EEG recognition through Time-varying Vector Autoregressive Model

Abstract

In this paper, the multi-task motor imagery EEG(electroencephalogram) signals are pretreated by principal component analysis and Fourier transform. By use of the methods of time series analysis and mathematic statistics, pretreated EEG signal series are separated into the deterministic part and stochastic part. Then, the stochastic part is analyzed by use of TVVAR (Time Varying Vector Auto-regressive) model to obtain the residuals. Therefore, the EEG signals are studied on the stochastic parts and the residuals of TVVAR model. EEG signals of 3 types of actions, 60 signals per action type, are sampled, from which a signal is in turn analyzed to be recognized. Experiments in this study indicate that the recognition rates of left, right, hold still are 93.33%, 98.33%, 96.67% respectively, and the average recognition rate is 96.11% through both the stochastic parts and the residuals of TVVAR model. It verifies the TVVAR model be useful to analyze autocovariance nonstationary vector process.