Partitioned Common Spatial Pattern Method for single trial EEG Signal classification in Brain-Computer Interface System

Abstract

Common spatial pattern (CSP) method is highly successful in calculating spatial filters for motor imagery-based brain-computer interfaces (BCIs). However, conventional CSP algorithm is based on a single wide frequency band with a poor frequency selectivity which will lead to poor recognition accuracy. To solve this problem, a novel Partitioned CSP (PCSP) algorithm is proposed to find the most relevant spatial frequency distribution with motor imaginary, so that the algorithm has flexible frequency selectivity. Firstly, we partition the dataset into frequency components using a constant-bandwidth filters bank. Then, a features selection method based on the Bhattacharyya distance is adopted for PCSP features ranking, selection and evaluation. Subsequently, the PCSP features are used to obtain scores which reflect the classification capability and being used for EEG signal classification. The experimental results on 4 subjects showed that the PCSP method significantly outperforms the other two existing approaches based on conventional CSP and Common Spatio-Spectral Pattern (CSSP).