Dimension Reduction Of EEG Signals Based On Feature Selection

Decoding EEG signals is a very difficult task,and one way to solve this difficult problem is to reduce the dimensions of the EEG signal.The multiclass and non-linear nature of EEG signals makes the process of dimension reduction difficult.In this thesis,two solutions are proposed to achieve dimension reduction for EEG signals.The first solution uses kernelized multiclassification support vector machines to extract multiclass nonlinear features to achieve dimensionality reduction of EEG signals;the second solution uses the property that the covariance matrix of EEG signals can be mapped to data points on Riemann manifolds to achieve dimension reduction of EEG signals on Riemann manifolds.The main studies are:(1)The class-wise optimisation kernelized multiclass support vector machine is proposed in the framework of support vector machines for EEG signals that are multicategory and non-linear.The algorithm selects features that work for discriminating all categories by combining a kernelized multiclass support vector machine with a modified recursive feature elimination algorithm.The recursive feature elimination algorithm has been improved to make its scoring criteria applicable to multiclass support vector machines,and also incorporates batch elimination and rescreening processes to improve the speed of feature selection.Experimental results on multiple datasets validate the effectiveness of feature selection and enable dimensionality reduction of EEG signals.The classification kappa is 15.5% higher than the original algorithms on EEG signals.(2)Based on the property that the covariance matrix of EEG signals is a symmetric positive definite matrix and can be mapped onto Riemannian manifolds,this thesis proposes EEG signal channel selection based on particle swarm optimisation.The algorithm selects the effective channels in the EEG signal by choosing the effective dimensions of the symmetric positive definite matrix,and realizes the dimension reduction of the EEG signal at the channel level,which makes the dimension reduction much more efficient.Experimental results on EEG signals validate the effectiveness of the proposed algorithm,and kappa is 11.8% higher than the original algorithms.The results are compared with brain work area knowledge to justify the selected channels.The thesis has 24 figures,13 tables,and 82 references.