Research On A Motor Imagination Brain-Computer Interface Algorithm Based On Granger Causality Analysis

Brain Computer Interface(BCI)technology does not rely on the brain's conventional neuromuscular output pathways.The brain directly outputs information to communicate with the outside world,providing a new way of human-computer interaction.The brain-computer interface system usually consists of a signal acquisition part,a signal processing part,a control device part and a feedback link.The signal processing part includes preprocessing,feature extraction and pattern classification.The most critical part is the feature extraction of the signal.Different types of signals can extract the most significant features to effectively identify and classify different signals.Therefore,studying how to extract better features is of great significance to the application of brain-computer interface systems.This paper focuses on the algorithm of brain-computer interface for motor imagination,and compares and studies different preprocessing,feature extraction and pattern classification algorithms.In the feature extraction part,in addition to traditional feature extraction algorithms,it also uses nonlinear granger causality analysis method and direction transfer function method for multi-channel motor imaging EEG data to identify the correlation between different channels and construct a brain network structure.Incorporating these features into the feature set extracted by traditional algorithms further improves the accuracy of the classification of motor imagery EEG signals.The main content of the paper is as follows:(1)The algorithms commonly used in motor imagery EEG signal processing are studied and compared and analyzed through experiments.In the preprocessing algorithm,the Butterworth filter and wavelet packet decomposition and reconstruction algorithm are used to compare the processing effects of the collected motor imagery EEG signals.In the feature extraction algorithm,features are extracted from motor imagery EEG signals through Autoregressive Model(AR),Wavelet Transform(WT)and Common Spatial Pattern(CSP).In the pattern recognition algorithm,Linear Discriminant Analysis(LDA)and Support Vector Machine(SVM)are used to classify the extracted features,and the processing effects of different algorithms are compared and analyzed.(2)Multivariate nonlinear Granger causality(GC)analysis method and direction transfer function(DTF)method are used to identify the causal relationship and information flow between different channels in different regions of the brain.The calculation results of the non-linear Granger causality of different kernel functions and the direction transfer function of different signal frequencies are studied.And according to the correlation between the EEG signals of different channels,a motor-imaging causal brain network is constructed.Using brain network features for pattern recognition,a higher classification accuracy rate is obtained.The validity of the classification of left and right hand motor imagery EEG signals by constructing brain network features is verified.(3)Combining the single-channel features extracted from motor imagery EEG signals by common feature extraction algorithms and the network features obtained by constructing brain networks further improves the classification accuracy of left and right hand motor imagery EEG signals.The experimental results show that for the EEG data of6 subjects,the average correct rate of classification using AR model parameter features is82.9%.The AR model parameter features combined with the brain network features obtained by nonlinear granger causality analysis and direction transfer function,the average classification accuracy rate reached 87.6% and 87.5%,respectively.