盲信号分离算法的研究

【作者】 徐尚志；

【导师】 叶中付；

【作者基本信息】 中国科学技术大学 ， 信号与信息处理， 2005， 博士

【摘要】 盲信号分离技术是现代信号处理领域的一个新的研究热点。目前，盲信号处理技术可以被应用于语音信号处理，图像信号处理，通讯信号处理，水声信号处理，医学信号处理以及数据挖掘等诸多领域。也正因为盲信号分离技术具有如此广阔的应用前景，盲信号分离问题从开始到现在的二十多年的时间内，得到了国内外专家学者的广泛的研究讨论。盲信号分离技术也因此获得了飞速的发展。然而，这一领域的研究工作还没有达到完全解决问题的程度。这促使我们将盲信号分离算法的研究作为本论文的研究对象。 针对瞬时线性混叠情况下盲信号分离算法中存在的一些问题，对盲信号分离技术的理论、算法进行了较深入的研究。主要工作和创新如下： (1)：针对FastICA类方法不能保证信号提取的有序或者有目的性以及容易丢失弱信号的问题；提出基于非高斯性极大准则和空间划分的原理，利用智能算法进行优化的盲分离算法。实验也显示：在需要有序提取以及存在微弱信号的情况下，本文算法分离效果优于FastICA算法。 (2)：针对欠定条件下S-ICA算法在源信号不是足够稀疏情况下算法难以达到满意的分离效果的问题，提出在估计混合矩阵的基础上，通过简化混合矩阵建立新的混合矩阵，利用信号的独立性，从观测信号中提取源信号的频谱完成源信号分离的方法。实验结果也说明了算法的有效性。 (3)：研究分析了白化处理对盲信号分离模型中混合矩阵的影响，提出在2源情况下可根据白化数据的峭度和中隐含的信息直接计算分离矩阵的参数的方法。该算法在多源情况下，也只需要少量迭代，且不需考虑迭代初试值对算法的影响。文中实验也证实：该算法可以达到快速有效的完成超高斯和亚高斯的混合信号的分离任务。 (4)：研究了Infomax算法中评价函数的作用以及Ext-Infomax算法中基于峭度的模型切换单一评价函数存在一些不稳定因素。提出根更多还原

【Abstract】 Blind signal separation(BSS) is an interesting project in the field of signal processing.The technique of BSS can be applied in speech signal process-ing,image processing,communication processing,hydroacoustic signal process-ing,biomedical signal processing,data mining and many others.Because of wide usage of BSS,many researchers have devoted to the research of BSS.and the technique of BBS has been greatly improved.research on BSS has been going on for about twenty years. However,the problem of BSS has not been totally solved.All above stimulate us to be engaged in researches on BSS algorithm.The main parts of thesis are research on BSS theory and modification to BSS algorithm in difference practical field based on linear mix model.The main work and innovation are abstracted as follows.Since the algorithm of FastICA can’t ensure the extraction of sources in order according to sources’s character and fail with weak source,a blind signal extraction algorithm based on intelligent algorithm and Maximization of Non-gaussianity is proposed.AU simulation results show that the proposed algorithm show better performance to above problem than the algorithm of FastICA.Since S-ICA algorithm requires that sources are very sparse,which leads to bad performance sometimes.A new algorithm is presented by using the characteristic of SPARSE of the signals.It simplifies channel matrix and set up a new one. Based upon the independence among sources, the frequency spec-trums of the sources can be estimated from the observed signals in a new channel model.So that sources can be separated. Simulation studies are available to support the proposed algorithm.By research white operation how to effect on mix matrix,herein a new simpler and faster method based on higher order statistics is introduced in the 2 sources instance, which can estimate the separation matrix straightly. In more sources instance,only a few iterative operation should be required and there is no need to consider initial value.Good performance of simulations proves the prac-更多还原