Research On Tensor Decomposition-Based Channel Selection For Motor Imagery-Based Brain-Computer Interface

Brain-computer interface(BCI)is a technique that allows for communication without using conventional neuromuscular mechanisms.It has been used in rehabilitation therapy,sports assistance and other fields because of its important application value.However,in the process of practical application,BCI is easy to be interfered by redundant signals,and a large number of electrode channels in a BCI affects not only its classification performance,but also its convenience.Despite many studies on channel selection in motor imagery(MI)-based BCI systems,they consist in matrix analysis of EEG signals,which inevitably loses the interactive information among multiple domains such as space,time and frequency.Tensor analysis can solve the above problems.In this paper,as for specific subjects,a tensor decomposition-based channel selection(TCS)method is proposed for MI BCIs.Firstly,the continuous EEG signal of a given subject was preprocessed.Secondly,a three-way tensor is obtained by wavelet transform of a single-trial EEG signal,and decomposed into three factor matrices of channel,frequency and sample by regularized canonical polyadic decomposition(CPD),of which pearson correlation analysis is done on the channel factor matrix to determine the importance of each channel for the motor imagery task,and the importance of each channel is ranked,according to the channel ranking table,different numbers of channels are selected from the beginning to extract training and test data for each channel subset.Based on the training data of the selected channels,feature extraction is performed using regularized common spatial pattern(RCSP).Finally,we use the training features to train a classification model by a support vector machine(SVM),then the model is used to classify the test features.The classification accuracy of each channel subset is obtained,and the one that yields the highest classification accuracy is selected as the optimal channels.The proposed TCS-RCSP algorithm is evaluated on three BCI data sets and compared with all channels(AC-RCSP)and the RCSP with selected channels by correlation-based channel selection method(CCS-RCSP).The results indicate that TCS-RCSP achieves significantly better overall accuracy than AC-RCSP(94.4% VS86.3%)with ρ﹤0.01 and CCS-RCSP(94.4% VS 90.2%)with ρ﹤0.05,proving the efficacy of the proposed algorithm for classifying MI tasks.