Study On Blind Source Separation Of Linear And Instantaneous Mixtures

Blind source separation (BSS) is a rising technology in signal processing, and has potential application values in seismic exploration, sound processing, image enhancing, sonar, radar and biomedical engineering etc. The linear and instantaneous mixing is the basic mixing model, while algorithms in this model can be expanded into other models such as the convolutive model. So this dissertation focuses on the blind source separation of linear and instantaneous mixtures, and the following is the summarization of the main work.1. Accurate estimation of number of sources plays the primary role in BSS. Based on the detailed analysis of well-known source number estimation methods, an eigenvalue clustering-based estimation method is proposed. Due to the expected process of clustering, a new distance between the different classes is defined. Simulations show that this method not only gives accurate number in estimating the number of stationary sources, but also gives accurate number more possibly than other well-known methods in estimating the number of non-stationary sources.2. BSS with unknown and dynamic changing number of sources can be classified by two models, i.e. sensors noise-free model and sensors noise model. In sensors noise-free model, this dissertation proposes that the demixing matrix could be set as N×N dimension, and the feasibility of this approach is proved by comparing the entropy of different mixtures. Obviously, the new approach needs less computation, simulations also show it has better convergence than the general one. In sensors noise model, PCA is introduced to the samples of observed mixtures, and then the obtained principal components are set as the input of demixing network. Simulations show that this approach can improves SNR of the output of demixing network.3. FastICA bases on Newton iteration, so it's sensitive to the random initial. To solve this problem, a Quasi-newton iteration-based batch BSS algorithm is proposed. The detailed process of this algorithm with negentropy as contrast function is also presented. Simulations show that this algorithm is not only insensitive to the random initial, but also has a rapid speed which is closed to the speed of FastICA.4. The choice of step-size always reflects a tradeoff between misadjustment and the speed of convergence in adaptive BSS. A chaotic step-size is proposed in this dissertation, which is obtained through adding a chaotic disturbance to exponential decreasing step-size. Compared with the exponential decreasing step-size, the chaotic step-size is relative with the separation states of the demixing network. Simulations show that the chaotic step-size-based adaptive BSS outperforms other related approaches.