Abstract
A brain-computer interface (BCI) enables people suffering from affective neurological diseases to communicate with the external world. Common spatial pattern (CSP) is an effective algorithm for feature extraction in motor imagery based BCI systems. However, many studies have proved that the performance of CSP depends heavily on the frequency band of EEG signals used for the construction of covariance matrices. The use of different frequency bands to extract signal features may lead to different classification performances, which are determined by the discriminative and complementary information they contain. In this study, the broad frequency band (8-30 Hz) is divided into 10 sub-bands of band width 4 Hz and overlapping 2 Hz. Binary particle swarm optimization (BPSO) is used to find the best sub-band set to improve the performance of CSP and subsequent classification. Experimental results demonstrate that the proposed method achieved an average improvement of 6.91% in cross-validation accuracy when compared to broad band CSP.