The Research Of Feature Extraction Algorithms Based On EEG Signals In Motor Imagery

Brain-computer interface (BCI) is a special human-computer interface technology that doesn’t depend on traditional information channels such as peripheral nerves and muscles, instead, it realizes direct communication and control between brain and external devices. As information carrier of neural activity, scalp EEG signals reflect changes of mental state in real time and can be easily detected, so they have been widely applied in non-invasive BCI system. Due to the effect of volume conduction on brain, the spatial resolution of scalp EEG signals is low. Meanwhile, the neural activity artifacts (such as EOG, EMG, ECG, etc.) and environmental noise also greatly reduce the SNR of useful information. Therefore, how to obtain real neural activity components related to thinking activity from the multichannel scalp EEG is a key technical link in the study of BCI system.Focusing on the realization of motor imagery BCI system, we have carried out research to EEG signal processing and new feature extraction methods, the main work is as follows:Firstly we designed experimental paradigms in motor imagery and collected rich EEG data for subsequent research.Then we studied four envelope detection algorithms in view of the detecting of task-related envelope of EEG rhythms, which included: the nonlinear energy operator (NEO), Hilbert transform(HT) and two kinds of sliding window Independent component analysis(ICA) algorithms. Based on the dataset of BCI2003competition, we analyzed and compared the application effect of four algorithms. After studying the influence of artifacts on the detection accuracy, we proposed corresponding improvement method.Thirdly a new spatial filtering method based on time, frequency and spatial domain was studied. As two important spatial filtering algorithms, ICA and Common spatial pattern (CSP) carried on filter processing on EEG signals with extracted spatial filters to obtain implicit signal sources related to neural activities. Because the design principles of spatial filters were different, the physical significance of implicit sources also differed greatly. Based on the measured EEG data of motor imagery, this thesis analyzed and compared the performance of two algorithms. On this basis, this thesis proposed a new EEG feature extraction method that combined ICA with CSP. The experimental results verified the effectiveness of proposed method.