Research On Feature Extraction And Classification Algorithms In BCI Based On ERD/ERS

BCI (Brain-Computer Interface) is a new kind of Human-Machine Interface. It doesn't depend on human peripheral nerve and muscle system but gains the external communication information from the brain directly. The major purpose of developing and researching BCI is to design a useful system based on EEG to help those seriously disability patients and communication disorders patients to recover.Studying BCI is a process of continuous cognition of EEG.The main purpose of this study is to search a practical BCI system, find the methods fit for acquisition of signal and signal processing and improve the accuracy rate of classification. So a stable, reliable and speediness on-line BCI system will be built to assist paralyzed patients to recover physical movement or communication with the outside.The data used in the experiment is collected by 5 healthy members through the EEG amplifier. This thesis researches the off-line BCI in feature extraction and classification algorithm aiming at small samples, analyses the mechanism of EEG produce and discusses the general method of brain electrical signal processing. The research mainly adopts three feature extraction methods, the wavelet coefficient, relative wavelet energy and the AAR (Adaptive Autoregressive model) coefficients based RLS. The wavelet coefficients can well reflect signal's characteristic. RWE can reflect the energy in different frequency band and different time. Compared with AR (Autoregressive) coefficients, AAR coefficients change with time, so they are used to express the characteristics of non-stationary signal, and this method has more advantages than windowing AR model. The experiments show that the three methods of feature extraction are fit for signal analysis.There are a number of species of pattern recognition algorithms. This thesis chooses Bayesian and SVM classifier based taking into account of number of samples and dimension of feature. Cause of the present experimental conditions and preprocessing, the last results are not satisfying, but still prove that the methods mentioned in the paper are available. This research for on-line BCI system experiment research accumulates experience and lays the foundation.