Research On Classification Of EEG Signals In Motor Imagery Based On Neural Network

Brain-computer interface is a direct communication channel between the brain and external devices,which can decode EEG signals and control external devices through EEG signals.This technology has developed rapidly in recent years,not only to help people in medical rehabilitation,but also has a wide range of applications in many fields such as military,education and aviation.The key to brain-computer interface lies in the classification of EEG signals,and the accuracy of the classification determines the performance of the brain-computer interface system.Motor imagery EEG signals are spontaneously generated without external stimulation and are one of the most frequently used EEG signals in brain-computer interface systems.However,due to the non-linear and non-smooth characteristics of the signal itself,the classification accuracy of the current motor imagery EEG signal is low.In order to improve the classification accuracy,two classification models based on neural networks are designed in this paper,and the specific research work is as follows:(1)Aiming at the problem of insufficient signal data volume,we designed a GAN and CNN-based motor imagery EEG signal classification model,which is named WGPConv Net.Unlike the traditional EEG signal classification,the WGP-Conv Net model introduces WGAN-GP for data augmentation of feature data after feature extraction,and WGAN-GP is able to learn the distribution of input data and pack the noise into real data to realize the expansion of data volume.After data augmentation,WGP-Conv Net uses AConv Net and B-Conv Net for classification.A-Conv Net can effectively extract features of the data for quadruple classification and B-Conv Net can effectively extract features of the data for binary classification.Experimental results on BCIC IV dataset 2a and dataset2 b show that the quadruple classification of WGP-Conv Net is 74.3%,and the binary classification of WGP-Conv Net is 83.4%.Both of the results are better than some existing methods.The comparison experiments on different feature data show that the data augmentation of WGP-Conv Net is effective for different feature extraction methods.(2)Aiming at the problem of spatial feature loss,we disigned a motor imagery EEG signals classification model based on CNN,GRU and attention mechanism,which is named 3D-DGCNN.The model uses wavelet packet decomposition and 3D transformation to extract features from the data.The wavelet packet decomposition can extract the time-frequency features of the signal,and the 3D transformation can restore the relative position relationship of each channel data according to the sampled electrode map of EEG signals,and calculate the the spatial features of the signal.After feature extraction,3D-DGCNN uses T-DGCNN with tree structure and P-DGCNN with parallel structure for classsification.The experimental results on BCIC IV dataset 2a show that the average classification accuracies of the two models are 74.9% and 74.6% respectively,and the highest classification accuracy for a single object can reach 89.7%,which is better than some existing methods.Meanwhile,it is concluded from the comparison experiments that the classification accuracy of the data after wavelet packet decomposition and 3D processing is significantly higher than that of the original data,and this processing can effectively extract the time-frequency features and spatial features of the data.