基于无监督学习的源数估计及盲分离算法

针对目前欠定盲分离问题中源数未知,采取"两步法"进行分离源信号.在第一步聚类算法中,一般都假设源信号个数已知,即事先给定聚类数目,这类算法成功与否依赖于源信号个数的先验知识.为了有效解决这个问题,提出了一种新的基于模糊聚类分析的无监督学习算法,它利用观测信号之间的相似关系来确定模糊相似矩阵进行迭代算法,不但可以精确估计源信号个数,同时也能获得对混叠矩阵的精确估计.该方法进一步完善了"两步法",仿真结果表明了算法的有效性及优异性能.

Sources number estimation and blind separation algorithm based on unsupervised learning

The two-step approach is often used to separate sources in underdetermined blind separation problem.The first step is to estimate the mixing matrix by clustering algorithms using the observations,in which it is often supposed that the sources number is known,so the two-step approach depends on the prior information about sources number strongly.A novel underdetermined blind separation algorithm based on fuzzy clustering and fuzzy similar matrix is proposed,which can accurately estimate the sources number and the mixing matrix respectively.The last simulations show the good performance of the proposed algorithm.