Motor Imagery Classification Based On Deep Learning And Whole-brain EEG Muti-domain Features

Brain-computer interface(BCI)systems have recently attracted much more attention in the medical field.It can bypass the human’s normal peripheral-nerve pathways and controls external devices directly.Motor Imagery(MI)-based BCI systems can monitor and extract the changes of EEG signals during the MI process,recognize different MI tasks through feature extraction and pattern recognition,complete the output of control instructions,and become a bridge between the human brain and external devices.Due to the low spatial resolution and low signal-to-noise ratio of EEG,how to effectively extract and accurately classify the features of MI-EEG signals is the focus of MI-BCI technology research.Aiming at the MI-BCI system,this dissertation focuses on the new MI-EEG feature extraction and classification methods.We proposed several effective MI-EEG feature extraction and classification models.The main work is divided into the following aspects:1.To handle the MI-EEG signals directly,we proposed a deep learning model that applies the local reparameterization trick in the convolutional neural network.The results show that the average accuracies of 20 subjects,50 subjects,80 subjects,and 109 subjects are 92.98%,98.53%,92.86%,and 92.37%,respectively.2.We developed a novel model,combining the recurrence plot(RP)and Bayesian Convolutional Neural Networks(BCNNs)for real execution/motor imagery classification.Twenty-five subjects were used on the Physio Net dataset to verify the model’s effectiveness.The results show that the average accuracies of 2-class,3-class,4-class,and 5-class during real execution are 97%,96.96%,95.76%,and 96.09%.As for motor imagery are 96.77%,96.52%,95.25%,and 96.28%.3.To analyze the differences between motor imagery and real execution from the perspective of the spatial domain,we used Phase-Locking Value(PLV)and Phase Lag Index(PLI)to construct functional brain networks.We statistically analyzed the characteristics of brain function networks corresponding to different motor tasks by calculating the clustering coefficient,degree distribution,and other complexity measures.Furthermore,we studied the functional connectivity between different regions during real execution and motor imagery.In the classification stage,we generated the Spatio-spectral features by fusing PLV and PLI matrices under different filter subsets,and the 3DCNN-LSTM model was used for features recognition and classification.The results show that the average accuracies of 20 subjects,50 subjects,80 subjects,and 103 subjects are 83.09%,76.3%,75.02%,and 74.54%,respectively.