Tensor Method Research Of Underdetermined Blind Source Separation Based On Higher-Order Statistics For Ocular Artifacts Removal Of Electroencephalogram

Electroencephalogram (EEG) can reflect a person’s psychology and physiological functions of the brain. It is widely used in clinical diagnosis of disease and psychological analysis. EEG is easily contaminatied by other physiological signals. Especially, uncontrollability of eye signals seriously affect EEG analysis. It is meaning full to effectively remove the eye signal for disease diagnosis and signal analysis. A series of studies focus on ocular artifacts removal based on blind source separation in this paper. It includes the following three aspects:(1) The artifact identification method is proposed based on parametric model. Firstly, according to the non-Gauss of ocular artifact, heteroscedastic mixture transition distribution model is built to identify ocular artifacts and normal EEG. The expectation conditional maximization method is used to estimate parameters of heteroscedastic mixture transition distribution model. Parameters of model are used to identify ocular artifacts as the feature of ocular artifacts. Furthe, for epilepsy patients and ocular artifact identification problem, because both of them are non-Gauss, the heteroscedastic mixture transition distribution model can not resolved that problem. The extremem learning machine is proposed to bulid pahse model for instantaneous phase of signal. The instantaneous phases are used as input samples of model for training parameters. Then, the ouput weights are used as feature of EEG to achieve classification between epilepsy EEG and ocular artifacts. Hilbert-huang transform method is used to obtain the instantaneous phase of signal.(2) The higher-order statistics CP tensor approach is proposed for removing ocular artifact base on underdertermined blind source separation (UBSS). The uniqueness decomposition of higher-order statistics CP tensor can guarantee that the estimation of the sources in the process of UBSS of ocular artifact removal is unique. When the observe signals have a strong correlation, the sources are difficult to be estimated. The observed signal can be dealt with principal component analysis method to reduce the second order correlation, so that it can be focused on the higher-order statistical analysis. The higher-order statistics CP tensor is constructed by the principal component matrix of the observed signal to improve estimation performance of the sources. Further, considering mixing matrix are nonnegative, an hierarchical alternating least squares approach with regularization is proposed to decompose tensor model in order to ensure that the decomposition of model is nonnegative. And then the nonnegative mixing matrix of UBSS is obtained.(3) The higher-order statistics Tucker tensor model is proposed for removing ocular artifact basd on UBSS. Because the ocular artifact is deep hidden in EEG, it is difficult to separate form the EEG. In order to solve that problem, the core tensor of Tucker model can be used to deep dig latent variable. The higher-order statistics Tucker tensor can be used to underdetermined blind source separation of ocular artifact removal. Because the higher-order statistics Tucker tensor is hard to decompose, an hierarchical alternating least squares approach is used to decompose tensor model for enhancing decomposition rate. On this basis, the Fourier transform is used to construct time-frequency higher-order statistics Tucker tensor in order to improving the estimation performace of the nonstationary source. The conjugate gradient method with orthogonalization is used to estimate the sources for improving estimation accuracy.