Power spectrum estimation via neural-type structured network

We emphasise the learning algorithm and the convergence capability of the structured network for solving linear equations (Wang et al., 1990). Based on this structured network, a new autoregressive (AR) modeling method is presented. Its basic idea is to solve Yule-Walker type matrix equations for model coefficients by the structured network. The advantages of this AR modeling method over other AR modeling methods are: a parallel architecture and algorithm, suitable for VLSI hardware realization; and no divisions are involved in the calculations, so that the method still works for ill-conditioned Yule-Walker type matrix equations. Simulation results illustrating the performance of the method are given for both narrow-band sources and combinations of narrow-band and broad-band sources subjected to various levels of Gaussian white noise.<>