Research On Classification Of Motor Imaginary EEG Signals Based On Deep Learning

Brain Computer Interface(BCI)is a human-computer interaction technology that establishes direct communication between the brain and computers or other electronic devices,without relying on any muscles and nerves of the body but only the brain.This technology has broad application prospects,including rehabilitation of disabled,exploration of human's brain,and faster targets detection.According to different ways of generating EEG,BCI system can be divided into different kinds.This dissertation studies the Motor Imaginary System in Spontaneous BCI,in which the subjects repeatedly imagine the movement of their fingers while actually keeping them still.The EEG generated during this process is collected,analyzed,then classified to complete the interaction with external devices.At present,the research of BCI systems using motor-imagined EEG are mostly processed in three steps.First,manually select the conductive pole channels with obvious features(usually around C3 channel and C4 channel)and do some proper preprocessing.Second,find the suitable feature extraction algorithm to extract the EEG features on the selected channels.Third,apply different classification algorithms and find the best-performing classifier.The average classification accuracy of this mode on a single subject generally ranges from 72% to85%,but there are many problems such as the selection of channels is based on experience,the extraction of features is incomplete,different subjects have different features.To deal with these problems,this dissertation uses the deep learning algorithms which can automatically learn features from input data.Thus,the experiment does not need to manually select the channels,feature extraction and classification are also merged into one step,which simplifies the design of BCI system.The work of this dissertation is mainly divided into three parts.In the first part,we design the EEG acquisition experiment based on the mechanism and characteristics of the motor imaging.Time-domain analysis,frequency domain analysis,and time-frequency analysis are all used in this part.After removing the abnormal samples,we use wavelet reconstruction method to extract the specific frequency band of the EEG.The second part focuses on two deep learning algorithms: Convolutional Neural Network(CNN)and Long-term Short-term Memory Network(LSTM).This paper shows the reason for choosing these two algorithms,and design different structures of network.After experiments,it chooses the best structure and try to analyze why this structure has advantages over other structures.In the third part,this paper presents a method which can combine time domain,frequencydomain and space domain of EEG.This method transforms the EEG into a series of power spectrum pictures,and then use the CNN algorithm combined with the LSTM algorithm to extract more comprehensive features.This method is proved to have a strong robustness among different people.Besides,it can be a more general method to process EEG,which also provides a novel idea for the classification of EEG.