【作者】 王尔馥；

【导师】 张乃通；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程， 2009， 博士

【摘要】 人们通过对自然界和现实生活中大量信息的分析和处理来获得认识和改造世界的能力。这些信息可能是已知的,但更多情况下对必要的信息却是未知的。将待认知的对象抽象为源信号,当源信号全部或部分已知时,根据已知信号通过适当的变换就可以将源信号还原出来。然而,在源信号和传输信道信息均缺失的情况下,仅依据有限的观测信号来分离、恢复源信号的过程,称为盲源分离。自从Herault和Jutten提出用神经网络方法实现语音信号盲源分离,在该领域作出开创性工作以来,关于盲源分离理论的研究层出不穷,并取得了丰硕的成果。盲源分离先后在语音识别、图像处理、地震勘探以及生物医学等领域获得了成功应用,引起了信号处理学界和神经网络学界的共同关注。随着应用化进程的不断推进,盲源分离算法所存在的问题也不断显现。对盲源分离算法进行深入研究并根据应用背景的需要提出改善性能的实用化技术,对进一步拓宽盲源分离的应用领域具有重要意义。本文围绕这一热点课题展开,并将研究重点放在卷积混合模型以及增强抗噪声鲁棒性和降低接收设备开销的实用化盲源分离技术上。按照传输信道混合特性的不同,盲源分离可分为线性瞬时混合模型和线性卷积混合模型,二者通过Fourier变换及其逆变换可以相互转换。学者们从不同角度出发,已经提出了很多改善盲源分离算法性能的方案。概括地说,现阶段尚存的有待进一步改进的问题主要集中在以下几个方面:如何克服模型中传输时延的影响求解出卷积混合信号,如何抵消加性高斯白噪声的影响保证分离算法的性能,如何利用尽量少的接收阵元来分离尽量多的混合信号。本文在对瞬时混合模型进行分析的基础上研究卷积混合模型及实用化盲源分离技术。盲源分离问题需要满足基本假设条件,其中独立性假设是核心所在,基于此建立起来的独立分量分析理论在用来分离独立源信号时与盲源分离是等价的。相似系数和性能指数是两个常用的评价算法分离性能的指标,分别从分离信号和分离矩阵的角度进行定义。盲源分离问题由分离准则和优化算法两部分构成,分离准则主要有基于信息论、基于最大信噪比和基于高阶累积量三类。本文采用自然梯度优化算法,分别对三类准则下的瞬时混合盲源分离算法进行仿真,结果表明高阶累积量准则对高斯噪声具有不敏感性,更适用于无线通信系统。当加权协方差矩阵的特征值接近时,基于高阶累积量的算法容易陷入局部极值而无法正确分离所有信号,基于归一化峭度的顺序盲源提取算法可以从混合信号中提取出具有特定属性的信号。线性卷积混合模型经过Fourier变换可以转化为频域的线性瞬时混合模型,因此解卷积至少可以有时域和频域两类方法。时域分离算法卷积计算量大、方法复杂,收敛性能也一般,频域分离算法可以利用FFT快速运算,但是需要解决各个频点上分离子信号排序不一致的问题。频域分离模型体现了噪声消除与信号分离之间的矛盾,本文通过对无噪情形下混合矩阵结构的分析,找到分离子信号排序不一致的根源并提出两项解决方案。定义邻频幅角比参数来寻找差错频点位置,纠正相应位置上的分离子信号到一个相同的顺序上,解决了频域算法的排列顺序不一致这个瓶颈问题。自适应耦合法引入耦合因子来保持相邻频点上分离矩阵的相关性,无需单独的排序过程,降低了出现顺序混淆的概率。仿真表明其效果虽然没有邻频幅角比好,但避免了额外的运算量开销,可以在精度要求稍低的场合下使用。总体上来说盲源分离算法尚处于理论研究阶段,推广到实际系统时存在着一些制约算法性能的因素。一方面,接收设备处存在加性高斯白噪声,无论是高阶累积量还是子空间理论,都只具备十分有限的噪声抑制能力。本文提出采用时频分析作为噪声预处理手段,首先提高观测信号的信噪比继而再进行盲源分离。分别设计了分阶段噪声预处理方案和时频联合两步消噪预处理方案,充分发挥了经验模态分解收敛速度快和小波变换消噪性能稳定的优势,两步化的预处理方案可以使观测信号的信噪比降低到一个较为理想的程度,提高了算法整体的抗噪声性能。另一方面,实际应用中多数情况下系统是一个黑匣子,无法根据源信号的数目来设计接收阵列的规模。即使已知源信号数目,当数目太大时接收阵元的规模也将同比增加,从而带来昂贵的设备开销。为简化接收设备的复杂度,有必要研究接收阵列低元化的盲源分离技术。本文在稀疏成分分析的基础上,着重于聚类环节的实现,提出一种将K-mean聚类和Kohonen神经网络聚类相结合的K-Kohonen混合聚类算法,不仅加快了Kohonen网络的收敛速度,而且可以获得精确的聚类结果。结合主分量分析,提出基于混合聚类算法的K-K-P欠定混合盲源分离算法,能够比较准确地估计信道混合矩阵,从而实现阵列低元化下的盲源分离。在算法实用化和降低设备消耗方面给出重要的研究结果。盲源分离在很多领域都获得了成功的应用,但应用于通信系统的盲源分离算法的性能却被一些实际因素所制约。作为应用性探讨,本文对卷积混合模型进行简化,在莱斯衰落信道下讨论了多径数、莱斯因子和最大传输时延三个参数对分离性能的影响,并通过仿真试验给出了数值结论。更多还原

