Underdetermined Blind Source Separation Algorithm Research

Currently, Blind Source Separation (BSS) is a new research direction in signalprocessing. Without the need to know the source signals and the channel characteristics,the source signals can be estimated effectively. This method has broad applicationprospects in communication, speech signal processing, and biomedical signalprocessing, and so on. Under-determined Blind Source Separation (UBSS) is achallenging and hot topic in BSS, this dissertation conducts a study and discussion onthis issue.Based on the above background, the first chapter describes the researchbackground and significance of the subject, and the second chapter gives a study anddiscussion of the classic BSS algorithms and common UBSS algorithms.Considering so many signals in real world are band-limited, a newunder-determined BSS method is proposed in the third chapter using frequencydecomposition method by complementary filters to estimate mixing matrix and recoversource signals effectively. The sensing signals are divided into different sub-bands, andeach sub-band uses common BSS algorithm. We can reconstruct the source signals bycombing the related signals together. Even if the number of source signals is muchlarger than that of sensing signals, this method could also recover the source signalswhen the number of source signals are not more than that of sensing signals in eachsub-band. Finally, computer simulation confirms the validity of this method.A classical UBSS algorithm has been improved in the forth chapter. Two-stepstrategy is the basic method of under-determined blind source separation currently, theestimation of mixing matrix is a prerequisite for recovering the source signals. Weconsider the situation that the mixing matrix has been estimated, matrix subspacemethod is applied to estimate source signals. In order to improve anti-noise capability ofthe underdetermined blind source separation algorithm, the relevant characteristics ofadjacent sampling points of signals is used to determine the ownership of the signals atany time. Finally, simulations demonstrate the performance and validity of the proposedalgorithm. The fifth chapter considers the case of the unknown number of users in DS-CDMAsystem, independent component analysis (ICA) based blind multiuser detection (MUD)cannot effectively implement the principal component analysis (PCA) and determine thenature of the independent component. To solve this problem, this dissertation proposesan algorithm for blind MUD based on estimation of the number of source signals inseparated components. First, pre-PCA is applied based on the DS-CDMA systemcapacity, and then ICA processing is used. The ICA independent signals are determinedas useful signal or noise signal according to standard kurtosis (or negative entropy). Theestimation of the number of users is used to reset the retain number of paths of PCA.Additionally, the useful signals are sent to the hard decision to complete the usefulinformation extraction. Compared with the existed source number estimation methods,this algorithm can more accurately estimate the number of users in system anddetermine the nature of the separated signals. Finally, the simulation confirms the goodestimation performance of the proposed algorithm.The last chapter is the summation of this dissertation and future work isprospected.