TY - JOUR
T1 - Efficient Blind System Identification of Non-Gaussian Auto-Regressive Models with HMM Modeling of the Excitation
AU - Li, Chunjian
AU - Andersen, Søren Vang
PY - 2007
Y1 - 2007
N2 - We propose two blind system identification methods that exploit the underlying dynamics of non-Gaussian signals. The two signal models to be identified are: an Auto-Regressive (AR) model driven by a discrete-state Hidden Markov process, and the same model whose output is perturbed by white Gaussian noise. For both models, exact EM algorithms are derived for the joint estimation of all system parameters. The exact EM algorithms are obtainable only by appropriate constraints in the model design, and have better convergence properties than algorithms employing generalized EM algorithm or empirical iterative schemes. The proposed methods also enjoy good data efficiency since only second order statistics is involved in the computation. When measurement noise is present, a novel Switching Kalman Smoother is incorporated into the EM algorithm, obtaining optimum nonlinear MMSE estimates of the system outputs. The signal models are general and suitable to numerous important signals, such as speech signals and base-band communication signals. Applications to speech analysis and blind channel equalization are given to exemplify the efficiency of the new methods.
AB - We propose two blind system identification methods that exploit the underlying dynamics of non-Gaussian signals. The two signal models to be identified are: an Auto-Regressive (AR) model driven by a discrete-state Hidden Markov process, and the same model whose output is perturbed by white Gaussian noise. For both models, exact EM algorithms are derived for the joint estimation of all system parameters. The exact EM algorithms are obtainable only by appropriate constraints in the model design, and have better convergence properties than algorithms employing generalized EM algorithm or empirical iterative schemes. The proposed methods also enjoy good data efficiency since only second order statistics is involved in the computation. When measurement noise is present, a novel Switching Kalman Smoother is incorporated into the EM algorithm, obtaining optimum nonlinear MMSE estimates of the system outputs. The signal models are general and suitable to numerous important signals, such as speech signals and base-band communication signals. Applications to speech analysis and blind channel equalization are given to exemplify the efficiency of the new methods.
U2 - 10.1109/TSP.2007.893935
DO - 10.1109/TSP.2007.893935
M3 - Journal article
SN - 1053-587X
VL - 55
SP - 2432
EP - 2445
JO - IEEE Transactions on Signal Processing
JF - IEEE Transactions on Signal Processing
IS - 6 (1)
ER -