Signal Feature Extraction Research Based On Kernel Method In Brain Computer Interface

A brain-computer interface (BCI) is a communication system in which messages orcommands that an individual sends to the external world do not pass through the brain'snormal output pathways of peripheral nerves and muscles. It will improve people'sability to communicate informations or control external device, especially for those withsevere neuromuscular disorders. Feature extraction is one of the most important steps inElectroencephalography/Magnetoencephalography (EEG/MEG) signal classification.Always, it is a hot question in current research. In order to solve the problem of featureextraction in signal classification, we use kernel method based on original CommonSpatial Patterns/Common Spatial Subspace Decomposition (CSP/CSSD) to extend threedifferent methods.Firstly, the kernel CSP method is extended from binary-class case to multi-class case.The CSP method has been proved to be very powerful in multi-channel binary-classfeature extraction of EEG. It is a presumption of strictly linear pattern. We adopt thestrategy of "One Versus the Rest" to extend the original CSP from binary-class case tomulti-class case, using kernel method. BCI competition Ⅲ dataset Ⅲ-3a has been used inthe experiment simulation. Comparing with other methods, this approach can extractrelavant classify feature from multi-class of single trial EEG.Secondly, the generalized kernel CSP feature extraction approach which bases ongeneralized singular value decomposition and kernel approach is extended in multi-classfeature extraction of MEG. In this paper, BCI competition Ⅳ dataset3has been used inthe experiment simulation. The average correct rate exceeds the second's in thecompetition.Thirdly, kernel CSSD approach is proposed in this paper. This approach is based onoriginal CSSD method appling in binary-class. It is extened to kernel space in this paper.BCI competition Ⅱ dataset Ⅳ has been used in the experiment simulation. The correctrate of test dataset is84%, which is equal with the winner's in the competition.