Neural network approach to spectral estimation of harmonic processes

A neural network approach is presented for the spectral estimation of random processes composed of closely spaced sinusoids in white noise. A linear programming formulation is adopted, determining the minimum L/sub 1/-norm solution of a set of linear constraints. Then, the optimisation problem is solved by a dedicated electrical neural network whose input is the estimated autocorrelation of the process, and whose output is the power spectrum. The time response is very fast since the network is analogue and has parallel architecture. Moreover the lack of a learning phase makes it suited both to real-time signal processing and to VLSI implementation. Results of SPICE simulations are presented.< >