-
近年来,盲源分离理论在无线通信领域中的应用受到了国内外学者的广泛关注[1-4]。盲源分离是指仅从观测混合信号中分离或提取期望的源信号和辨识混合矩阵。在盲源分离技术的辅助下,无线通信系统中频繁使用的导频序列可以免除或减少,进而提高系统的频谱效率;同时在先验信息缺乏的情况下,如信道未知或不可预知干扰,也能成功地分离出源信号信息,提高系统源信号恢复的鲁棒性。因此,研究无线通信盲分离方面的问题具有重要意义。
本文关注欠定盲源分离中欠定盲辨识问题,它对于实现欠定中的源信号恢复是至关重要的一步。现有的欠定盲辨识算法可以大致分为两大类,第一类是基于源信号稀疏性的聚类算法[4, 5],需要假设信源是稀疏性的或可以通过一些预处理线性变换得到稀疏源信号,如短时傅里叶变换、小波变换等。典型的代表性算法是degenerate unmixing estimation technique (DUET)算法和time frequency ratio of mixtures (TIFROM)算法[4-5]。这类算法依据观测混合信号稀疏化后的散点图特点,利用聚类算法穷举搜索来估计混合向量空间,从算法复杂度看,代价较高,尤其当观测通道数大于2时,实现比较困难。而且源的稀疏性要求一定程度上限制了此类算法的在实际中的应用范围。第二类算法是基于统计特征代数结构的张量分解[6-10],这类算法借助张量模型分解的唯一性特性,不需要源的稀疏性约束,应用范围更广。此类算法的一般原则是,首先基于统计特征建立一系列核函数,接着利用核函数堆叠三阶张量模型,最后利用平行因子分解求混合矩阵。典型的代表算法有基于二阶协方差的second-order covariance based blind identification of undertetermined mixtures (SOBIUM)[6]算法和基于四阶累积量的fourth-order cumulant based blind identification of undertetermined mixtures (FOBIUM)[7]算法。对于上述算法的核函数建立,由于四阶累积量的估计较为复杂,且需要较长的采样才能更好地提取统计信息;而二阶协方差又失去了四阶累积量抗噪声性。实际中为了更好地提取统计信息需要建立大量的核函数,再利用平行因子分解,其实现的复杂度较高。
基于提取更优的统计信息和降低上述算法的复杂度两方面考虑,本文提出了两个新的策略。一方面,基于一种新的统计方式建立核函数,即广义协方差[11-12],能够更好地从采样的数据中提取统计信息,它不仅含高阶的统计信息,而且维持着二阶协方差简单的二维数值结构。另一方面,借助塔克(Tucker)张量分解[8, 13]用于估计混合矩阵,原来构建的张量模型被压缩为一个低维的核张量,再进行基于交替最小二乘的平行因子分解估计混合矩阵,可以有效地减少计算量和运行时间,降低算法的复杂度。理论分析和仿真实验证明了提出的算法generalized covariance based blind identification of undertetermined mixtures (GCBIUM)与SOBIUM和FOBIUM算法相比,在计算复杂度和辨识性能上具有更好的竞争优势。
Underdetermined Blind Identification Algorithm Based on Generalized Covariance and Tensor Decomposition
doi: 10.3969/j.issn.1001-0548.2016.06.003
- Received Date: 2015-03-23
- Rev Recd Date: 2015-10-23
- Publish Date: 2016-11-01
-
Key words:
- blind source separation /
- cumulant /
- generalized covariance matrix /
- tensor decomposition /
- underdetermined blind identification
Abstract: In view of the estimation problem of mixing matrix in the underdetermined blind source separation (UBSS), a novel underdetermined blind identification algorithm is proposed. This proposed algorithm employs the statistical and structure properties of generalized covariance and the compressive characteristic of Tucker decomposition. Firstly, the core functions are built based on generalized covariance matrix. Then the core functions are stacked as a three-order tensor, and the tucker decomposition of constructed tensor is executed to estimate the mixing matrix. The proposed algorithm has not only the better identification performance, but also the lower computational complexity. At last, the simulation experiments demonstrate the effectiveness of the proposed algorithm.
Citation: | LUO Zhong-qiang, ZHU Li-dong. Underdetermined Blind Identification Algorithm Based on Generalized Covariance and Tensor Decomposition[J]. Journal of University of Electronic Science and Technology of China, 2016, 45(6): 893-897. doi: 10.3969/j.issn.1001-0548.2016.06.003 |