Research On Underdetermiend Blind Source Separation In Complex Electromagnetic Environment

The extensive use of the radar, communication and other eletronic equipments makes the electromagnetic environment in the war field more and more complex, where large numbers of signals are overlapped in the frequency domain at any same time. In this circumstance the received signal is always non-disjointed in time frequency (TF) domain. Furthermore, in practical field the number of sources is larger than that of sensors sometimes as a result of unknowing the number of the potential sources beforehand and that the number of the sensors is finite. To estimate the parameters and extract the transmitted information from the each interceped signal, the original signals must be separated from the mixtures firstly. The problem of the underdetermined signal blind separation in the complex electromagnetic environment is investigated in this paper. The underdetermined problem is transformed into over or well determined problem by exploiting the sparsity of sources in the TF domain, independence, cyclostationarity, periodicity and other statistic property, and then separate all the original signals. The main contributions are detailed as follows:In chapter 2, the underdertemined blind separation algorithm based on single source detection and subspace projection is proposed in the case of that there should be some single-source TF neighborhoods for each source and the number of active sources at any TF point is strictly less than the number of sensors. Fristly, detect the single source neighborhoods by eigenvalue decomposition and then estimate the mixing matrix via clustering the principal eigenvectors of the covariance matrixes corresponding to the single source neighborhoods without knowing the number of original sources. Extract the original signals by the modified subspace algorithm when the mixing matrix has been estimated. Because of the inequality between the assumed number and real number will lead to estimation performance degradation of source signals, to overcome the drawback this paper improve the based-subspace algorithm for extracting the original signals via estimating the number of active sources at any TF point. Simulation results verify the effectiveness of the proposed algorithm.In chapter 3, the novel underdertemined blind separation algorithm based on second-order cyclic correlation and covariance matrix diagonalization is proposed in the case of that the sources are independent and the number of active sources in any TF neighborhood does not exceed that of sensors.Fristly, calculate cyclic correlation matrix of the mixtures and stack the cyclic correlation matrices corresponding to different cycle frequencies and time lags into a three order tensor, achieve the estimation of mixing matrix by canonical decomposition. To exploit the cyclostationarity of the sources sufficiently, the proposed method improves the estimation accuracy when the cyclic frequency of the signals are different. Separate the source by measuring the diagonalization degree of the covariance matrix after the mixing matrix has been estimated. This method considers the independence of the sources sufficiently and relaxes the sparsity condition of sources in TF domain further, which allows the number of the active sources in any TF neighborhood simultaneously equals to that of sensors. Simulation results demonstrate that the proposed algorithm performs well.In chapter 4, the Underdetermined blind source separation algorithm based on TF distribution is proposed in the case of that the auto-source TF point and cross-source TF point is disjointed. Fristly calculate the spatial TF distribution matrix of the mixture and fold the TFD matrices into a third-order tensor corresponding to the auto-source TF points, then estimate the mixing matrix by tensor canonical decomposition or joint diagonalization. Secondly, transform the underdetermined problem into the over-determined case with knowing the mixing matrix. Finally we obtain the sources by calculating the pseudo-inverse matrix and TF synthesis techniques. It is not necessary that assume the sources are sparse in time domain or independent for the proposed algorithm. The simulations verify the capabilities of the proposed method.In chapter 5, the spreading sequences estimation of the synchronous and non synchronous DS-CDMA signal in the case of the single channel is investigated. For synchronous DS-CDMA, the estimation algorithm based on ICA and overlapped segmentation is proposed. Firstly, the baseband samples are obtained with chip period by carrier frequency synchronization and chip timing, and then divide the intact spreading sequences into many overlapped short-time segments so as to recover them in a sequential procedure. Estimate the spreading sequences segments of each user respectively by ICA and then solve the order permutation and amplitude ambiguity by the correlation of the overlap. The intact spreading sequences are estimated by splicing the segments and the closed-form expression that characterizes the estimation performance of the proposed algorithm in theory is obtained. For non-synchronous DS-CDMA, the baseband spreading wave with frequency offset of each user is estimated by the method based on Complex-ICA and overlapped segmentation, and then extract the spreading sequences of each user by carrier frequency synchronization and chip timing. Simulation results verify the effectiveness of the proposed algorithms.