الفهرس 
chapter 1Only 14 pages are availabe for public view
 |  | from | 187 |  |  |
| | | from | 187 | | |
Abstract
Blind Source Separation is a rising research filed in late 20th century, as a kind of new data processing method. It combining with artificial neural network, information theory and computer science, has been applied in many areas, especially in the biomedical engineering, biometrics, security, medical imaging, water marking, speech processing, communication system, data mining, and so on.
The goal of BSS can be considered as estimation of true physical source signals and parameters of a mixing system. The approaches developed by researchers in the last few years can be classified into two methodologies, namely over-determined BSS and underdetermined BSS, according to the number of source signals and observable mixed signals. BSS that has fewer sensors or observable mixed signals than source signals is called underdetermined BSS while a BSS that has more sensors than sources is called over-determined BSS. Traditional blind source separation is based on over-determined; however we focused on the underdetermined which is more consistent with actual situation and more challengeable where the number of sensors or mixtures (observations) is less than the number of source signals.
A commonly used method for BSS problem is the independent component analysis (ICA). ICA does not consider sensor noise which generally should be taken into consideration. The performance of ICA algorithm deteriorates and the separation quality decreases when the noise level increases. Moreover in many practical situations, there are a large number of source signals but few sensors. In this case, one would like to lump the low-intensity sources together and treat them as noise while focus on the high-intensity ones. Another major drawback of ICA is that the mixing matrix and the magnitude of original source signals cannot be estimated due to its ambiguities and that the variances (energies), sign, and the order of the independent components cannot be determined. Although some methods based on the ICA were proposed to overcome the ambiguities, the independent components and the mixing matrix of the ICA cannot represent the original source signals and mixing matrix respectively. Consequently, ICA fails to solve underdetermined BSS problems.
One of the important approaches that will be used to separate a set of signals from a set of mixtures and obtain one signal is empirical mode decomposition (EMD) that represents the coordinates of all the input signals. It can be considered as a source to derive ICA. The recent trends in blind source separation is to consider problems in the framework of matrix factorization and exploit a priori knowledge about true nature, morphology or structure of latent (hidden) variables or sources such as non-negativity, sparseness, spatio-temporal decorrelation, smoothness or lowest possible complexity. The blind source separation functions or similarity functions and convergences are used to make update rules to estimate the mixing system and the true signals. In this thesis matrix factorization combined with machine learning techniques and some preprocessing steps will be used as a hybrid system in the separation of underdetermined mixtures. A criticism of some matrix factorization approaches is the high computational cost of the resulting algorithms. Therefore, this thesis aims to develop algorithms that are computationally undemanding so that they can realistically be applied when rapid processing of the data is required.
Using this hybrid system the thesis develops four new algorithms for separating sparse and non-signals in case of the number of source signals is known or unknown. The first algorithm is developed employing projected sequential sub-coordinate optimization approach in the framework of semi-nonnegative matrix factorization combined with fuzzy clustering, which is based on utilizing prior information about the sparseness of the human prints speech signals of noise free and use the multi-layer mechanism to improve the separation performance. The second algorithm is developed employing S transform with rough set clustering and modified projected gradient approach in the framework of general matrix factorization without resorting to any sparsity conditions to separate super- and sub- Gaussian. The third algorithm is developed based on single source detection employing S transform with shadowed clustering and modified gradient descent conjugate gradient approach in the framework of general matrix factorization without resorting also to any sparsity conditions to separate super- and sub- Gaussian, sparse and non-sparse source signals but in this time with additive different types of noises such as Gaussian white noise, pink noise, blue noise, violet noise, grey noise, Brownian noise as the first thesis to consider this case. The fourth algorithm is developed employing short-time Fourier transform with rough fuzzy clustering and modified Lin’s projected gradient algorithm likewise in the framework of general matrix factorization. Finally we make performance evaluation and analysis for the proposed algorithms and compare the results with recent references.