Research On Classification Algorithm Of Motor Imagery EEG Based On Deep Learning

Motor imagery brain-computer interface system has been a hot research topic in the field of brain-computer interface,which can translate neural activities into control commands of assistive devices to help disabled patients to complete some human commands.However,there are still many difficulties in the work of accurately understanding brain dynamics and classifying different motor imagery EEG signals,especially in the classification problem of multiple actions,how to effectively extract features and improve signal classification accuracy is still a problem that needs to be solved nowadays.In this paper,we mainly use deep learning methods to classify and analyze the motor imagery EEG signals of four kinds of actions,and our work includes the following aspects:(1)In this paper,we first classify four motor imagery EEG signals from the Data 2a dataset(hereafter referred to as the Data 2a dataset)in the BCI Competition using the more classical co-space pattern feature extraction algorithm and support vector machine classifier,and the results are used as a comparative benchmark for deep learning methods to evaluate the deep learning network model effects.(2)A deep learning-based approach to classify motor imagery EEG signals.In this paper,we designed a long and short-term memory neural network model and designed to use three convolutional neural network models,Deep Net,EEGNet and MI-Net,to classify the four motor imagery EEG signals in the Data 2a dataset as well,and the classification accuracy obtained was improved compared with the co-space pattern algorithm.The average classification accuracy of nine subjects could reach 81.01%.From the classification results of these deep learning network models,it is demonstrated that the deep learning approach can spontaneously extract the features of motor imagery EEG signals and can achieve better classification results.(3)The proposed methods above all focus on feature extraction for raw EEG signals,and the input form is a combination of motor imagery EEG signals from multiple electrodes into a feature vector.This paper also investigates the classification of motor imagery EEG signals based on time domain,frequency domain and spatial domain information,using a dataset of imagery EEG signals from 14 subjects in the Physionet database for four types of imagery actions: imagined left hand,right hand,double fist and double foot.The deep learning network model used is a VGG-based improved network model.During the training and testing process,the classification results of single-frame and multi-frame input,and with and without spatial information input were compared,and the average classification accuracy of the four types of motor imagery EEG signals in the case of multi-frame and with spatial information input was about 85%.The average classification accuracy of the four motor imagery EEG signals with multiple frames and spatial information input is about 85%,and the classification accuracy is the highest,which indicates that both time domain features and spatial information features play a positive role in the improvement of the classification accuracy.