【作者】 刘建强；

【导师】 冯大政；

【作者基本信息】 西安电子科技大学 ， 信号与信息处理， 2009， 博士

【摘要】 盲源分离是在源信号与传输信道未知的情况下仅利用接收天线的观测数据提取或恢复统计独立的源信号。盲源分离因为在语音处理、医学信号处理,图像增强与无线通信等诸多领域具有广泛的应用前景,从而引起了信号处理学界和神经网络学界的共同兴趣。本文围绕这一热点课题展开,并把研究重点放在非平稳环境中的盲源分离算法研究,本文的主要工作包括以下几个方面:1.提出一种变步长、变动量项因子的自适应改进的自然梯度算法实现源信号瞬时混合的盲分离。在后向传播算法的启发下,在自然梯度学习过程中结合动量项以加快收敛速度,同时改善自然梯度算法的稳定性。然后在分离模型中提出一个合适的测度函数自适应控制步长和动量项因子,由此得到的变步长和变动量项因子的改进的自然梯度算法适合解决时变环境下的盲源分离问题。实验表明与经典的自然梯度算法和其它改进的自然梯度算法相比,即使在信源个数很多的情况下,本文所提自适应改进的自然梯度算法有更快的收敛速度和更好的稳态精确性,当混合矩阵突变或信号功率突变时自适应改进的自然梯度算法依然有较好的跟踪能力。此外,结合自然梯度和非线性主分量分析提出了一种块递归的盲源分离方法,构造出按块递归更新的矩阵方程,然后用QR分解和回代法求解该矩阵方程得到最优分离矩阵。所提方法与已有递归型盲源分离算法相比适于实时处理且遗忘因子的选择相对简单,与其它块处理算法相比有较快的初始收敛速度。2.针对混合矩阵发生突变的情况,提出一种基于时变遗忘因子递归广义特征分解的非白源盲分离算法。首先给出一种新的协方差矩阵的逆矩阵和时延相关矩阵乘积的递归更新方程,广义特征向量的估计经由近似幂迭代法和压缩处理在线实现,得到的在线算法可以避免同时估计协方差矩阵及其逆矩阵。为跟踪混合矩阵的突然变化,提出一种新颖的基于广义特征向量关于协方差矩阵正交特性的在线决策规则,可以判别混合矩阵是否发生突变,从而利用时变遗忘因子的递归广义特征分解算法在混合矩阵发生突变后能获得较好的跟踪能力及精确的稳态性能。3.充分利用人类发音的特点和语音信号的非平稳特性,提出一种数目未知的语音信号瞬时混合的盲分离和信源数目检测算法。首先利用递归广义特征分解在线估计对应于最大广义特征值的广义特征向量,利用向量相似度定义广义特征向量的平均相似度用以拟合信道互扰性能曲线,并根据平均相似度曲线提取出“高相似度区间”近似“分离区间”。然后进一步提取分离性能较好的“高相似度区间”并剔除部分“混合区间”,得到“高分离度区间”。然后对“高分离度区间”中的广义特征向量进行多阶段聚类确定源信号的数目,在实现估计信号数目的同时完成信号的盲分离。此外,所提算法可以克服盲提取问题中先提取出的信源性能好而后提取出的信源性能差的缺点。4.针对现有卷积混合盲源分离的频域算法存在的问题,即恢复出的信号是源信号和一个未知的滤波器的卷积,且由于部分频率点处盲源分离方法的失效和无法精确解决排列和尺度模糊问题,分离出的信号中会泄露进其余的信号分量,影响分离性能。提出一种基于多信道语音增强的频域盲源分离后处理方法以消除空间干扰和背景噪声。该方法有机结合盲源分离技术与阵列处理技术,既可充分利用空域信息,又无需增加其它先验信息。首先在频域盲源分离方法中利用语谱分裂技术得到M×N个分离信号,将多输入多输出(MIMO)混合系统分裂成N个单输入多输出(SIMO)系统,泄露进该信号的其它信号分量可以视为干扰信号,然后分别用频域多信道语音增强方法重构出各个语音信号,以消除不同信源间的空间干扰和环境噪声。仿真结果表明该算法具有良好的性能。更多还原

