The Application Of Global Optimization Algorithm Based On Nonlinear Dimensionality Reduction In EEG Problem

The research of EEG based on potential of scalp is one of noninvasive research methods on brain science. It causes the concern of many scholars. EEG signal is non-stationary and non-linear random signal. In many cases, EEG data processed by computer usually has high dimension and a large quantity of data. If such data is processed by traditional linear or local algorithms directly, the problem will become much complex. Take into account of the characteristics of EEG, Nonlinear dimensionality reduction based on manifold learning to preprocess the EEG is introduced. From the perspective of the global and nonlinear, low-dimensional data that can reveal the intrinsic structure of high-dimensional EEG data can be get, then the EEG data is processed by global optimization algorithm.The current commonly used linear dimensionality reduction technology can not reveal the actual nonlinear structure of data sets. Nonlinear dimensionality reduction technology, such as isometric mapping, locally linear embedding, et al, is superiors to the traditional linear dimension reduction technique. Nonlinear dimensionality reduction algorithm, the global optimization algorithm and EEG problem have been researched and the application of nonlinear dimensionality reduction algorithm in EEG is researched. Three aspects of the application in EEG problems were researched: LLE and BP Neural Network is applied in the recognition of epilepsy spikes and background EEG and the method of automatically detecting epilepsy spikes based on nonlinear dimensionality reduction is proposed; The classification of awareness mission of moving the cursor upward or downward; The research on dipole localization used of ISOMAP and Genetic Algorithm. The correct recognition rate of epilepsy spikes that reached 97 per cent is showed by simulation results. The correct recognition rate of awareness mission of moving the cursor upward or downward reaches 92 per cent. Dipole localization from the EEG data processed by ISOMAP is calculated by genetic algorithm, When the evolutional generations are 500, the average fitness is 4.4808×10^-3. The convergence rate of genetic algorithms is raised.