Extension Of CSP Algorithm Under Multi-class Condition And Its Application To Brain-computer Interfaces

Feature Extraction is the most important segment in BCI system, which influences system's classification performance. Common spatial pattern is a spatial filtering algorithm, which is used to extract spatial feature of multi-channel EEG signals. Although the classification performance of CSP algorithm is very excellent, it is a binary algorithm which only is fit for feature extraction of two meantal tasks. In order to improve the information transfer rate of BCI system, it's necessary to extend CSP algorithm to multi-class paradigms.Simultaneous diagonalization CSP algorithm is an algorithm for extending binary CSP algorithm to multi-clss cases, which does approximate joint diagonalization of muitlple covariance matrices and uses principal component analysis and spatial subspace to extract the feature of EEG signals. In the paper, we mainly introduce three approximate joint diagonalization algorithms:Jacobi algorithm, ffdiag algorithm and UEDGI algorithm. Paper analyzes these three algorithms' performance in convergence rate, gray image of multiple covariance matices after processing by each algorithm, classification accuracy and runnig time and gets UEDGI algorhitm as the best choice with average classification accuracy 78.31 and average information transfer of 13.62 bits/min for five subjects. In addition, paper detailedly describes another two multi-class CSP algorithms:one versus one CSP algorithm and one versus the rest CSP algorhitm. In four-class cases, the performance of CSP algorithm based on UEDGI is superior to another two multi-class CSP algorithms both in classification result and runnig time, althogh the performance of CSP algorithm based on UEDGI is inferior to OVO CSP algorithm in three-class paradigms. Therefore, CSP algorithm based on UEDGI will be the first choice in multi-class paradigms with the increase of recognizable class number.