Manifolds-Based Multivariate Time Series Signals Classification Algorithm

As the control center of whole body,human brain handles all kinds of human activities such as emotion,cognition,sense,and behavior.Magnetic multivariate time series signals generated by human brains and captured by special devices are called MEG(magnetoencephalography)signals.Investigation in MEG signals will reveal mysteries of brain,and provide a solid basis for treatment of brain diseases.Therefore,research of brain mechanisms,in this dissertation,is of important significance and practical value.This dissertation proposes a novel classification algorithm of MEG signals based on Riemannian manifolds,on the basis of researching relevant MEG signals' processing theories.As a matter of fact,there lies a huge disadvantage that MEG signals are high-dimensional.Therefore,after filtering operation in preprocessing,we first propose a dimension reducing algorithm which utilizes covariance matrix,and we could get result signals of much less dimensions.Then we adopt a pre-classification procedure based on SVM and Lasso methods,and get primary classification results.At last,we propose a new method combining Bhattacharyya distance and Riemann distance to adjust classification labels,and we get results of high accuracy,as contemporary methods' accuracy is not very good.The accuracy of our algorithm is 80.11% while the accuracy of the champion is 75.8%.As to algorithm complexity,on the same computer,our algorithm only needs less than an hour to run the codes while its counterpart consumes about 16 hours to complete.