Research On EEG Signal Recognition Method Based On Convolutional Neural Network

Brain-computer interface technology(BCI)decodes the EEG signal of the patient’s hand motor imaging collected by the electroencephalogram,and converts it into control instructions to control the hand rehabilitation equipment,thereby driving the hand for rehabilitation training,and continuously activating the motor function area of the patient’s brain to reconstruct the damaged cerebral cortex,and finally achieve the purpose of rehabilitation.In the BCI rehabilitation system based on motor imaging-EEG signal(MI-EEG),the most important are MI-EEG preprocessing,feature extraction and recognition classification processes.Due to the inevitable introduction of the most difficult to remove Electrooculography(EOG)in the process of collecting EEG signals,it is difficult to characterize MI-EEG and low classification recognition rate.In order to solve these problems,this thesis studies the pretreatment and classification identification of MIEEG,and the main work is as follows:(1)In view of the problem that EOG is difficult to remove in MI-EEG,this thesis proposes an ocular electrical artifact removal method based on improved discrete wavelet transform-independent component analysis-cosine criterion-signal matching(IDWT-ICA-CC-SM),which decomposes EEG signals by using discrete wavelet transform combined with independent component analysis,and uses cosine criterion to judge the pseudo-traces and signal matching mechanism to solve the defect that the components obtained by independent component analysis do not match the position order of the source signal.By experimenting on a mixed signal with linear superposition of pure EEG signals from BCI competition II datasets III and BCI competition IV datasets 2b,and using mean squared error and signalto-noise ratio evaluation indicators to measure the advantages and disadvantages of the proposed method,the experimental results show that the proposed method for removing ocular electrical artifacts is the best.(2)Aiming at the problem of difficult extraction of MI-EEG features and low recognition rate,a convolutional neural network(CNN)MI-EEG classification recognition method based on attention module is proposed,that is,Efficient channel attention(ECA)module and Coordinate attention(CA)module are inserted in CNN respectively.This method combines CNN adaptive extraction and attention module to pay attention to the advantages of MI-EEG-related feature signals,so that it can improve the accuracy of model recognition without increasing the model calculation overhead.Experiments are carried out on BCI competition II datasets III and compared with the methods of EEG classification using CNN models in recent years,and the experimental results show that the proposed method can obtain better recognition accuracy in MI-EEG classification recognition.In addition,compared with the ECA module,the CNN MI-EEG classification and recognition method based on CA module is more stable.(3)The data sets based on DW-ICA,DWT-ICA-CC,IDWT-ICA-CC,IDWT-ICA-CC,IDWT-ICACC-SM were applied to CNN and CNN+CA models respectively,and the experimental results showed that IDWT-ICA-CC-SM corresponds to the highest recognition accuracy on CNN+CA models and CNN models.(4)In order to obtain the optimal parameter combination of CNN+CA model,a CNN+CA parameter optimization method based on symbiotic biological search is proposed to optimize the structural parameters of convolutional layer and pooling layer,and the experimental results show that when the CNN+CA model is set with the optimal solution as the structural parameter,it can reach the maximum test recognition accuracy with the least number of iterations.At the same time,this method greatly reduces the parameter adjustment time,and avoids the inability to obtain optimal structural parameters due to insufficient human experience and knowledge.