Research On Underdetermined Blind Sources Separation And PCMA Blind Signal Separation

This dissertation is devoted to a research on key technologies of blind separation of mixed signals,primarily including the blind separation of mixed multi-source signals and the blind separation of satellite PCMA signals while laying emphasis on the latter. The work finished in this paper was supported by the nation’s 863 programme and the army’s 9201 project.The mission of blind separation is to recover the original signals from a mixed multi-source signal. Blind separation, as a fast developing new technology, has been widely used in many fields such as communication signal processing, voice signal processing and medical signal processing, however still facing many crucial hard nuts to crack urgently.The chief difficulty the blind separation of mixed multi-source signal comes up against is the underdetermined mathematical condition. Most of the comparatively matured algorithms targeting at this kind of topics in the field are based on sparseness of the signal. Thus,the performances of the algorithms will be poor if the sparseness is weak, especially with the estimation of the total number of the source signals.The blind separation of PCMA signal is a special case of underdetermined blind source separation. The increased bandwidth-and-power efficiency due to bandwidth multiplexing, and the high anti-interception performance make the application of PCMA signals grow very fast. The blind separation of PCMA signal is of very high degree of difficulty both in theory and in practice. A lot of researches have been implemented, and many important achievements have been gained. Among them the algorithms based on particle filtering and the algorithms based on Per-Survival Processing (PSP) are universally recognized as the best two categories of algorithms while the PSP is the better due to its lower computational complexity. However the current PSP algorithms still have two problems.One is its still very high computational complexity that makes it difficult to be implemented on the current hardware platforms. The other problem is that the soft information contained in the received signals has not been tapped for further improvements of the comprehensive performances of the algorithms.The work of this thesis was just set up under above background.In the paper, the basic theories, algorithms and applications were analyzed, discussed and summarized first. And then several innovative algorithms targeting at above two topics were provided. Simulation results have proved the feasibilities and the effectiveness of the proposed algorhthms.The innovative achievements gained in this dissertation are summarized as follows.1.In respect of underdetermined blind source separation mainly for voice signals:A novel algorithm is proposed, in which the concept and the corresponding determination approaches of single source signal interval (SSI) were suggested. The observed signals are preprocessed and the data outside SSI are removed. And then the number of source signals of the preprocessed data is estimated by an improved K-means clustering algorithm,and the mixing matrix is then estimated. Compared with the conventional method of directly processing the observed whole data, the algorithm is of higher precision and can get relatively smooth curve of the probability distribution of the data within SSI with fewer filtering operations, thus is easy to obtain the local peak by peak detection and the number of source signals.Simulation results show that compared with the conventional clustering algorithms, the new algorithm is of lower complexity and higher precision of estimation.2.As regards PCMA blind signals separation:(1) An algorithm of SOVA-PSP based on feedforward nonbinary code is proposed which is more suitable for PSP implementation. The algorithm can provide soft information output that is very important for performance improvement, however is not available in the currently existent PSP algorithms. Simulation results show that the new algorithm can get 2 dB performance gains with the aid of channel decoding.(2) A BCJR-PSP algorithm based on maximum a posteriori is proposed, which is targeting at the problem of low reliability at the end of the data blocks in conventional SOVA-PSP algorithm.Simulation results have proved the performance improvements over the SOVA-PSP.(3) A new PSP algorithm with lower complexity is proposed, in which the timing is set to one of the constituent PCMA signals that is assumed to be perfectly synchronized with the sampling signals. Because the ISI with one of the two signals has already eliminated, the number of state in trellis graph is now immediately reduced from M2(L-1) to M(L-1) (where M is the order of the modulator, L is the length of equivalent channel response), thus the computational complexity is greatly reduced. This result makes the algorithm much easier to be implemented on hardware platform. Simulation results proved the reduction of the complexity with almost no performance loss.(4) Enlightened by the idea of CHASE decoding for linear block code, a new PSP algorithm was proposed. In the algorithm the results of soft output of PSP were sorted according to the corresponding reliability indices, and then only the mixed symbols with low reliability in the sorted results were reconstructed. Errors will be corrected by the comparison in terms of Euclidian distance between reconstructed signal and the received signal.Simulation results show the novel algorithm could obtain about 2 dB performance gain over the existent SOVA-PSP algorithm.(5) A new algorithm for blind separation of single channel PCMA signal based on MCMC was proposed. The algorithm employs Gibbs sampling to get the set of important sampling sequence of source signal, and then calculates the posteriori LLR values of the set, thus avoiding ergodic processing of the whole sequence of the source signal and remarkably reduced the computational complexity of the algorithm. Furthermore, an improved MCMC algorithm was proposed, in which the convergence was speeded up by multi-symbol joint updating. Simulation results show that the new algorithm can get about 1dB performance gain over the existent PSP algorithm when the time delay difference between two signals is about 2/8T.