| As one of the key technologies of signal processing,single-channel blind source separation has broad application prospects in biomedical signal processing,communication signal processing,and underwater acoustic signal processing.It is a hot research topic in the field of signal processing.Since some traditional single channel blind source separation algorithms can't separate linear or nonlinear mixed signals with less a priori information,this thesis starts from the signal model,based on the signal with time series structure characteristics,using state space estimation theory and Bayesian statistical analysis theory.Focusing on Kalman filter,Extended Kalman Particle Filter(EPF),Unscented Kalman Particle Filter(UPF)and Recurrent Neural Network(RNN)methods to achieve single-channel blind source Separation.The specific research work is as follows:(1)Single channel blind source separation algorithm based on Kalman filtering.Based on Kalman filtering and single-channel blind source separation theory,this thesis improves the Kalman filtering structure and proposes a single-channel blind source separation algorithm based on Kalman filtering.By constructing the observation equation and state equation of the system,the algorithm defines the initial iteration value,and uses the continuity of the discrete time points to estimate and iteratively update the signal state at each moment to obtain the optimal estimation and realize the single channel blind source separation.The simulation verification of the three-channel signal in Matlab environment shows that the Kalman filtering algorithm has better separation performance for the linear instantaneous mixed system.(2)Single-channel blind source separation based on EPF and UPF algorithm.The Kalman filter has a good separation effect on the mixed signal of the linear system,however,due to the limitation of the state equation,the separation performance of the mixed signal of the nonlinear system drops sharply.In order to solve the nonlinear system problem,this thesis deduces extended Kalman particle filter and unscented Kalman particle filter single-channel blind source separation algorithm by analyzing Kalman filter,particle filter and unscented transform characteristics.The use of low-order terms in the Taylor series expansion in the EPF algorithm improves the performance of importance sampling.The UPF algorithm uses the unscented transformation to control the nonlinear transfer of the mean and covariance of the system.The particle filter is used to obtain the optimal estimation of the multi-channel source signal,and then thereby the source signal is recovered and separated.The simulation experiments are carried out under the general nonlinear and strong nonlinear systems.The experimental results show that the EPF and UPF algorithms can have good separation effect on nonlinear mixed signals.(3)Research on single-channel blind source separation algorithm based on recurrent neural network.Since the EPF and UPF algorithms need to perform importance sampling and resampling for each discrete time,there are two problems during processing large amounts of sampled data that high computing complexity and long calculation time.Considering the neural network has the characteristics of self-learning and nonlinear input signal structure generalization,this thesis will apply it to the nonlinear single-channel blind source separation,optimizes the structure of traditional recurrent neural network by improving the self-recurrent of hidden layer into the recurrent from loss layer to hidden layer,thus reduces the computational complexity and makes the network structure more suitable for the solution of nonlinear hybrid single-channel blind source separation problem.In the Python environment,the simulation experiment will be carried out on the three-channel mixed signal.The experimental results show that this method can effectively solve the problem of blind source separation of the nonlinear mixed signal under the three-layer recurrent neural network. |
PDF Full Text Request