Study On Classification Methods Of Multi-class Mental Tasks Based On Support Vector Machine

Brain Computer Interface(BCI) can offer a communicating and controlling channel between the brain and the external world. Study of the modern brain-science shows that the electroencephalogram(EEG) contains very rich physiological and pathological information, and can reflect brain's function state. In order to help patients with serious behavior obstacles to facilitate exchanges with the outside world, people begin to research to recognize several simple mental tasks and form the comparatively complicated control command which can accomplish controlling to auxiliary equipment such as wheel chair, artificial limb.However, due to the complexity and non-stationarity of EEG, It is difficult to obtain sufficient sample data for training and classifying. Statistical learning theory is the theory of estimating and projecting which suits to small samples, it can solve many problems such as model select, overfitting, nonlinear and dimension disaster[2]. As the realization method of statistical learning theory, SVM not only has simple structure, but it's many properties especially extended ability is improved. So it's a feasible choice to apply it in the field of mental tasks classification.Although the advantages of SVM in theory is very prominent, but there are two crucial problems in it's application. First, for a given sample, the performance of SVM mainly influenced by the kernel function and parameters, so how to select them is the key of a good SVM classifier. Then, SVM is proposed originally specifically for the two-category classification, for multi-category problem, we need to construct multi-category SVMs. Therefore, revolved the mental tasks classification based on BCI, we studied the key problems of feature selection, the selection of kernel function and parameters, designing of multi-category SVMs in the papaer.Firstly, the paper introduced the background, research situation and scientific significance of mental tasks classification, and elementary theory of statistical study theory and SVM. In order to discuss how to choose kernel parameters, we first studied it's change rule, then proposed adopts the way of "rough adjust then intensify adjust" to seek the optimal parameters fastly, it is more suitable for the practical application.When reducing the dimensions of EEG feature vector, for feature extraction is through some mapping to compress data, which will destroy thr original physical significance, and is disadvantageous to the classification. Therefore this article has carried on the feature selection to the best electrode combination and the EEG feature.In order to recognize three mental tasks, a new multi-category SVMs combined~2with decision tree and SVM is proposed in this paper. The decision direction is determined by the separability measure based on class distribution, which can reduce the influence of "error accumulation" coming form decision tree. This method not only simplifies the structure of classifier and expedite the speed of training and identification, overcome the problem of rejecting classification, but parameters of every sub-SVM can be selected respectively, whiche will better to the improve of classification accuracy. The result of analyzing a data set of BCI competition 2005 provided by IDIAP Research Institute shows that, the classification accuracy is higher than traditional multi-category SVM. The method presented in this paper is verified to be effective and it may provide a new approach to recognize multi-class mental tasks.