基于频域盲解卷的噪声信号分离

由于机械噪声传播过程中存在反射等多种因素影响，大多数情况下混合噪声分离更适合采用卷积模型，为此提出了一种多频点盲解卷算法。有别于传统的频域盲解卷算法，新算法利用有限的少数几个频率点直接从频域模型恢复出时域噪声信号。算法为瞬时混合盲分离。主成分分析一瞬时混合盲分离结构，首先对给定的每一个频率点执行瞬时混合盲解卷算法，获得噪声源的基本估计，然后再经过主成分分析和第二次盲源分离。提高分离性能和增加算法鲁棒性。由于算法不需要对所有频率点执行瞬时混合分离，计算量小，同时也不存在传统频域盲解卷算法排列顺序不确定性的缺点，具有较好的应用价值。仿真实验证实了新算法能有效地分离机械噪声信号。

Noise Seperation Based on Blind Deconvolution in Frequency Domain

Because of effect of many factors such as reverberation, it is appropriate in general to formulate machine noise spreading process as a convolution model.To separate the convoluted mixed signals,a multi-frequency points blind deconvolution approach is presented,which uses limited frequency points to directly recover time domain noise signals from its frequency domain model.Being different from traditional frequency domain blind deconvolution algorithms,using this new approach firstly runs an instantaneous mixture blind source separation(BSS) step on each frequency point to get a rough estimation of sources,and then refines the sources by running principal component analysis(PCA) and running BSS step again.Because the new approach only operates on a few frequency points,it is efficient in computation and has no permutation indeterminacy problem which is great trouble for taditional frequency domain blind deconvolution algorithms.It has extensive application prospect including machine noise separation.Analysis and simulation demonstrate its good performance in mixed noise separation.