Researchon Motor Imagery Brain Computer Interface System Algorithm Based On Convolutional Neural Networks

Brain Computer Interface is a brand new system connecting Brain with Computer,which has the potential value of changing the way of human-computer interaction.Brain-computer interfaces have grown exponentially in recent years,people were able to send two bits of data per minute to a computer via a BCI system in 2002.Four years later it was 40 bits,or five letters per minute.According to an article published by Nature in 2019,the transmission speed through BCI system has reached 25 letters per minute.This idea was mentioned in the2002 national science foundation and U.S.department of commerce report ‘Technology summary of strengthening human beings',Brain Computer Interface can supplement speech and even replace conversation when it is needed to maximize the efficiency of consciousness communication".When the brain-computer interface system allows the computer to communicate ideas faster than the vocal cords,and the speed of communication between people is faster than the human brain itself,artificial intelligence must be used to assist the human brain in communication.In the 13 th five-year plan,China proposed to list "brain science and brain-like research" as one of the country's major scientific and technological innovation and engineering projects.Nowadays,non-invasive BCI system which is able to collect scalp electroencephalogram(EEG)signals has been widely applied.EEG signals can be collected only by contacting with the scalp instead of implanting into the brain,which is convenient and non-invasive.A head-mounted device is used to collect electroencephalogram,which has a certain information transmission rate under the condition of human safety and relatively low cost,and is close to the ideal simple and portable BCI.It is therefore commonly used in research about the brain computer interfaces.Motor imagery is the most common communication method for non-invasive BCI system when collecting EEG signals.The EEG signals obtained by stimulating the eeg rhythm changes of the cerebral motor cortex through the motor imagination task serve as the theoretical basis for the BCI system research based on motor imagination.Because of the low signal-to-noise ratio and high feature dimension ofElectroencephalogramsignals,it is difficult to improve the classification rate of BCI.This research aims to systematically develop the motor imagery Brain Computer Interface system algorithm based on Convolutional Neural Network.An EEG signal innovative processing frame was presented in this paper firstly,it's consisted of 11-layer CNN builted by PyTorch for feature extraction,The feasibility of the proposed algorithm is proved by using a set of data set in the 4th international competition database of brain-computer interface.Finally,the experimental classification results can reach 95%,which proves the innovation feasibility of the algorithm.This research aims to systematically develop the Brain Computer Interface(BCI)system algorithm based on Convolutional Neural Network(CNN).The research based on the motor imagery BCI system was carried out in terms of the pretreatment,feature extraction,classification and identification methods of electroencephalography(EEG)signals.It is innovative that the 11-layer CNN model was constructed in this research by using PyTorch for feature extraction.In the signal pretreatment part,the Least Mean Square(LMS)adaptive filtering algorithm was employed for the filtering.Principal Components Analysis(PCA)and Fast Independent Component Analysis(FastICA)were employed to reduce the dimensionality,and the better option was selected.The classification and identification was conducted by using Gradient Boosting(GB)(95%)algorithm and Bayesian Linear Discriminant Analysis(BLDA)in order to achieve the further classification,and the better one was selected.At the same time,the use of BCI competition database to train 11 layer model to optimize the structure and parameters of CNN,and compare with the three other kinds of traditional classification method proved the advantage of CNN on brain electrical signal classification.Firstly,joint training of CNN was carried out by employing the international standard BCI competition database.The appropriate pretreatment method PCA and classification identification method GB were selected.Then,the EEG signals collected were applied to the PCA+CNN+GB joint training algorithm.The feasibility can be proved by classification accuracy rate of 95%.Finally,compared with CNN(86%),CNN+PCA(90%),and CNN+GB(91%),the performance of joint training turned out to be the best when the pretreatment PCA and classification identification GB were added on the basis of CNN.The experimental results show that the best classification result(95%)can be obtained when theLMS adaptive filtering and PCA dimensionality reduction were initially used for the pretreatment of the EEG signals,and then the 11-layer CNN was used to extract the EEG features.Finally,the EEG signals were reclassified by the GB algorithm identification method.The experimental results show that the signal is preprocessed by LMS adaptive filtering and PCA,and then the eeg signal recognition method combining CNN and GB algorithm is built by Pytorch to get the best classification results.Moreover,the PyTorch framework is simple in structure and fast in operation,which is very suitable for the training of CNN model,and proves the innovation of this paper and the usefulness of the algorithm used for EEG signal recognition.