The independence priori is very often used in the conventional blind source separation （BSS）. Naturally, independent component analysis （ICA） is also employed to perform BSS very often. However, ICA is difficult to use in some challenging cases, such as underdetermined BSS or blind separation of dependent sources. Recently, sparse component analysis （SCA） has attained much attention because it is theoretically available for underdetermined BSS and even for blind dependent source separation sometimes. However, SCA has not been developed very sufficiently. Up to now, there are only few existing algorithms and they are also not perfect as well in practice. For example, although Lewicki-Sejnowski＇s natural gradient for SCA is superior to K-mean clustering, it is just an approximation without rigorously theoretical basis. To overcome these problems, a new natural gradient formula is proposed in this paper. This formula is derived directly from the cost function of SCA through matrix theory. Mathematically, it is more rigorous. In addition, a new and robust adaptive BSS algorithm is developed based on the new natural gradient. Simulations illustrate that this natural gradient formula is more robust and reliable than Lewicki-Sejnowski＇s gradient.
Underdetermined blind signal separation （BSS） （with fewer observed mixtures than sources） is discussed. A novel searching-and-averaging method in time domain （SAMTD） is proposed. It can solve a kind of problems that are very hard to solve by using sparse representation in frequency domain. Bypassing the disadvantages of traditional clustering （e.g., K-means or potential-function clustering）, the durative- sparsity of a speech signal in time domain is used. To recover the mixing matrix, our method deletes those samples, which are not in the same or inverse direction of the basis vectors. To recover the sources, an improved geometric approach to overcomplete ICA （Independent Component Analysis） is presented. Several speech signal experiments demonstrate the good performance of the proposed method.
Bofill et al. discussed blind source separation （BSS） of sparse signals in the case of two sensors. However, as Bofill et al. pointed out, this method has some limitation. The potential function they introduced is lack of theoretical basis. Also the method could not be extended to solve the problem in the case of more than three sensors. In this paper, instead of the potential function method, a K-PCA method （combining K-clustering with PCA） is proposed. The new method is easy to be used in the case of more than three sensors. It is easy to be implemented and can provide accurate estimation of mixing matrix. Some criterion is given to check the effect of the mixing matrix A. Some simulations illustrate the availability and accuracy of the method we proposed.
