| The purpose of brain-computer interface(BCI)is to establish a direct connection between the human brain and external communication devices.It can convert different modes of brain activity into computer commands by analyzing and identifying electroencephalogram(EEG)signals recorded in specific psychological tasks.The effective combination of BCI technology and other science and technology can promote the development of medicine,military,education and other fields,so it has been widely studied.Among them,the BCI system based on motor imagination(motor imagery,MI)has been widely used in medicine.The system mainly uses motor imagination EEG as signal input and carries out signal processing through computer,thus realizing the communication between the brain and the outside world.It makes it possible for the severely disabled to regain the ability to control the external environment.In recent years,a large number of literatures have reported the classification algorithms of MI signals used to identify left and right hands,tongues and feet.Most of the EEG signals used come from open source data sets.However,only a few literatures have given the design of leg-related BCI system,and there is no public data for research.Therefore,the purpose of this study is to collect relevant data independently and try to prove the feasibility of using classical algorithms and deep learning networks to classify signals.The methods used in this paper are as follows: MI EEG signals of left hand,right hand,tongue,foot,left ankle,right ankle,left knee and right knee of 10 volunteers were collected,and EMG,EOG and other artifacts were eliminated by independent component analysis.The correctness of self-collected data is verified by combining transfer learning theory and a highprecision classifier.Because the features extracted from a single domain can only provide limited information,while the features from different fields may contain more useful information for classification,this paper chooses to extract time,frequency and spatial features to form a multi-domain feature matrix and send them to support vector machines(Support vector machine,SVM)for classification,in order to avoid "dimension explosion" and feature redundancy affecting the classification results.In the process of training,particle swarm optimization algorithm(Particle Swarm Optimization,PSO)is used to extract the most relevant classification features and the best channel.The verification results of collected data show that except for one person,most of the volunteers can complete the excellent MI task,and the EEG signals of 9 qualified volunteers are used for further research.Compared with the classification results of other existing algorithms,the average classification accuracy of PSO-SVM for 9 volunteers is 76.33%,and the recognition accuracy is higher than other methods.Among the 9 volunteers,the average classification accuracy of Bayesian linear discriminant analysis(Bayesian linear discriminant analysis,BLDA)was 49.15%.The classification algorithm based on deep learning also did not show good advantages.The classification accuracy of convolution neural network(Convolutional Neural Network,CNN)and CNN-long-term and short-term memory artificial neural network(Long Short-Term Memory,LSTM)in 9 volunteers was 33.47% and 33.38% respectively,close to the random probability value.By comparing the algorithms with or without feature selection,the number of features used after optimization is reduced by 24.71%,and the accuracy is improved by 25%,which proves that the signal processing method in this paper improves the classification performance of the BCI system.In this study,the accurate classification of lower limb MI MI signals expands the effective action range of MI experiment,and shows that lower limb EEG signals can be correctly recognized.In addition,the importance of feature optimization to related signal processing is explained.This study has a certain reference value for the development of BCI system based on leg MI signal. |
