脑电信号眼电伪迹去除的高阶统计张量欠定盲分离方法研究

发布时间：2018-06-07 15:05

  本文选题：脑电信号 + 眼电伪迹　； 参考：《大连理工大学》2016年博士论文

【摘要】：脑电信号能够反映人的心理状态和大脑的生理功能,被广泛地应用于心理分析以及临床疾病诊断等研究领域。脑电信号在采集过程中容易受到其它生理信号的干扰,其中具有不可控性的眼电信号的干扰(即眼电伪迹)最为严重。因此,有效去除脑电信号中的眼电伪迹对信号分析具有重要意义。本文在盲源信号分离方法的基础上,针对眼电伪迹去除问题展开研究。研究内容包括以下三个方面：(1)提出基于参数模型的眼电伪迹识别方法。首先,研究正常人脑电信号中眼电伪迹识别问题,针对眼电伪迹具有非高斯性,建立异方差转移混合分布模型,利用条件期望最大化算法估计该模型的参数,获得眼电伪迹特征,从而区分正常人脑电信号和眼电伪迹。其次,针对癫痫病人脑电信号和眼电伪迹识别问题,由于两者都具有非高斯特点,上述模型难以有效解决该问题,则根据癫痫脑电信号相位同步特点,利用极限学习机建立信号的瞬时相位模型,将模型的输出权值作为信号特征以用于区分癫痫病人脑电信号和眼电伪迹,同时采用希尔伯特-黄变换解决非平稳信号瞬时相位难以获取的问题。(2)提出基于高阶统计标准(Candecomp/Parafac, CP)张量欠定盲分离的眼电伪迹去除方法。高阶统计CP张量模型分解唯一性的特点可以保证在眼电伪迹去除欠定盲分离过程中源信号的估计具有唯一解。针对欠定盲分离中观察信号具有较强相关性导致源信号难以估计问题,引入主成分分析方法降低观察信号的二阶相关性,使信号集中于高阶统计分析中,利用观察信号的主成分阵构建出高阶统计CP张量模型并分解,从而提高欠定盲分离源信号估计性能。对于欠定盲分离中混合矩阵的非负性问题,采用一种基于正则化分层交替最小二乘方法分解高阶统计CP张量模型,保证模型分解过程为非负分解过程,进而求出欠定盲分离中的非负混合矩阵。(3)提出基于高阶统计Tucker张量欠定盲分离的眼电伪迹去除方法。针对眼电伪迹深入隐藏在脑电信号中难以有效分离的问题,利用Tucker张量模型中核张量能挖掘隐变量信息的特点,采用高阶统计Tucker张量模型,实现眼电伪迹去除中的欠定盲分离过程。由于高阶统计Tucker张量模型难以分解,采用分层交替最小二乘(Hierarchical Alternating Least Squares, HALS)算法提高模型分解速度。在此基础上,改善高阶统计Tucker张量欠定盲分离非平稳源信号估计性能,在高阶统计Tucker模型建立过程中引入傅里叶变换,构造一种时频高阶统计Tucker模型并结合最小残差共轭梯度算法,提高非平稳源信号估计的准确度。
[Abstract]:EEG signals can reflect the psychological state and physiological function of human brain and are widely used in the field of psychoanalysis and the diagnosis of clinical diseases. EEG signals are easily disturbed by other physiological signals in the process of acquisition, especially the uncontrollable Eye-electric signals (i.e., Eye-electric artifact). Therefore, the removal of Eye-electric artifacts from EEG signals is of great significance to signal analysis. On the basis of blind source signal separation method, the problem of eye electrical artifact removal is studied in this paper. The research includes the following three aspects: 1) an eye electrical artifact recognition method based on parametric model is proposed. First of all, the problem of Eye-electric artifact recognition in normal human brain electrical signals is studied. Aiming at the non-Gao Si property of Eye-electric artifact, a mixed heteroscedasticity transfer distribution model is established, and the parameters of the model are estimated by using conditional expectation maximization algorithm. Eye-electric artifact features are obtained to distinguish normal human brain electrical signals from Eye-electric artifacts. Secondly, in view of the problem of EEG and Eye-electric artifact recognition in epileptic patients, the above model is difficult to solve the problem effectively because both of them have the characteristics of non-Gao Si, and then according to the characteristics of phase synchronization of epileptic EEG signals, The instantaneous phase model of the signal was established by using the extreme learning machine, and the output weight of the model was taken as the signal feature to distinguish the EEG signal from the eye electrical artifact in the epileptic patient. At the same time, Hilbert-Huang transform is used to solve the problem that the instantaneous phase of non-stationary signal is difficult to get. The uniqueness of higher-order statistical CP Zhang Liang model can guarantee the unique solution of source signal estimation in the process of ocular-electric artifact removal from under-determined blind separation. In order to reduce the second-order correlation of observation signals and concentrate them on high-order statistical analysis, it is difficult to estimate the source signals due to the strong correlation of observation signals in under-determined blind separation, so that the principal component analysis (PCA) method is introduced to reduce the second-order correlation of observation signals. The higher-order statistical CP Zhang Liang model is constructed by using the principal component matrix of the observed signal and decomposed to improve the performance of the under-determined blind source estimation. For the problem of nonnegativity of mixed matrix in underdetermined blind separation, a regularized hierarchical alternating least-squares method is used to decompose the higher-order statistical CP Zhang Liang model to ensure that the decomposition process of the model is a non-negative decomposition process. Furthermore, the nonnegative mixed matrix in underdetermined blind separation is obtained. (3) an eye electrical artifact removal method based on high-order statistical Tucker Zhang Liang subblind separation is proposed. Aiming at the problem that Eye-electric artifact is difficult to be separated effectively in EEG signal, using kernel tensor in Tucker Zhang Liang model to mine hidden variable information, high order statistical Tucker Zhang Liang model is adopted. The process of undetermined blind separation in the removal of eye electrical artifacts is realized. Because the higher-order statistical Tucker Zhang Liang model is difficult to decompose, the hierarchical alternating least square algorithm is used to improve the decomposition speed of the model. On this basis, the estimation performance of high order statistical Tucker Zhang Liang underdetermined blind source signal is improved, and Fourier transform is introduced in the process of establishing high order statistical Tucker model. A time-frequency high-order statistical Tucker model is constructed and the minimum residual conjugate gradient algorithm is used to improve the accuracy of non-stationary source signal estimation.
【学位授予单位】：大连理工大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TN911.7;R338
，

本文编号：1991608