【Abstract】 Human obtains the ability to understand and change the world by acquiring and analyzing the mass information from the nature and our daily lives. However, most of the time, the necessary information is unknown. In signal processing, we abstract these unknown subjects as source signal. With the complete or partial knowledge of the source signal, the source can be extracted signal according to some proper transform of the known information. Blind source separation (BSS) is regards as the process of separating and recovering source signal based on the limited observed signals, without the knowledge of source signal and channel information.Since Herault and Jutten have found the neural network method to realize the blind source separation for speech signals, a great many achievements emerge based on the BSS theoretical research whose application has succeed in the fields of speech recognition, image processing, seismic exploration, biomedical science and etc, as attracts the common attention from the circles of signal processing and neural network. With the promotion of the BSS application process, the problem of the algorithm is being clearer. It has important significances that to deeply research on the BSS algorithm and to find out a practical technology for the performance improving, as well as to further widen the application field of the BSS. This article is based on the hot subject, and pays more attention to the convolved mixture model and the practical BSS technology which increases the robust speaker feature and decreases the cost of receiving equipments.The signal transmission models for narrowband and broadband wireless communication system can be abstracted as BSS linear instantaneous mixture model and linear convolution mixture model, which can be transformed to each other by Fourier and its inverse transform. Researchers had studied several solutions to improve the BSS algorithm performance based on different angles. Generally, the questions being to promote now are based on the following parts: how to overcome the influence of the transmission delay in the model and solve the convolved mixture signals; how to cancel the influence of the additive white Gaussian noise to make sure the performance of the algorithm; how to separate more mixture signals with less use of receiving array elements.This paper studies the linear convolution mixture model based on the analysis of instantaneous mixture model. There are some assumptions for BSS problems, among which the independence assumption is the core. The independence component analysis theorem which is based on this assumption is equivalent to BSS when used to analyze independent source signal. Similarity coefficient and performance index are two parameters usually used to evaluate the performance of the separation algorithm, defined from the aspects of signal separation and matrix separation.BSS problem consists of separation criterion and optimization. The criterion can be based on information theory, maximum SNR or high order cumulants. The nature gradient optimization algorithm is used to simulate the instantaneous mixture BSS algorithm under these three criterions. The result shows that the criterion based on high order cumulants is robust under high Gaussian noise, and hence more suitable to wireless communication system. When the eigenvalues of the weighted covariance matrix are very close, the algorithm based on high order cumulants criterion is more likely to be stuck in local optimum and fail to separate signal. The sequential blind extraction algorithm based on normalized kurtosis. And the algorithm can extract the expected signal with certain properties from the combined signals.The linear convolution mixture model can be transformed to the linear instantaneous mixture model by Fourier transform. Therefore, there are two ways to solve the convolution, i.e., in time domain or frequency domain. Time domain separation algorithm is complicated and has moderate convergent performance, while frequency domain algorithm makes use of FFT, but needs to handle the inconsistent ordering problem of separated sub-signals. The frequency domain separation model shows the contradiction between noise elimination and signal separation. According to the analysis of mixture matrix structure under noiseless assumption, this paper finds out the reason of the inconsistent ordering and proposes two schemes as the solution. A neighbor frequency breadth-angle ratio is defined to find out the wrong frequency points and correct the separated sub-signals’order correspondingly. Coupling operator method introduces coupling factor to maintain the correlation of the separation matrix on the neighbor frequency without separate ordering process, as decreases the probability of ordering chaos. The simulation result shows that although it performs no better than neighbor frequency breadth-angle ratio, it avoids extra computation consumption, which can be applied in the case when high accuracy is not required.As a whole, because of some performance restricting factors of application in real systems, BSS algorithm is still in the stage of theory study. On one hand, receiving equipment of communication systems is under AWGN condition, so either higher-order cumulant or subspace theory only has limited noise suppression ability. In this paper, the time-frequency analysis preprocessing is applied as noise preprocessing method to increase SNR of received signals and separate blind source. A grading noise preprocessing scheme and a joint time-frequency two-step noise cancellation preprocessing scheme are designed to make full use of empirical mode decomposition’s fast convergence advantage and wavelet transform’s stable performance advantage. Two-step preprocessing scheme can achieve desired SNR of observed signals and improve the noise immunity performance of the algorithm. On the other hand, in most cases, the system is a black box, so that is impossible to design the size of a receiving sensor array according to the number of source signals, even if the number is known. If the number of source signals is too large, the number of sensor elements will be increased correspondingly. In order to simplify the complexity of receiving equipment, study on BSS based on low number of sensor technology is very necessary. Based on the sparse component analysis, this paper has proposed a mixture clustering algorithm combining K-mean clustering algorithm and Kohonen network clustering algorithm. The proposed algorithm focuses on the implementation of clustering, not only increases the convergence speed of Kohonen network, but also achieves accurate clustering. Combined with principle component analysis, K-K-P underdetermined BSS based on mixture clustering algorithm is proposed in this paper, which can estimate the channel clustering matrix accurately and implement BSS based on low number of sensor. Some important research results are given on algorithm practicability and reducing equipment consumption.BSS has been applied in many fields successfully, but the BSS algorithm applied in communication system is constrained by some practical factors. As a discussion for application, this paper applied BSS to rice fading channel, explained the effects of multipath on separation performance by simulation test.更多还原