Research On Motor Imagery EEG Classification By Multidimensional Feature Analysis Based On Graph Convolutional Networks

Motor Imagary Electro-Encephalo Gram(MI-EEG)has received a lot of attention in the medical field as the number of patients with motor disabilities continues to increase.Brain Computer Interface(BCI)is an emerging technology that can directly control assistive devices through the recognition of MI-EEG,thus allowing patients with motor disabilities to establish contact with the outside world and help them repair and enhance their motor functions through aids such as wheelchairs.Therefore,there is great value in the study of MI-EEG,and the key to the research lies in how to improve the classification and recognition rate of MI-EEG.However,the complex characteristics of EEG signals increase the difficulty of feature extraction and classification,which can be summarised as follows: the original EEG contains many impurities,such as electrooculography and electromyography;the EEG has multi-dimensional spatio-temporal characteristics,which cannot be considered only in a single dimension;and there are certain differences in the selection of features in the spatio-temporal dimension between different single participants.To address the challenge of motor imagery EEG classification,this study adopts a multi-dimensional fusion deep learning method to extract the global features of EEG,completing the feature extraction of MI-EEG from three dimensions: spatial,temporal and frequency,to achieve accurate classification.The main work is as follows.(1)To address the problem of single dimension of feature information extracted from MI-EEG,a motion imagery EEG classification method based on spatio-temporal feature analysis of graph convolutional networks is proposed.Taking graph convolutional networks as the underlying framework,the method achieves spatio-temporal feature extraction for MI-EEG by fusing a self-attentive mechanism.Specifically,based on the multi-channel spatial dimension of MI-EEG,the spatial domain features of MI-EEG are extracted using graph convolutional network;based on the different temporal association dimensions of the same channel of MI-EEG,the encoder structure of Transformer is used to extract its time-domain features using the self-attention mechanism.Finally,the extracted features are fed into the fully connected layer to complete the classification process.(2)To address the problem that EEG signals from different participants present variability in spatio-temporal feature dimensions,a motor imagery EEG classification method based on multidimensional feature analysis of graph convolutional networks is proposed.The method uses wavelet packet decomposition reconstruction technique to decompose the MI-EEG into different frequency band ranges.The different frequency band energy values are calculated and reconstructed to form new correlation information to complete the information reconstruction of the MI-EEG in frequency dimension.Finally,the reconstructed information data are fed into the graph convolutional network for subsequent processing.In summary,this thesis proposes a fused deep learning model that simultaneously combines wavelet packet transform,graph convolutional network,and self-attentive mechanism to extract feature information at different levels with their respective advantages,and finally achieves an average accuracy of 90.35% in the international public dataset EEG Motor Movement/Imagery Dataset.The experimental results show that the fusion model proposed in this study has a good classification recognition rate for MI-EEG.