On The Single-Channel Few-Trial Extraction And Classification Method Of Event-Ralated EEG Signals

An internally or externally paced event results not only in the generation of an event-related potential (ERP) but also in a change in the ongoing electroencephalogram (EEG) in form of an event-related desynchronization (ERD) or event-related synchronization (ERS). The ERP on the one side and the ERD/ERS on the other side are different responses of neuronal structures in the brain. While the former is phase-locked, the latter is not phase-locked to the event. The ERD/ERS is highly frequency band-specific. Therefore, it is import to study the single-channel few-trial ERP extraction and motor imagery EEG signals feature extraction and classification problems. It will not only help to enrich the contents of the human brain cognitive science and neuroscience, but contribute to the development of artificial intelligence technology. Aiming at these targets, the contribution of this thesis includes the following work.(1) A virtual signal channel independent component analysis (ICA) method was proposed to realize ERP few-trial extraction from space domain viewpoint. After analyzing the applicability of ICA for the single channel ERP few-trial extraction problem, a virtual channel ICA model was presented to make it comply with the premise of the ICA method. With this new model, the record signals from single channel can be realized with only four-trials.(2) A subspace matrix filtering approach was advanced to estimate single channel ERP with few trials from time domain viewpoint. Considering the limitation of the traditional convolution (vector) filter in noise reduction, a matrix filter for few-trial ERP extraction was given. Based on the principle of the generalized subspace approach (GSA), the filter matrix could be expressed as the multiplication of projection matrix, coefficient weighted matrix and reconstruction matrix. Then the projection matrix was obtained with the observed noisy signals and the coefficient weighted matrix was calculated under the minimum mean square error (MMSE) criterion. The ERP signal was then obtained by averaging the signals estimated with the reconstruction matrix. The algorithm can estimate the ERP signal with only two trials observable noisy signals which greatly reduced the number of trials required.(3) A weighted-thresholding wavelet analysis was described to estimate ERP signals with single trial from transform domain viewpoint. The underlying principle was to filter the noisy signal and turn the EEG into white noise. Then, through wavelet transform, the coefficients of the noise were distributed in all scales and shifts while those of the expected ERP signal concentrated only in several scales. The expected signal was estimated by weighting the coefficients and inverse transformation. This algorithm is free from the dependence on the prior information of the EP signals, eliminating the choice of wavelet coefficients by experience which is in line with the application background. Moreover, it can extract ERP signals with only one trial record.(4) For ERD/ERS in the motor imagery (MI) EEG. two adaptive feature extraction methods were given to extract the EEG features that reflect the different mind states. A sequential likelihood ratio test (SPRT) was combined to classify the features. The proposed technique not only improved the classification accuracy and transformation rate, but also and balanced the tradeoff between speed and accuracy. Experimental results suggested the possibility of improving the quality of BCI.