Reaserch On The Methods Of Blind Source Separation For Communication Signals

The extensive use of the radar, communication and other electronic equipmentsmakes the electromagnetic environment more and more complex, where large numbersof signals are overlapped in the frequency domain at any same time. Thus, theprobability of receiving mixing signals overlapping in the time frequency domain byvarious communication systems increases dramatically. In order to estimate theparameters and extract the information from the signals of interest, the correspondingsignals must be separated from the mixtures firstly. Blind source separation (BSS) is asignal processing technique that consists of retrieving n unobservable sources fromm linear combinations provided by sensors. The adaption ability of the communicationsystems can be greatly improved by utilizing the BSS methods, which consists of majorsignificance in practical field. This paper is aiming to investigate the following fourproblems according to two different factors (e.g., the reception structure of multiplesensors and the sparseness property of signals):1) the problems related to both the DOAestimation and BSS for sparse signals based on controlled sensor arrangement;2) theproblems related to both the DOA estimation and BSS for non-sparse signals based oncontrolled sensor arrangement;3) the problems of BSS based on random receptionstructure of multiple sensors;4) the problems of extraction of the specific target signals.The content of the dissertation is as follows.In chapter2, the problem related to both the DOA estimation and BSS, whichconsists in estimating the waveforms of signals and their physical directions ofpropagation and given a controlled sensor arrangement is investigated. A novelunderdetermined number of sources, DOA and waveforms of sources estimationalgorithm in the TF domain is proposed under the assumption that the number ofsources is less than that of sensors in each time frequency partition domain by utilizingthe concept and method of sparse signal representation. Firstly, partition the timefrequency support domain of the sources into disjointed time frequency neighborhoods,and estimate the DOA and the corresponding waveform of the active sources in eachtime frequency neighborhood domain by the sparse reconstruction method, and thenestimate the number of the source signals and the their DOA simultaneously by thecluster validation technique based on k-means clustering algorithm. Finally, the timefrequency transformation of each signal can be obtained by splicing the estimated resultin each time frequency partition domain while the permutation problem betweendifferent time frequency partition domain can be solved by the utilization of the DOAinformation and then the waveforms in the time domain of the source signals can beobtained by the inverse transform. The proposed algorithm can estimate the number of sources, the DOAs and waveforms of the source signals simultaneously in theunderdetermined case and the major advantages of the proposed algorithm are asfollows:1) Compared with the clustering-based DOA estimation algorithms, theproposed algorithm relaxes the necessary condition that the number of the signals is lessthan that of the sensors. Moreover, the number of the signals needs not to be known as apriori and can be obtained along with the DOA estimation process. Compared with thehigh-order cumulants based DOA estimation algorithms, the proposed algorithmrequires less number of samples and achieves higher DOA estimation accuracy.2) Theproposed algorithm provides better performance in separating source signals than thesubspace-based method which requires the same sparseness assumption as that in theproposed algorithm since it can estimate the number of active sources at any timefrequency partition domain. The simulation results show the efficacy and accuracy ofthe proposed algorithms.In chapter3, the problem related to both the DOA estimation and BSS, whichconsists in estimating the waveforms of signals and their physical directions ofpropagation and given a controlled sensor arrangement is investigated, where the sourcesignals are not sparse in time or time frequency domain. A new method fordirection-of-arrival (DOA) estimation of cyclostationary signals is proposed. Based on ajoint diagonalization of combined set of extended cyclic autocorrelation matrices, theproposed DOA estimation algorithm can address for the underdetermined case. Firstly,calculate cyclic correlation matrix of the mixtures corresponding to different cyclefrequencies and time lags, stack the cyclic correlation matrices with the same cyclefrequency and different time lags into the extended cyclic correlation matrix form byalgebraic method. Then estimate the signal subspaces and the noise subspaces by jointdiagonalize a set of extended cyclic correlation matrices with different cycle frequencies.Finally estimate the DOAs by the conventional subspace-like methods. The proposedDOA estimation algorithm does not require the source signals to be sparse in the timefrequency domain and can estimate the DOAs in the underdetermined case bygenerating a virtual of both the effective aperture and the number of sensors of theconsidered array through the algebraic method. Furthermore, it is insensitive to thespatially correlated noise and its estimation performance can be improved greatly byutilizing multiple extended cyclic autocorrelation matrices of source signals. After theestimation of DOAs, a new method for source separation based on ExpectationMaximization (EM) algorithm and optimum beamforming in the time frequency domainis proposed. The source signals in each time frequency neighborhood can be extractedby obtaining the corresponding optimum beamformers. The probability densityparameters of the sources and noise required in the calculation of the optimumbeamformers can be estimated by the expectation maximization method. Simulationresults confirm the validity and high performance of the proposed algorithm. In chapter4, the problem of underdetermined BSS in the case that the receptionstructure of multiple channels is random is investigated. A novel underdeterminedseparation algorithm based on single source point detection and improved matrixdiagonalization is proposed in the case of that there should be some single source pointsfor each source and the number of active sources at any time frequency neighborhooddoes not exceed that of sensors. Firstly, calculate the time frequency ratio matrix of themixtures and detect single source point of each source signal by hist stat. of the real partand imag part of the time frequency ratio matrix, then estimate the number of the sourcesignals and the mixing matrix simultaneously by clustering the mixing vectors in thecorresponding single source point set by utilizing the cluster validation technique basedon k-means clustering algorithm. The proposed mixing matrix estimation algorithm hastwo major advantanges:1) relaxes the sparsity requirement of the source signals and canestimate the mixing matrix under the assumption that there exist disjointed single sourcepoints for each source signal.2) user-parameters in the estimation process of number ofsources are easy to set. After the mixing matrix has been estimated, the source signalscan be obtained by the improved matrix diagonalization algorithm. For the single sourceneighborhood, the corresponding mixing vector can be obtained by the subspaceprojection method, and for the multiple source neighborhood where there are more thanone active sources, the corresponding mixing matrix can be obtained by measuring thedifferences between the diagonal elements and non-diagonal elements of the covariancematrix. Finally, the source signals in each time frequency neighborhood can beestimated by calculating the peuside inverse of the corresponding mixing matrix. Theproposed algorithm improves the accuracy over the matrix diagonalization basedalgorithm via estimating the number of active sources at any time frequencyneighborhood. Simulation results confirm the validity and high performance of theproposed algorithm.In chapter5, the problem of the specific signal extraction in theoverdetermined/determined case is investigated. The constrained independentcomponent analysis (CICA) framework under the linear instantaneous mixing model isgeneralized to the linear delayed mixing model, the types and the sources of theconstraints are summarized, and two novel solutions to the CICA framework areproposed. Based on the new CICA framework, two new CICA algorithms are proposedby utilizing different prior information of the communication signals. One is CICAalgorithm with cyclosationary constraint which exploits the cyclostationary property ofthe target signals as prior information. The convergence condition and the selectionrules of user parameters are analyzed. The other is CICA algorithm with spatialconstraint which exploits the spatial information corresponding to the Directions ofArrival (DOAs) of the SOIs as prior information. The convergence condition and theselection rules of user parameters are analyzed. The CICA algorithms incorporate the a priori information about the desired signal as the additional constraints into theconventional ICA learning process and means that only a single statisticallyindependent component will be extracted for the given constraint, which effectivelysolve the order ambiguities in conventional BSS problem. The correspondingexperiment results show the efficacy and accuracy of the proposed algorithms.