TMAnet: An Innovative Approach for Classifying Motor Imagery EEG Signals Related to Motor Movements

The Brain-Computer Interface (BCI) system utilizing Motor Imagery (MI) has diverse applications in medical rehabilitation. This study aims to validate the reliability of self-collected data and classify open-source electroencephalogram (EEG) signals. We introduce a temporal convolution network with multihead self-attention (TMAnet), designed for decoding EEG signals. This network combined with a data expansion approach based on cyclic translation strategy, which effectively expands experimental data without losing original information. To enhance the processing of self-collected signals, the EEG signals undergo the application of a spatio-temporal filtering technique and a low-rank approximation for generalized eigenvalue decomposition to effectively eliminate multiple artifacts. The proposed model could perform more accurate recognition of self-collected EEG signals than the existing methods, and the highest classification accuracy achieved was 83.75 on self-collected dataset. The effectiveness of this system was proven through the accurate classification of the self-collected MI EEG signals in this study.