改进白化的MDL快速独立分量分析算法

MDL-FastICA algorithm based on improved whitening

摘要:

采用主分量分析法(PCA)进行的白化处理，可能会错误估计信号子空间维数，且未考虑噪声影响。提出了一种基于最小描述长度(MDL)准则信源个数估计改进白化的盲分离算法。通过信源个数估计确定信号子空间的维数，区分信号与噪声子空间，并估计噪声平均方差，对信号特征值进行修正，进而减小噪声影响，提高算法分离性能。仿真表明，在信噪比高于5 dB时，MDL估计正确估计概率趋近于1，改进白化的MDL快速独立分量分析(FastICA)算法比经典FastICA算法分离性能有较为明显的提高。

Abstract:

The whitening processing by Principal Component Analysis(PCA) may inaccurately estimate the signal subspace dimension without considering the noise impact. This paper proposes a blind signal separation algorithm based on improved whitening by Minimum Description Length(MDL) for estimating the source signal number. It determines the subspace dimension based on the source signal number estimation, and distinguishes between signal and nosie subspace. It can improve the separation performance. The eigenvalues of signal will be corrected by noise average variance estimation，thus it can reduce the impact of noise. Simulation results show that the exact estimate probability by MDL will approach to 1 at hyper-5dB SNR，and the separation performance of improved whitening MDL Fast Independent Component Analysis(FastICA) algorithm can be improved distinctly compared with that of traditional FastICA algorithm.