EEG Spatial-temperal Feature Extraction And Classification For Imitating Reading Based BCI

With the rapid development of cognitive brain science, neuroscience, electronictechnology, computer science and signal processing, the brain-computer interfaces haveprovided a new way to communicate with the outside world and drew a considerableattention from the society.Typically, a brain-computer interface system consists of a signal acquisition part, asignal pre-processing part, a feature extraction part, then a pattern recognition part and aexternal control part. And, the feature extraction and pattern recognition of EEG signals arethe cores of the system. The ability to extract the signal feature impact on the design of theclassifier and its performance tremendously, so this part is a major issue of the researchers.According to the commonalities and differences of the subjects’ individual EEG signals,and the spatial information, this paper presents to use common spatial pattern algorithm tofind an optimal spatial filter automatically. Considering the instability of EEG signals, therequirements of BCI system and the adaptive capacity of the online BCI system, this paperpresents an online method to update the spatial filtering model which is based on CSPalgorithm. Finally, we use Matlab tool to complete the procedures of EEG signalspre-processing, space-time feature extraction and pattern recognition. Therefore, this papercontains the following parts:1) We acquire the EEG signals which is based on the "imitating reading” event relatedpotentials evoked, establish the EEG data sets, and complete the signal pre-processing.2) According to the feature of EEG signal, we select the signal associated with the task,using different target expression of CSP algorithm to extract the spatial feature, and find theoptimal number of the feature vector. Took the classification accuracy and execution rate intoconsideration, we found that the algorithm based on Rayleigh coefficients has moreadvantage.3) We use the CSP algorithm which based on Rayleigh coefficients,Sherman-Morrison-Woodbury matrix inverse formula and different sub-component extractionscheme to update the spatial filtering model, then we compare the updated projectioncomponents and the component which are not updated to find out whether this update isefficient or not.4) After the feature extraction, we use the SVMclassification to classify the EEGsignal. The SVM classification’s kernel function is polynomial kernel function. In order to getthe best classification accuracy, we use the cross-validation function of the LibSVM, and withthe method of Leave-20%-Out for optimizing the parameters of the kernel function and the classification.In brief, we use the characteristics of the EEG signal in time and space to extract thefeature and update the spatial filtering model; finally, we achieved a acceptable result, and aimprovement of the robustness of our BCI system.