Motor Imagery and Mental Arithmetic Classification Based on Transformer Deep Learning Network

Non-invasive electroencephalography (EEG) signals find widespread application in brain-computer interfaces (BCI), with paradigms such as motor imagery (MI), mental arithmetic (MA) and emotion recognition being particularly common. This study aims to explore feature extraction and classification methods for MI and MA tasks. We employed convolutional neural networks (CNN) combined with Transformer networks and Fasternet Block to extract features. Through our reaearch, we obtained features for MI and MA tasks, and used softmax classification for binary classification of these tasks. We conducted on a publicly available dataset consisting of data from 29 subjects. The experimental results demonstrate that our method achieved high classification accuracy in MI and MA tasks. The final accuracy rates for MI and MA tasks were 88.67% and 91.23%, respectively.