Study On Four Classification Of Motor Imagery EEG Based On CNN+LSTM

Brain computer interface realizes the direct interaction between human brain and external devices by encoding and decoding EEG,which is of great significance for patients with motor nerve injury.EEG signal classification has always been a very important topic in the field of brain computer interface.Among them,the research of EEG signals based on motor imagination has been widely concerned.In the traditional research of EEG decoding,it usually has to go through two processes: feature extraction and feature classification,and the factors leading to human selection account for a large part.With the successful application of deep learning in the field of brain computer interface,the form of EEG decoding based on neural network,that is,end-to-end spontaneous learning,has become the mainstream of research.However,due to the low signal-to-noise ratio and non-stationary characteristics of EEG signals,the classification accuracy is often not ideal,especially the accuracy of multi classification is still relatively low.In order to improve the classification accuracy of multi classification motor imagery EEG,a four classification motor imagery EEG signal recognition method based on convolutional neural network(CNN)and long short term memory(LSTM)network model is proposed in this thesis.It includes the design and comparison of three four classification network models.The specific work is as follows:(1)In order to explore the effectiveness and feasibility of a single network model for the recognition of four types of motor imagery EEG signals,two four classification network models-four classification CNN(FCCNN)model and four classification LSTM(FCLSTM)model are proposed in this thesis.Firstly,the raw EEG signal is transformed into a form suitable for model input after preprocessing and data enhancement.Then two network models are designed respectively,and the method is tested on the data set 2a of BCI competition IV.The experimental results show that the classification effect of fccnn model on four types of motor imagination EEG is better than that of FCLSTM model.The average classification accuracy of FCCNN model is about 80%,while that of FCLSTM model is only about 70%.Finally,the experimental results are deeply analyzed,and the relevant factors affecting the classification effect of the model are summarized.(2)In order to fully learn the EEG features of motor imagery EEG signals and further improve the accuracy of four classification,a recognition method of CNN-LSTM hybrid network model based on EEG features is proposed in this thesis.Firstly,the raw EEG signal is preprocessed to eliminate the ocular electrical interference,and then the signals of each channel are extracted by fast Fourier transform ? Wave(0.5-3hz)? Wave(4-7hz)? Wave(8-13hz)? Wave(14-30hz)and ? Wave(31-50hz)five EEG frequency bands related to motor imagination EEG,and the Power Spectral Intensity(PSI)of each frequency band is calculated.Secondly,the frequency band PSI of each channel is extracted by multi-layer one-dimensional convolution,and then the EEG features are input into the LSTM network to extract the temporal features.The results of the classifier are divided into four categories: the left hand,the right hand and the tongue.Finally,the results of the classifier are input to the softmax network.This thesis uses the data set 2a of BCI competition IV competition to verify the proposed method.The experimental results show that this method can effectively improve the accuracy of four classification.The average accuracy of four classification is90.38% and the average kappa coefficient is 0.87.Compared with other excellent methods using this dataset,this method has a higher performance in the four classifications.