Least-squares spectrum estimation through a neural network-inverse predictor structure

A method for the spectral estimation of random processes composed of sinusoids in white noise is proposed, based on the least-squares solution of an overdetermined set of linear equations representing the relationship between the power spectrum and the autocorrelation of the process. As operation in real-time is assumed, a problem to be faced is the accurate estimation of the autocorrelation, using few samples of the process. This problem is solved by resorting to an inverse predictor. The proposed approach provides a powerful computational architecture, composed of a neural network and an inverse predictor, that is suited for VLSI fabrication. Numerical examples are presented to illustrate the performance of the proposed method.