Research On Key Technologies Of Brain Computer Interface Based On Motor Imagery

Brain computer interface(BCI)is a new technology to establish communication between brain and peripherals.It has a broad application prospect in rehabilitation,military and personal entertainment.Among them,Motor Imaging(MI)signal,as an active human-computer interaction paradigm.The motor state has formed a fixed pattern in the brain,and distinguishable signals can be obtained without training.So,it has become one of the hot research directions in many BCI systems.However,due to the complexity,dynamic and low signal-to-noise ratio of Electroencephalogram(EEG),the classification accuracy of MI-BCI system is low,cross subject model is difficult and online application performance is poor.In this paper,Convolutional Neural Network(CNN)combined with self-attention mechanism and graph embedding method are used to design two end-to-end networks for accurate analysis of four classification MI signals.The MI-BCI online platform has been developed.(1)Aiming at the problem of low classification accuracy of MI-BCI system,this paper analyzes the temporal and spatial correlation of EEG signals based on the temporal and spatial characteristics of EEG,and designs CNN with parallel spatial-temporal self-attention mechanism to classify EEG.Specifically,taking advantage of the characteristics of selfattention mechanism,two self-attention modules are designed by capturing the elements which has strongly relevance to the certain task and combining with the temporal and spatial information of EEG.The temporal attention module enhances the temporal representation ability of the original signal by learning the similarity between time steps and recoding the EEG in time dimension.Similarly,the spatial self-attention module completes the global channel selection by giving higher weights to the MI related channels.The encoded EEG is spliced with the original EEG,and CNN is used to extract the high-level local features in the EEG.Finally,fully connected layer is used for classification.The classification accuracy is used to evaluate the performance of the model basing on the 2A data set of the fourth BCI competition and high gamma data set.The results are 78.51% and 97.68% respectively.(2)Because EEG is a dynamic signal,the signals of different subjects are different,and the signals of the same subject at different times are also different.So,the design of MI-BCI cross subject model has always been a challenging topic.In this paper,from the spatial point of view of the signal,referring to the distribution of electrical poles of 10-20 system collected by EEG,the method of graph is used to embedding MI signal.Compared with the traditional twodimensional arrangement of EEG signals,the special data structure of the graph is more consistent with the arrangement of electrode points in the EEG signal space.Therefore,when designing the adjacency matrix of the graph,the value between neighbor nodes is one,and the value between non neighbor nodes is set to zero.Then,the normalized adjacency matrix is used to embedding the original signal.Finally,CNN is used to classify the signals.A cross subject experiment was conducted in the 2A data set of the fourth BCI competition.The experimental results show that this method can achieve an accuracy of 63.37%.(3)Aiming at the problem of poor on-line performance of MI-BCI system,this paper designs an on-line MI controlled UAV platform.The platform uses the trained CNN model to realize the online decoding of MI signal to control the Air Sim virtual UAV.This system successfully applies the self-attention CNN model proposed in this paper to the online system,which proves the availability of this method.At the same time,the test shows that the BCI system is fast,stable and safe.To sum up,this paper not only improves the accuracy of signal classification,provides a research direction for cross subject model,but also proves the reliability of online application of deep learning system.