Analysis Of EEG And Multi-model Brain Computer Interfaces

EEG is one of the body’s most important biological signals. On the one hand theanalysis of EEG increases understanding of the function of the brain and itsphysiological characteristics for practical use in clinical disease; on the other hand, EEGbased brain-computer interface(BCI) system can communicate and control externaldevices without peripheral nerves and muscles, which has broad application prospects.The paper includes the following content.An important application of EEG is in BCI systems, EEG feature extraction andclassification is the focus of BCI system. We do some researches about motor imagerybased EEG and then present three EEG feature extraction methods: W-CSP、PSO-CSPand W-RCSP. The first method is a combination of wavelet analysis and CSP algorithmand it is a time, frequency and spatial domain fusion method. Wavelet analysis withtime-frequency characteristics make it more suitable for processing non-stationary EEGsignals. The second method is based on the PSO algorithm, firstly the PSO optimizationis done for searching the best bands, then the best bands are used for signal processingapplications. The third method is the wavelet based regularized CSP, introducing tworegularization parameters, to enhance the stability of CSP. Comparative analysis showsthat the proposed method can be more obvious to distinguish characteristic vectors.For the classification of EEG features based on motor imagery, comparative studyof the typical linear and nonlinear classifiers-LDA and SVM-is completed. Then wecompared the classification results in the case of two classifiers using different featureextraction algorithms. Experimental results show that the W-RCSP+SVM achieves thehighest average classification accuracy, W-CSP+LDA gets the lowest averageclassification accuracy, for inefficient subjects of brain-computer interface,W-RCSP+SVM achieves the best classification accuracy. It indicates that W-RCSP ismore conducive to handling noisy signal, the accuracies of algorithms based on SVMare higher than LDA averagely, illustrating the advantages of SVM for nonlinearproblems, but when the feature vectors are clearly discriminative in the projectionspace, the classification efficiency of LDA is also high based on W-RCSP、PSO-CSP+LDA.In experimental observation, we have found that significant EEG characteristics ofeye movements. Through the analysis of EEG produced by eye movements, the real-time extraction of upward and downward eye movement characteristics achieveshigh recognition accuracy. We achieve the letters spelling, accumulating experience forthe follow-up study. Then, we propose an asynchronous BCI system combined ofimagination and eye movement, the results show that motor imagery and eyemovements based asynchronous BCI achieve fine results.