Study On Featraction And Classification Of EEG Signals In Brain Computer Interface

With the increasing aged and disabled people, more and more people demand a brand new way of communication instead of traditional sensory modalities, thus brain-computer idea emerged as the times require and come to be a hot topic worldwide. Brain-computer interface is a method that doesn't rely on traditional peripheral nerve and muscles, but the direct orders from brain to control the equipment.Nowadays, BCI is still in the lab research, and the brain signal is hard to extract precisely because of its brain signal's instability and diversity. Base on the physiological features of brain signal, this paper tries to find a more perfect signal processing system through many experiments of various brain signals. In details, considering the key processes of signal processing, I pay more attention on the study of feature extraction and pattern recognition.First, since brain signals are non-stationary, we put forward the idea of combining wavelet transform coefficients, coefficient averages and wavelet entropy as the feature vectors, which joins the energy information of brain signal. Through this method, features extracted from the P300 signal could express the transient components of the signal to effectively increase the classification accuracy with the aim to control the machines through brain activity. Secondly in feature classification, we improved two self-training semi-supervised algorithms based on support vector machine (SVM) and K-means. In the experiment of BCI2003 datasets, compared with traditional BP neutral network, it was showed that the two proposed methods improve the classification accuracy with much lower dimension of the extracted features and shorter convergence process, which makes real-time BCI system possible.Based on the former studies, this paper also did some research on feature extraction and classification of the prevailed motor imagery signal.In the experiment of BCI Competition IV datasets, I realized the combination of Common Spatial Pattern method and BP neural network. Compared with the Champion results, my methods have further improved the classification accuracy.