Research On Brain-computer Interaction Based On Visual Imaging EEG Signals

Brain computer interface(BCI)is a kind of communication system which controls computer or other external devices by measuring brain signals.It aims to bypass peripheral nerves and muscles of brain to realize direct communication between brain and external world.It is a new way of human-computer communication.This technology is expected to provide alternative and new communication or control methods for patients with severe sports disability or healthy people in specific situations.In brain computer interface(BCI),the traditional imagination task is motor imagery,which requires the subjects to imagine moving a certain part of their body.This kind of imagination task has certain difficulty,even there is motor imagery blindness,which makes this kind of BCI difficult to be applied.In this study,BCI was constructed by a less studied and easily completed psychological imagination(visual imagery).However,the classification of this kind of BCI is very difficult,so it is necessary to design effective visual imagery task combination and explore effective feature extraction methods.The main contents of this paper are as follows:(1)The first experimental paradigm of this paper is designed.EMD and AR model are combined to extract EEG features of visual imagination.EMD is suitable for the analysis of non-linear and non-stationary signals.In this study,EMD is used to decompose EEG signals into stable Intrinsic Mode Function(IMF)functions,and then AR parameters are established for each IMF component,then AR parameters and residual variance are extracted as eigenvectors.In addition,the AR model with high resolution in time domain is used to extract the auto regressive parameter features of EEG signals,and Hilbert Huang transform(HHT)with high resolution in time domain and frequency domain is used to extract the average instantaneous energy features of EEG signals for comparison.(2)The second experimental paradigm of this paper is designed,which uses the EEG data collected by microstate analysis.In this paper,an improved k-means clustering algorithm is used to analyze the micro state of EEG signals,and the EEG of18 subjects who visually imagined the lower limb movements is studied.Four typical microstates were extracted from multichannel EEG signals.At the mean level,the microstate time series parameters of the two kinds of signals are analyzed and compared,and the differences between them are obtained.The results showed that the most significant difference between the two VI tasks was in microstate 1 and microstate 4.(3)The construction and characteristic analysis of brain network.In this paper,32 electrodes are selected as network nodes to construct brain function network with mutual information and coherent calculation correlation coefficient.The graph theory analysis method is used to calculate the node degree,clustering coefficient,feature path length,global efficiency and local efficiency of brain network.The feature vectors are constructed by the network attribute characteristics and the spatial characteristics of adjacency matrix of different dimensions.It is found that there are significant differences in node degree,clustering coefficient and feature path length between the two types of VI tasks,but little differences in global efficiency and local efficiency.From the perspective of adjacency matrix,the difference of total average adjacency matrix of EEG data based on mutual information is more obvious.From the perspective of brain function network,the main core nodes of brain network are F8,FT8,FCZ,CP4,tp8,po7 and O1,that is,the connection from right forehead to left occipital lobe.(4)Support vector machine(SVM)was used to classify EEG signals of two kinds of visual imagery tasks.The above features are used as input features of SVM,and the classification results are obtained and compared.The results show that the combination of EMD and AR model is better than HHT and AR model,and the average classification accuracy is 78.06 ± 2.07%.In this paper,the visual imagery paradigm of lower limb movement is designed.The time parameters of microstate 1and microstate 4 with significant difference are extracted as feature vectors by microstate method.The results show that the average classification accuracy of microstate is 80.6 ± 2.58% respectively.Based on the EEG mutual information and the spatial characteristics of the coherent adjacency matrix,the visual imagery of the lower limb motion is recognized.The results show that the spatial features based on EEG mutual information and coherent adjacency matrix are the effective features of the VI task.The average classification accuracy of the spatial features constructed by8-D mutual information adjacency matrix and 6-D coherent adjacency matrix is90.12 ± 5.43% and 87.47 ± 7.01%,respectively.This study is expected to provide ideas for the construction of a new on-line visual imagery brain computer interface for the rehabilitation of lower extremity motor disorders.