The Study Of Sparse Measurement Based Feature Extraction And Classification In Motor Imagery Eeg

In 21 st century, one of the research hot spots is brain protection and related information mining. The research of brain function is the fundamental window to understand brain mechanisms. Electroencephalography(EEG) acts as a predominant method in the cerebral-function information mining, providing the millisecond temporal resolution for the related researches. Recently, the use of brain computer interface(BCI) based on EEG have obtained much interest ranging from medicine to entertainment. However, the signals collected by amplifiers from brain usually have higher dimension with less sample number, which are non-stationary and contain many outliers. As for BCI, these problems get sharper. In order to alleviate these defects, one necessary step for preprocessing is dimension reduction with feature extraction. Currently, some standard feature extraction and pattern recognition methods had been proposed, such as Common spatial pattern, Principal component analysis, linear discriminant analysis, but most of them are L2 norm based methods, which indicate that they are prone to outliers with strong amplitude. These interferences will restrict the further development of BCI. In the perspective of statistics, L1 norm is more robustness to outliers than L2 norm. Therefore, it is very essential and of significance to develop relative approaches for extracting efficient brain information with stabilized EEG systems.Based on the inherent characteristics of outliers introduced by EEG acquisition, this dissertation ameliorates the L2 norm based feature extraction and classification algorithms to lower the effects of outliers and improve the robustness of BCI system. The contribution of this dissertation is as follows:1. For feature extraction in BCI system, the conventional least square based methods will exaggerate the influence of outliers so as to distort the features. This dissertation used L1 norm based singular value decomposition to estimate the common spatial filters. The application to both the simulated and actual MI datases demonstrated that the proposed approach can robustly extract the related MI features for BCI system.2. For pattern recognition in BCI system, the outliers with strong amplitude will also influence the the classifiers if they are constructed in L2 norm space. This dissertation build the new Rayleigh quotient object function in Lp(p<=1) norm space so as to get a more robust classifier to outliers. The developed Lp norm discriminate analysis is applied to various datasets including the simulation, actual MI dataset, Gene dataset, face recognition dataset and UCI datasets, and the results consistently proved that the developed method is robust to outliers, resulting in relatively higher recognition accuracy than other conventional LDAs.