A Study On Signal Processing Algorithms For Visually Evoked P300 Potential In EEG

Brain-computer interface (BCI) is a direct information communication and control channel established between human and computer or other electronics devices. It has a wide application in the medical, rehabilitation, entertainment and military. P300 component embedded in visually evoked EEG signals is a widely used component for the BCI interface system. Due to the complexity and nonstationarity of EEG, it is difficult for traditional averaging methods to extract the P300 signal feature information accurately, causing more superposition on P300 EEG signal processing and low classification accuracy. Motivated by these problems, this dissertation focuses on the P300 brain signal processing technology. The main study achievements include:Firstly, we use traditional algorithms for P300 EEG data processing, including low-pass Butterworth filter to smooth EEG data and optimal weighted average algorithm to remove the random noise. After choosing standard P300 waves as templates, template matching is then used for signal classification, yielding a character prediction accuracy of only 64.52%. In order to improve such low prediction accuracy, we improved the original template matching processing methods, resulting in a higher accuracy of 77.42%.In traditional template matching approach, the data of only 1 of 64 channel is used, hence resulting in low correct rate of classification. To take full advantage of the data from other channels, this dissertation proposes methods combining wavelet transform and support vector machine (SVM) for feature extraction and classification. We use the optimal weighted average to increase the signal to noise ratio when extracting P300 component. We also apply wavelet decomposition to data form 10 channels in order to remove noise and compress data. By doing so we can effectively extract features of P300 potentials. We then apply the kernel SVM (RBF kernel) to take advantage of the ability of SVM of dealing with small samples. Cross-validation is adopted to determine the best parameters of the training model. Using LIBSVM and MATLAB tools, we achieve a character prediction accuracy of 83.87%, which is significantly better than the traditional template matching method.In this dissertation, we combine wavelet transform with support vector machines for the feature extraction and classification of P300 EEG. Comparing to the traditional template matching algorithm, our method improves the accuracy of classification and lays a solid foundation for real time BCI systems based on P300 component extracted from visually evoked EEG signals.