基于张量正则分解的时频混叠信号欠定盲分离方法

Canonical decomposition approach for underdetermined blind separation of non-disjoint sources

针对时频同时混叠条件下的欠定盲源分离(UBSS)问题,提出了一种基于四阶累积量(FO)与张量正则分解相结合的算法。首先构建观测到的混合信号的四阶累积量,并利用高阶累积量的"半不变性"将其表示成四阶张量的形式,然后采用线性搜索迭代最小二乘算法对张量进行分解并获得混合矩阵的估计,最后根据估计出的混合矩阵,采用最小均方误差波束形成器算法,完成源信号的恢复。仿真结果表明该方法的有效性,与已有算法相比提高了信号盲分离的性能。

This paper proposes a method of underdetermined blind separation of non-disjoint sources（UBSS）based on fourthorder cumulant（FO）and tensor decomposition.By semi-invariance of high-order cumulant,the FO is presented as statistics of the observed signal as fourth-order tensor;hence the mixed matrix is estimated by tensor decomposition with line search alternating least square.Finally,with the estimated matrix,sources are recovered by minimum mean-squared error-based beamforming.Simulations illustrate the validity of the method and show that the proposed method outperforms the existing methods in performance significantly.