Study On Space Mapping Based KCSP Algorithm And Its Application To Brain-computer Interfaces

Brain-computer interface (BCI) is a communication system which does notdepend on normal output channel consisting of peripheral nerves and muscles, and itis used to transmit information and command to the outside world. This system hasgreat significance for paralyzed patients, because it can be used to control thewheelchair or other external devices, and then realize their intentions.Feature extraction is the most important part in BCI system, which influencessystem’s classification performance. Common spatial pattern (CSP) as a commonlyused spatial filtering algorithm is used to extract spatial feature of multi-channelelectroencephalogram (EEG) signals. Because the CSP algorithm in brain-computerinterface assumes that the relation between the observed EEG signal and the brainsource signals are strictly linear and very limited nonlinear kernel functions exist,both linear CSP and nonlinear kernel CSP cannot characterize the brain patternaccurately.In this dissertation, a new hybrid kernel CSP algorithm was proposed in order toimprove the performance of pattern classification, which is based on the combinationof linear and nonlinear kernel functions. K-means clustering and Nystromapproximation were used to relieve the computational complexity and therequirement for memory in eigenvalue decomposition of kernel matrix. The bandpassfiltered EEG data were clustered for dimensionality reduction and then Nystromapproximation was performed to ensure the validity of low rank decomposition ofkernel matrix in a high dimensional feature space. The performance of the algorithmwas tested on a four-class dataset and compared with that of linear and kernel CSP. Inorder to improve the operation efficiency, use a pairwise classification strategy fortesting. The experimental results show that in terms of classification accuracy theproposed algorithm has more superior performance.