methods and filtering in different frequency bands is studied based on the performance of the base network. Also, we investigate the effect of frequency-domain transforms, such as wavelets, and combining them with the time-domain as the network's input. After that, channel selection is studied to reduce computational and hardware costs and increase the system's performance. To this end, the base network's accuracy is considered as the eva‎luation criterion. According to this eva‎luation metric, two sequential search algorithms that work faster among other search algorithms are tested. Based on them, two-channel selection methods named incremental and decremental sequential channel selection algorithms are proposed, and the effect of using them is investigated based on the base network. Then, to search a wider space in channel subsets, a genetic algorithm as a metaheuristic search algorithm is proposed due to genetic's capability to search wider and finding a better subset of channels. As the proposed genetic algorithm takes a long time, combining it with the incremental sequential channel selection method is proposed to make the genetic algorithm converges faster and to a better subset of channels. The proposed combined method has shown superiority over the state-of-the-art channel selection method, showing excellent performance on the used dataset.

شادرخ سماوي

نادر كريمي

فرزانه شايق بروجني، محمدعلي خسروي فرد