【Abstract】 Blind source separation (BSS) aims to extract independent but unobservered source signals from their mixtures captured by a number of sensors without knowing the mixing coefficients. Over the past two decades, the problem of BSS has received much attention in various fields, such as speech and audio processing, image enhancement and biomedical signal processing. The main works can be summarized as follows:1. An adaptive improved natural gradient algorithm for blind separation of instantaneous mixtures of independent sources is proposed. First, inspired by the well-known back-propagation algorithm, we incorporate a momentum term into the natural gradient learning process to accelerate the convergence rate and improve the stability. Then, an estimation function for the adaptation of the separation model is obtained to adaptively control a step-size parameter and a momentum factor. The proposed natural gradient algorithm with variable step-size parameter and variable momentum factor is therefore well suited to blind source separation in a time-varying environment. The expected improvement in the convergence speed, stability and tracking ability of the proposed algorithm is demonstrated by extensive simulation results in both time-invariant and time-varying environments. The ability of the proposed algorithm to separate extremely weak (or badly scaled) sources and many sources is also verified. In addition, a block recursive approach for blind source separation is presented. Firstly, based on natural gradient and nonlinear principal component analysis, a matrix equation is obtained by block recursive updating,and then the matrix equation is solved using QR factorization and back substitution to obtain the optimal separating matrix. Compared with other existing recursive-type BSS methods, the proposed algorithm is feasible to real-time processing, and the choice of the forgetting factor is simple. Compared with other block processing methods, the proposed algorithm has fast initial convergence speed.2. An efficient variable forgetting factor recursive generalized eigen-decomposition algorithm is developed for blind separation of non-white sources when the mixing matrix changes abruptly. We derive a new recursive update equation for the multiplication of a cross-correlation matrix and the inverse of a covariance matrix with compact form and low computational complexity. The generalized eigenvectors are recursively estimated by using the approximated power method and the deflation procedure. Without additional priori information of the mixtures, we propose a novel on-line decision rule to track the abrupt variations of the mixing matrix and then a variable forgetting factor recursive generalized eigen- decomposition algorithm for BSS is presented for the time-varying environments. The improved tracking ability and steady-state accuracy of the proposed algorithm are validated by the computer simulation results.3. By exploiting of speech nonstationarity, a method for estimating the number of sources from instantaneous mixtures of speech signals with unknown source number is presented, and then the sources are extracted. The first dominant generalized eigenvector is estimated by recursive generalized eigen-decomposition. The mean similarity curve of the estimated generalized eigenvector is introduced to fit inter-channel interference curve, and then“High Similarity Intervals”are extracted to approach“Separation Interval”. Moreover,“High Separation Intervals" are obtained by extracting“High Similarity Intervals”with better separation performance and eliminate“Mixtures Interval”. Final, the number of the sources is estimated with multistage clutstering techniques and the corresponding sources are extracted. The proposed algorithm can avoid suffering the error propagation of the deflation technique, which exists in all sequential algorithms.4. A new post-processing method for convolutive mixtures blind speech separation is proposed. It utilizes multi-channel signal enhancement to suppress spatial interference and background noise. Due to imprecision for solving the permutation ambiguity problem, frequency domain blind source separation has its fundamental limitation in separation quality. To overcome that, by splitting spectrograms, the M×N multi-input multi-output (MIMO) system will be converted into N single-input multi-output (SIMO) system in frequency domain blind speech separation system. Furthermore, to attenuate spatial interference from competing sources and background noise, the multi-channel signal enhancement method is exploited to reconstruct source signals from the N SIMO system respectively. The separation performance of the proposed algorithm is demonstrated by experiments.更多还原