Time-frequency Based EEG Motor Imagery Signal Classification with Deep Learning Networks

In this research, a wavelet transform-based feature extraction approach with time-frequency analysis is proposed for motor imaginary EEG signal classification. The proposed approach selects specific channels such as C3 and C4 to identify event-related synchronization (ERS) or event-related desynchronization (ERD) phenomenon to filter out the artifacts and noisy data from signals. As EEG dataset is noisy and size of the dataset reduces after filtering, the proposed approach adopts multi-scale analysis ability of wavelet transform to utilize small input. It allows to extract features from the dataset and generate input images for training the models. Considering abstraction ability of Convolutional Neural Network (CNN), deep CNN with two convolutional layers, and VGGnet with six convolutional layers are employed. The model performance is evaluated in terms of accuracy, loss, and epochs. The proposed approach is applied to EEG dataset III from BCI competition II. The primary results show that VGGnet performs better than deep CNN with respect to training loss and training accuracy.