The Extraction Of Single Channel Evoked Potentials

The detection and analysis of Evoked Potentials (EP) is one of the important ways of diagnosing neural disease and injury. EP signals obtained from clinical application are contaminated by the noise of Electroencephalogram (EEG), and the SNR of EP is from -10dB to -20dB. Besides, EP signals have time-varying and spatial characters, so the extraction of single channel EP with less or even one stimulation is very important.This thesis studies the methods based on radial basis function neural network (RBFNN) and model-based methods for estimating EP. The methods based on RBFNN use RBFNN to simulate the nonlinear character of cerebral neural system for estimating EP. The model-based methods use the priori knowledge about pure EP and model EP with the weight sum of several basic vectors, and then some kinds of estimation algorithms are used for estimating the weights of basic vectors. In this thesis, adaptive noise canceller method based on RBFNN, RBFNN filter method and subspace regulation method, least mean square method, recursive least square method and Kalman filter method based on linear observation model are used for estimating and tracing EP and the performance of these methods were analyzed.Recently it is accepted that Alpha stable distribution is better for modeling highly noise contaminated EP than Gaussian distribution. In this thesis, academic analysis and computer simulation prove that the performance of RBFNN filter method and model-based methods will degenerate for estimating EPs with Alpha stable distribution noise. This thesis proposes the improved RBFNN filter method and the improved model-based methods. The improved RBFNN filter method uses least p -norm algorithm for adjusting the weights of output layersof RBFNN, and the new algorithm can converge well for estimate EP with Alpha stable distribution noise. The improved model-based methods use nonlinear transformation to effectively restrain the pulse noise. Computer simulation results show that the proposed new methods are more robust than conventional methods in Alpha stable noise and Gaussian noise conditions.