Research On Channel Selection And Training Data Quality Assessment For Motor Imagery BCI

Noninvasive Brain-Computer Interface(BCI)provides an alternative communication pattern for those people who are suffering from severe physical disabilities by recording and decoding bioelectrical signals of human and transforming them into control commands of external devices.These people can control external devices without depending on the normal muscles and nerves system.Electroencephalography(EEG)is most frequently used to measure brain signals for BCI applications due to its convenience and safety.However,EEG signals are inevitable contaminated by non-brain activity artifacts and environmental noises.In addition,the spatial resolution of EEG signals is relative low due to the volume conduction effect of the human brain.Therefore,the signal-to-noise ratio of the EEG signal is relatively low.Independent Component Analysis(ICA)is an effective spatial filtering method and widely used in BCI denoising and feature extraction.However,the ICA-based motor imagery BCI system(ICA-MIBCI)is very sensitive to the channels and training data quality,which will affect the classification performance of the motor imagery EEG signal(MIEEG).In this thesis,automatic channel optimization approach and training data quality assessment method were proposed for ICA-MIBCI.The details are as follows:(1)Since the performance of ICA filters is closely related to the data quality of the training samples,a training data selection method,called self-testing,was used to eliminate the influence of training data on optimal channel selection.ICA-MIBCI system and self-testing were constructed based on single trial.After that,we ranked all single trials according to their self-testing accuracies.Then,top 10 trials were selected and concatenated as training data in the subsequent channel selection procedure.(2)An automatic channel optimization method for ICA-MIBCI system was investigated in this thesis.Since the optimal channels are subject-dependent and may change due to different experimental environment.Inappropriate channel combination would lead to suboptimal performance.Therefore,we proposed a channel optimization method.Firstly,ICA-MIBCI and common spatial pattern-based MIBCI(CSP-MIBCI)were constructed based on commonly used channels.Given selected candidate channel combinations,ICA spatial filters were designed and applied into corresponding ICA-MIBCI system.The optimal channel combination was selected according to the self-testing results.The experimental results demonstrated the effectiveness of the proposed method.Average classification accuracies with optimal channels during self-testing,session-to-session transfer and subject-to-subject transfer are 93.8%,87.7%and 94%.Compared with fixed channels and CSP algorithm,the increments are 8.5%and 2.8%under self-testing,9.5%and 14.4%under session-to-session transfer,26.7%and 36.2%under subject-to-subject transfer,respectively.(3)An EEG signals quality assessment method based on accuracy-matrix was proposed in this thesis.Since the EEGs are weak signals under strong background noise which are usually contaminated with non-brain activity artifacts.The amplitudes of these artifacts are usually higher than that of EEG signals,which will lead to signal quality problems and seriously affect the classification of EEG and the control of external devices.Accurate evaluation of EEG signal quality is beneficial to select appropriate EEG signal to improve the performance of BCI system.Therefore,we proposed an EEG signals quality assessment method.Firstly,the accuracy-matrix was designed,then several signal quality parameters and corresponding threshold were selected and set based on accuracy-matrix.Threshold evaluation was used to classify the signal quality of 32 EEG datasets.The experimental results shows the effectiveness of the proposed signal quality assessment method,and the highest average classification accuracy reaches 91.6%.