EEG Motion Classification Combining Graph Convolutional Network and Self-attentiion

The study of EEG motor imagery adds a new therapeutic approach for patients with motor disorders, and the key to the problem study is how to improve the classification recognition of EEG motor imagery. The complex characteristics of EEG signals and the existence of multi-channel spatio-temporal properties increase the difficulty of their feature extraction and classification. There are spatial correlations between different channels and temporal correlations between different time series signals, so the selection process of signal features is complicated, resulting in low recognition rate. In this paper, we propose a spatial graph convolutional neural network based on a self-attentive mechanism. For the spatial characteristics of signals with different channels, we extract spatial features by constructing a graph structure and then by information aggregation; for its temporal characteristics, we use time slicing to calculate the importance weights of different time periods in the input signal by using the self-attentive mechanism, and then update the time segments by weighting and summing, so as to minimize the influence of other interfering signals, complete feature extraction and improve the The classification recognition rate is improved. From the experimental results, the recognition rate of this model reaches over 88% in the existing open EEG motion imagery dataset, which has good practicality and applicability.