Magnetoencephalography Visual Target Classification Based On Semisupervised Algorithm

Electromagnetic signals can be generated from human activities,and they could be captured by devices,which are called MEG(magnetoencephalography)signals.It is of profound meaning to research MEG signals for curing kinds of human diseases and understanding advanced features of human activities.It is inevitable to be interfered by much noise in the process of the collection and transportation of MEG signals.Meanwhile the high dimensional features,which is the result of multi-channel and multi-time sampling,become the obstacles of the MEG research.As the result of much interference of noise,according to characteristics of MEG,it calls for well-designed filters to denoise.The high-dimensional data of MEG needs effective method for dimensionality reduction,and it is urgent to find an effective way to catagorise the processed data.In this thesis,we adopt combinational filters to denoise MEG signals on the basis of analyzing many filtering methods.We choose the way of covariance to reduce dimensions of MEG signals and extract their features,after comparing linear and nonlinear dimensionality reduction.We utilise SVM and Lasso models to catagorise features of data and use them to make a comparison,and then we propose combinational filters to improve models’ generalised features.On the basis of supervised combinational classifiers,unsupervised Bhattacharyya and Riemann distances are used to cluster them unsupervisedly,so the precision could be raised.The innovation of this thesis lies in the combination of supervised and unsupervised algorithms,which is the method of semi-supervision.The cross validation accuracy of the experiment is 80.11% and 99.2%,respectively,higher than 73.6% and 97.1%,which are the best records of the original competition.