基于改进FastICA算法的混合语音盲分离

发布时间：2018-02-26 19:01

本文关键词： 盲分离独立成分分析负熵峭度快速固定点算法　出处：《上海交通大学》2015年硕士论文　论文类型：学位论文

【摘要】：目前信号处理领域中最热点的研究问题之一就是盲源分离问题(BBS),它的主要研究工作就是,在对系统的源信号和混合系统都未知的情况下,处理一组以时间序列或者并行信号的形式表示的观测变量,最终分离出想要求解的源信号。盲源分离的典型例子有很多:手机中的射频干扰信号、传感器记录的脑电波以及麦克风录取的混合语音信号等。而处理这类问题最有效的方法就是独立成分分析(ICA),它是随着盲分离问题研究的不断发展而引起广泛关注的。ICA方法的基本思想是,首先通过传感器采集到观测信号,再根据信号的统计特性寻找一个合适的目标函数,通过选取迭代算法对目标函数优化,得到最优的解混信道,将观测信号通过解混信道的处理就可以得到想要求解的源信号的估计。基于盲源分离问题的研究背景下,本文简述了独立成分分析的研究意义和发展历史,通过对ICA基本原理的研究分析得到基本模型存在的不确定性以及约束条件:研究对象必须是非高斯信号,并且满足相互独立的统计特性。ICA问题可以简化为通过一定的优化算法得到选定的目标函数的最优解,从而分离出待求解的源信号,由此从理论上重点介绍了独立成分分析的几种典型的目标函数和优化算法。在ICA模型估计之前,必须要对测量信号作预处理,包括中心化和白化这两个处理过程,本文从理论上说明了预处理可以有效地减少ICA中需要预估模型参数的数目。快速固定点算法(FastICA)是独立成分分析中最常用的一种快速算法,根据非高斯性的评价指标的不同,本文介绍了两种FastICA算法:基于峭度的FastICA算法和基于负熵的FastICA算法。本课题的核心研究工作在于,通过对两种原有算法的原理分析,针对其存在的问题,分别提出了相应的改进方案。对于基于峭度的FastICA算法存在的收敛不稳定问题,本文提出了通过共轭梯度法对原算法进行改进,实验结果表明,改进后的算法不仅分离效果更佳,而且改善了原算法收敛不稳定的问题。对于基于负熵的FastICA算法,本文分别提出了通过最速下降法以克服原算法易受初始值影响的缺点,用差商法代替求导以降低了原算法的复杂性。仿真实验结果表明,改进后的算法不仅分离出的信号更逼近源信号,而且解决了初始值的问题,算法速度也变得更快。最后,将本文中提出的两种改进的FastICA算法分别应用于分离人工混合和实际混合的语音信号,对分离结果进行分析得出,基于峭度的改进算法的分离效果更佳,而基于负熵的改进算法的收敛速度更快,但由于两种改进算法的分离效果差距并不大,基于负熵的改进算法以其收敛速度的优势更具有实用性。
[Abstract]:At present, one of the hottest research problems in the field of signal processing is the blind source separation (BSS) problem. Its main research work is that the source signal and hybrid system are unknown. Processing a set of observation variables in the form of a time series or a parallel signal, and finally separating the source signal to be solved. There are many typical examples of blind source separation: radio frequency interference signals in mobile phones, Brain waves recorded by sensors and mixed speech signals recorded by microphones, etc. The most effective way to deal with this kind of problems is independent component analysis (ICA), which has attracted wide attention with the development of blind separation research. The basic idea of the ICA approach is, First, the observed signals are collected by the sensor, then a suitable objective function is found according to the statistical characteristics of the signal, and the optimal unmixing channel is obtained by selecting the iterative algorithm to optimize the objective function. The estimation of the source signal to be solved can be obtained by processing the observed signal through the de-mixing channel. Based on the research background of blind source separation problem, the research significance and development history of independent component analysis (ICA) are briefly described in this paper. Based on the analysis of the basic principle of ICA, the uncertainty and constraint conditions of the basic model are obtained: the object of study must be non-#china_person0# signal, The ICA problem can be simplified to obtain the optimal solution of the selected objective function by a certain optimization algorithm, and the source signal to be solved can be separated. Therefore, several typical objective functions and optimization algorithms of independent component analysis (ICA) are introduced in theory. Before ICA model estimation, the measurement signal must be preprocessed, including centralization and whitening. In this paper, it is theoretically explained that preprocessing can effectively reduce the number of estimated model parameters in ICA. Fast fixed point algorithm (FastICA) is one of the most commonly used fast algorithms in independent component analysis (ICA), according to the difference of non-#china_person0# evaluation indexes. This paper introduces two kinds of FastICA algorithms: FastICA algorithm based on kurtosis and FastICA algorithm based on negative entropy. For the unsteady convergence of FastICA algorithm based on kurtosis, the conjugate gradient method is proposed to improve the original algorithm. The experimental results show that the improved algorithm not only has better separation effect. Moreover, the unsteady convergence of the original algorithm is improved. For the FastICA algorithm based on negative entropy, this paper proposes to overcome the shortcoming that the original algorithm is vulnerable to the influence of initial value by the steepest descent method. The difference quotient method is used instead of the derivative method to reduce the complexity of the original algorithm. The simulation results show that the improved algorithm not only approximates the source signal, but also solves the problem of initial value, and the speed of the algorithm becomes faster. The two improved FastICA algorithms proposed in this paper are applied to the separation of speech signals with manual mixing and actual mixing respectively. The analysis of the separation results shows that the improved algorithm based on kurtosis has better separation effect. The improved algorithm based on negative entropy has a faster convergence speed, but because the separation effect of the two improved algorithms is not big, the improved algorithm based on negative entropy is more practical because of its advantage of convergence speed.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN912.3

【共引文献】