Study On Deep Learning Sleep Stages Based On EEG Signals

Sleep is a relatively important and indispensable item in people's physiological activity,and a good sleep quality is an important guarantee for our work,life and study.Sleep staging are the basis of sleep research,which can help people to understand the characteristics of various sleep stages and evaluate sleep quality,and play a great role in the diagnosis of sleep-related diseases.Physiological signals in different sleep stages have different characteristic changes,so physiological signals are often used for sleep staging.In this dissertation,electroencephalogram(EEG)is used for sleep staging.At present,the sleep staging research based on deep learning method has become more mature and has a higher accuracy than the traditional method.Therefore,this dissertation uses the end-to-end network model to conduct the sleep staging research.The main research contents of this dissertation include:(1)The original EEG data of Fpz-Cz channel in the selected Sleep-EDF data set was filtered after 30 s segmentation,then,the traditional sleep staging algorithm of support vector machine(SVM)based on wavelet transform is verified,which uses wavelet transform to extract the information and energy characteristics of EEG signals,and finally inputs them into SVM classifier to realize sleep stage classification.(2)This dissertation constructs the complete convolution neural network(CCNN)used in the phase of Sleep stages,dissertation adopts the CCNN network to replace the convolutional neural network(CNN),remove the pooling the CNN network layer,retains more useful features,and use the Sleep EDF data set Fpz-Cz and Pz-Oz channel20 fold cross validation experiments,it is concluded that to improve CCNN network to the improvement of Sleep stage classification with certain accuracy;In addition,on the basis of CCNN network,Long Short Term Memory Network(BiLSTM)was added to construct the CCNN-BiLSTM network model for sleep staging,as the comparison network of the CCNN-Bi GRU network in the later depth residual.(3)In this dissertation,a depth residual CCNN-Bi GRU network model is constructed,compared with the previous CCNN-Bi GRU network,the depth residual CCNN-Bi GRU network model adds residual connection to the CCNN network to enhance information features and alleviate gradient disappearance.In addition,a Mask layer was added to enhance the final feature classification result by filtering out the incomplete feature sequence.The model is tested with Fpz-Cz and Pz-Oz channels in the Sleep-EDF data set,and the final overall accuracy is 96.31% and 92.15%,respectively.The feasibility and advance of the depth residual CCNN-Bi GRU network model model in this dissertation are verified by comparing it with the high-level dissertation model in the same experimental data set in recent years.