Research On Motor Imagery EEG Classification Based On Graph Convolutional Network

Brain Computer Interface(BCI)is an emerging technology,which allows patients with motor dysfunction,such as stroke patients who have motor function loss or partial loss but whose brain is still working normally,to control wheelchairs,mechanical arms and other auxiliary equipment through EEG signals without the help of traditional interfaces such as mouse or keyboard,so as to improve their independence and the quality of their daily lives.Because Motor Imagery(MI)EEG signal has the characteristics of non-invasive,low cost,high time resolution and spontaneity,and the motor imagination ability of patients with motor disorders is not significantly different from that of normal people,the BCI system based on Motor Imagery EEG signal(MI-EEG)has been widely concerned and studied in medical and non-medical applications.The core problem of MI-BCI system is how to accurately decode MIEEG.However,MI-EEG has the characteristics of low signal-to-noise ratio(SNR)and large individual difference,so MI-BCI system has low classification accuracy on subject-dependent and poor performance on subject-independent.To address the above issues,the main contents of this thesis are as follows:(1)Aiming at the problem of low classification accuracy caused by low signal-to-noise ratio of MI-EEG,the model based on spatio-temporal adaptive graph convolution network is proposed for subject-dependent in this thesis.The model uses the raw EEG signal as the input to extract features and classify them from the spatial-temporal dimension.The model includes four modules: spatial adaptive graph convolution module,temporal adaptive graph convolution module,feature fusion module and feature classification module.First,the spatial distribution of ERD/ERS is different when the same subject performs different motor imagery tasks,and when the subjects imagined motion,any channels with similar motion-dependent features could promote each other.Therefore,using spatial distance as a measure to generate graph representation is easy to ignore the role of other nodes with distant spatial positions but similar functions,resulting in the loss of spatial information.The spatial adaptive graph convolution module proposed in this thesis automatically constructs the spatial graph representation through the feature similarity between channels,rather than through a priori knowledge or artificially constructing a fixed graph representation.Then,graph convolution is used to strengthen the spatial feature of MI signal by aggregating the features of adjacent nodes.Secondly,because of low SNR of MI-EEG and a single time point is susceptible to noise,the time series is sliced into different time segments.And the similarity between time segments and time segments is calculated to eliminate the effect of noise.The time adaptive graph convolution module uses the similarity between time segments as a measure to adaptively construct the time graph representation of MI-EEG.Then,the feature of temporal dimension is extracted by graph convolution,which aggregates the features of different time segments.The feature maps with two dimensions of spatial-temporal are input into the feature fusion module to merge and concatenate,and finally input to the feature classification module for classification.The proposed model is evaluated in subject-dependent classification experiment on dataset BCIIV2 a,and the average classification accuracy is 90.45% by using the 10-fold cross validation method and the average classification accuracy is 91.64% when verified on dataset HGD,which proves the effectiveness of the model.And the experimental results show that the average accuracy of using transfer learning method is 1.66% higher than that without transfer learning method,which proves that transfer learning method can achieve the effect of data augmentation.At the same time,subject-independent experiments are carried out on the model,and the average classification accuracy is 66.07% on dataset BCIIV2 a.The experimental results show the robustness of the proposed model.(2)Aiming at the problem of poor performance of the model in subject-independent caused by the large individual differences of MI-EEG,this thesis proposed a self-attention network based on space-time-frequency feature fusion for subject-independent classification.The features of EEG signals from different MI tasks of the same subject can be fully extracted and classified in the spatial-temporal dimension.However,EEG signals of the same MI task of different subjects have great differences in the spatial-temporal dimension.Therefore,the model which only considers the spatial-temporal features has poor adaptability on subjectindependent.The frequency feature distribution of EEG signals of different subjects in the same MI task is similar,so the proposed model introduces the frequency domain information of EEG signals to solve the problem of large individual differences.The raw EEG signal and EEG timefrequency maps are used as network input to explore the MI-EEG decoding ability of the model on subject-independent.The input module includes two parts: raw EEG signals after preprocessing and the time-frequency feature maps of each channel by Morlet wavelet transform.The frequency feature extraction module extracts frequency domain features from the three-dimensional time-frequency map,and the spatial feature extraction module and temporal feature extraction module extract spatial and temporal features from raw EEG signals.The self-attention mechanism of the three modules strengthens the MI related features in the three dimensions of frequency,space and time,while the irrelevant features are ignored.Then,the features obtained from the feature extraction module are input into the feature fusion module for fusion,and finally the fused features are input into the classification module for classification.The proposed model is evaluated by subject-independent classification experiments on the dataset BCIIV2 a,and the average classification accuracy is 67.04%,which is superior to current mainstream methods.The experimental results show that it is more beneficial to improve the subject-independent classification performance when taking into account the three dimensions of spatial-temporal-frequency of MI-EEG.