Study On Methods And Applications Of Blind Signal Processing In Stable Noise

Blind signal processing (BSP) has become a widely concemed subject in recent years and has shown an absorbing prospect in the fields of seismic exploration, mobile communication, aliasing speech separation, brain signal processing, space array signal processing, and so on. In practical applications, many signals and noises are not exactly of Gaussian distribution. Furthermore, spikes and impulses often accompany these signals and noises, resulting in their deviation from Gaussian distribution. Alpha-stable distribution was introduced to model the signals and noises of spikes and impulses. Based on the contribution of many people in the last decade, a new theory named blind signal processing comes into being.This dissertation concentrates on the study of blind signal processing with alpha stable distribution. The main work and conclusion of this dissertation are listed as follows:(1) This dissertation proposes a robust method for radial basis function neural network (RBFNN) beamforming technology based on fractional lower order statistics (FLOS) pre-processing. Beamforming technology is a key part in array signal processing, and the information of beamforming is contained in FLOS. As a result, FLOS makes beamforming process as a nonlinear mapping from input space to output space that approached by RBFNN. Because covariance is sensitive to additive impulsive noise, conventional MUSIC algorithms become degenerative to various degrees and the convergence rate is slow. In order to compress data, eliminate initial phase and avoid the effects of impulsive noise, here it takes advantage of FLOS pre-processing before RBF network. The simulation experiments show the proposed new algorithm fits the mapping process very well. The outputs of RBFNN have a good approximation to the generalized Wiener solution (GWS) and its direction-of-arrival (DOA) estimation matches the result of FLOM-MUSIC algorithm. So it represents the same robustness and high resolution as these classical algorithms. At the same time, the algorithm proposed in the dissertation has the advantages of fast convergence, small amount of the computation and robust nonlinear approach, so it is easy to come true in the project.(2) This dissertation proposes the unscented Kalman Filtering (UKF) algorithm based on RBFNN pre-processing. Kalman filtering method is a classical linear filtering algorithm and extended Kalman filtering (EKF) algorithm can solve the filtering problems about non-linear system. But, the method will bring large tnmcation errors in its linearization. In order to solve this problem, foreign scholars propose the theory of unscented Kalman filter. However, the variance matrix of computation error gradually loses positive definite and symmetry in the recursion, which induce output divergence of system filtering. This dissertation analyses the reason of divergence, and makes use of RBF network to adjust residual between predictive value of state vector and filtering estimated value, and gain matrix to minish output error and avoid output convergence. The simulation results demonstrate that the proposed method accomplishes tracking-target much better and is very promising in nonlinear filtering.(3) Referring to the traditional least squares (L2-norm) filtering algorithm, it loses its robustness in dealing with the issue of infinite variance lower-order stable distributed noise filtering. Therefore, a new least absolute deviation (L₁-norm) filtering algorithm is proposed. The adaptive learning of neural network can fulfill the minimum absolute deviation criteria, so this dissertation puts forward the algorithm that can be implemented by the neural network. Compared to traditional method, it has more advantageous convergence and better effect of separation when extracting EP signal from the EEG noise of non-Gauss distribution. Moreover, the independent component analysis (ICA) algorithm based on a minimum dispersion coefficient and rotation transformation is given, which can separate independent component sources from brain signals under low signal-noise-ratio (SNR). As the traditional ICA algorithm can not ascertain the amount of sources in observed signals, blind source separation algorithm based on the peak coherence measure detection is indicated and obtains good performance while it is applied to aliasing speech separation. Furthermore, based on innovation process with infinite variances in fractional poles system, this dissertation proposes a inverse filtering algorithm which can well rebuilt source signal under alpha-stable distribution noise, and analyze its convergences.