Optimization on CSP-based Joint Feature Extraction Algorithm

Common Spatial Pattern method is restricted to the abundant input channels and few frequency information. This paper focuses on CSP algorithm improvement from the following three aspects: improving CSP filter, constructing CSP-based combined features, and optimizing pattern recognition process. First of all, component selection algorithm based on S-transform is proposed to build a new CSP-filter. Then, CSPS respectively combined with EMD, EEMD, Bispectrurm. Specially, constructions and comparisons in CSP-based joint features included these three ones: EMD-CSP, EEMD-CSP, Bispectrum-CSP. At last, with these joint features, SVM classifier is improved to determine the optimal range of penalty factor and the most stable kernel function. On the other hand, further classification recognition and comparison are carried out between SVM and LDA. In addition, a thinking-task experiment about motor imagery was designed to acquire laboratory data. And this data, combined with BCI competition dataset, were used to analyze the efficiency of proposed methods. The experimental results show that: in term of the classification accuracy and pattern recognition time, when using the LDA classifier, the Bispectrum-CSPS feature is the best optimization among the other ones with the highest accuracy rate and the shortest classification time.