Research On SSVEP Signals Based On Complex Network And Convolutional Neural Network

Steady state visual evoked potential(SSVEP)based on brain computer interface(BCI)has been widely studied in the field of rehabilitation,smart car,entertainment,military,smart home and so on.The human-machine interaction efficiency of mentioned systems is closely related to the classification accuracy.However,the difference of individual brain structure and appearance of fatigue symptoms will lead to the decrease of the classification accuracy,which affects the performance of BCI systems.In order to improve the general applicability of BCI system,a series of studies have been carried out in this dissertation.The main work are as follows:1)In this dissertation,an adaptive optimal-Kernel time-frequency representation(AOK-TFR)-based complex network method is proposed for characterizing fatigued behavior based on SSVEP signals.First,the experiments for subjects under normal and fatigue states are designed.Second,the classification accuracy of signals under different states is evaluated by Fisher linear discriminant analysis.Third,the brain network is constructed based on the time spectrum of EEG signals obtained from AOKTFR.The small-world-ness and nodes local efficiency of the brain networks generated under two different states are calculated to characterize the fatigue mechanism of brain under SSVEP paradigm.2)The convolutional neural network with long short-term memory(CNN-LSTM)is used to classify EEG signals of normal subjects in fatigue state and EEG illiterate subjects in normal state from a SSVEP-based BCI system.Firstly,fast Fourier Transform(FFT)is used to calculate the power of EEG signals.Secondly,the power spectrum of EEG signals is converted into two-dimensional images with spatial structure preserved.Thirdly,four different CNN-LSTM network models are used to extract the spatial,frequency and temporal features of EEG signals,and the analysis results are compared with the existing classification algorithms.The results show that the method can increase the accuracies of all subjects to about 95%.3)A joint frequency and phase modulation SSVEP excitation interface with 15 excitation targets is designed to improve the applicability of BCI system.Based on this interface,experiments are designed and the experimental data are analyzed by CNNLSTM network,and the classification accuracies of all subjects are about 90%.The results show that the CNN-LSTM network model designed in this dissertation can be used to solve the multi-classification problem of SSVEP,and the joint frequency and phase modulation excitation interface can improve the performance of BCI system.