【作者】 马守科；

【导师】 何选森；

【作者基本信息】 湖南大学 ， 信号与信息处理， 2007， 硕士

【摘要】 在源信号和传输信道未知的情况下,仅利用传感器接收的观测数据恢复出源信号,称为盲源分离。因其在天线阵列信号处理、多用户通信、语音和医学信号处理等领域具有广阔的应用前景,从而成为当今信号处理学界的研究热点。盲源分离的算法包括离线批处理和自适应在线处理两种,在信道不恒定或实时性要求比较高的场合,必须采用自适应在线算法,而在线算法存在收敛速度慢和稳态误差大的缺点。本文正是以提高算法的收敛速度和稳态精度为目标,对盲源分离中的在线算法进行研究的。本文首先系统阐述了盲源分离问题的主要解决方法——独立分量分析,从信号模型出发,分析了独立分量分析的不确定性和基本假设,研究了独立性测度准则及其一致性。总结了常用的优化算法及算法性能评价测度。通过分析影响自适应盲分离算法性能的主要因素,并结合相关算法本身的特点,提出了基于自然梯度算法的步长迭代规则和基于扩展Infomax算法的步长更新规则:(1)针对自然梯度算法提出的两阶段步长迭代规则,是通过估计信号之间的相依性测度,在不同阶段采用不同的步长算法,并能有效的检测信道矩阵是否改变,使得算法无论是在信道矩阵恒定环境下还是在信道矩阵改变的环境下均能有效的分离出服从单类分布的信源。(2)针对扩展的Infomax算法,提出了判断信道矩阵是否改变的检测机制,改进了峭度的在线估计模型,并通过分析峭度收敛曲线与分离状态之间的关系,提出了随峭度方差而变化的步长更新规则。与以往算法相比,该算法不仅收敛速度快,稳态误差小,且更能适用于信道矩阵改变的环境。最后,对更复杂的超完备盲分离进行了初步的研究,深入分析了各算法的特点,将其分为两类,介绍了各算法的主要思想、优缺点及其适用的环境,并讨论了下一步的研究方向。更多还原

【Abstract】 Under the situation with unknown transmit channel and unknown sources, only the observation data received by the sensors can be used to resume the source signals, which is called the blind source separation (BSS). Because its wide application prospects in many fileds, such as aerial array signal processing, multiuser’s communication, the speech and medical signal processing, thus it becomes the research focus in the signal processing recently. The algorithm of blind source separation includes offline batch processing and online selfadaptation processing. Under the situation of changeable channel or realtime processing necessary, the online selfadaptation algorithm must be adopted. However, there are two crucial problems for online algorithm: slow convergence speed and the fluctuation in steady state. Any research has been done of the online algorithm for blind source separation in this paper, exactly regards to improve the convergence speed and precision in steady state.This paper has explained the main solution of the blind source separation problem systematically at first :The independent component analysis (ICA) . From the signal model, the uncertainty and basic supposing of ICA was analyzed, the measurement criterion of independent degree and its consistency was studied. Some kinds of optimized algorithm and performance index of algorithm are summarized.By analysized the main factor which Influente the performance of online algorithm, combined the characteristic of the relevant algorithm itself, two differente variable step-size update rule based on natural gradient algorithm and extended Infomax algorithm respectively were proposed: (1) Two stage step-size update rule was proposed based on natural gradient algorithm. It divides the process of separation into two stages by using the independent measurement of the outputs. We adopt different variable step-size algorithm at different stage. It can detect whether the channel matrix is changed or not efficinently. So this algorithm can separate single distribute source efficinently under the situation for variable channel or invariable channel. (2)Based on extended Infomax algorithm, an improved online estimation model of the kurtosis was introduced, which is according to the detection of mixed channel matrix.Through analysing the relation between the kurtosis curve and separate state, proposed a novel step-size update rule based on the covariance of kurtosis. Compared with past algorithm, this algorithm not merely restrains the speed fast, the stable state error is little, and can be suitable for the environment that channel matrix change.Finally, more complicated over-complete BSS have been studied, analyzed the characteristic of every algorithm deeply, divided them into two kinds, introduced the main thought, advanced and disadvanced,the suitable environment of most algorithms, and has discussed the next research direction.更多还原