Research On EEG Signal Recognition Of Motor Imagery Based On Deep Learning Method

With the development of science and technology and the advancement of data analysis capabilities,brain-computer interface(BCI)technology has become increasingly mature,which has been applied in various fields such as medical rehabilitation,aerospace,life and entertainment,etc.It is also a hot research topic in the field of engineering technology and medical rehabilitation.In this paper,the deep learning method is used to study the motor imagery EEG signal of the spontaneous EEG signal.The main research is as follows:(1)A denoising method based on improved wavelet soft threshold is proposed.Aiming at the problem of that interference noise often appears in the EEG signal acquisition,a new wavelet threshold selection rule and threshold function are adopted over the traditional wavelet soft threshold denoising method.The signal-to-noise ratio and root mean square error are taken as evaluation indexes.The improved soft threshold algorithm not only improves the shortcomings of traditional methods,but also obtains higher signal-to-noise ratio and lower root mean square error comparing with the traditional threshold processing method.(2)A recognition method of motor imagery EEG signal based on boosted convolutional neural network is proposed.First of all,STFT is used to obtain time-frequency image of EEG.Secondly,the CNN model is designed and dropout layer is added to prevent the network training from over fitting.The random gradient descent algorithm with momentum is used to speed up the parameter updating.Then,the CNN model is combined with the improved Ada Boost algorithm.Several pre-trained CNN models are used as the base learners.The improved Ada Boost algorithm will update the base learners automatically according to the classification error rate,and the boosted convolution neural network model is obtained by linear combination of the base learners.Finally,boosted model and voting method are used to classify the test data.The boosted model verification is carried out on the BCI competition IV 2b data set.In the case of selecting the first three sessions for each subject,the average accuracy rate and the average standard deviation are 76.4% and 3.7%respectively,which are better than the other two traditional methods;the kappa value is 0.63 with all the sessions are selected,which is 0.3 higher than that of the winner of the competition.The results proves that the proposed method is effective in motor imagery EEG recognition,which can enhance the robustness of the classification model and improve the classification result.(3)A multi-channel motor imagery EEG recognition method based on CNN and LSTM is proposed.Firstly,wavelet packet transform is used to filter ERD / ERS’s related characteristic frequency bands,and sliding window clipping is used to obtain multi-frame images in time dimension.Then the CNN model is constructed,and the multi-frame images are converted into the feature sequence matrix which is used as the input of the LSTM model;the LSTM model is constructed,and the Bi LSTM unit is used in the hidden layer,which can obtain the association information of the time series from the front to the back as well as the information from the back to the front.Finally,the CNN-LSTM model is used to classify the test data.The model verification was carried out on BCI competition IV data set 1.The average classification accuracy of the four subjects was 87.4%,and the average standard deviation was 2.7%,which were better than other comparison models.The results prove that the model can effectively improve the accuracy of binary classification,and has strong robustness.(4)A CNN-LSTM model EEG recognition research based on wavelet coherence analysis is proposed.In order to extract more correlation information between channels and improve the accuracy of the model,the correlation information between channels is obtained by wavelet coherence analysis from the perspective of brain functional network.The average classification accuracy is 90.3%,the average standard deviation is 2.5%,and the accuracy rate is improved by2.9%.The results prove that the model is further optimized and the classification performance is improved.