LAUSR

Different from a conventional artificial neural network (ANN) typically being used as a data-driven classifier, the “quasi-neural network” (Quasi-NN), with its inputs, outputs, and activation functions being artfully defined to have explicit physical meaning, is recently proposed in the literature as a modeling tool. In this paper, we show that the classic problem of line spectral estimation can be modeled by a Quasi-NN with its weights being the frequencies and the coefficients of the sinusoids to estimate. Owing to the layered structure the same as an ANN, the weights of the Quasi-NN can be optimized using the celebrated back-propagation (BP) algorithm initialized by the fast Fourier transform (FFT) based spectral estimation. The optimized weights are the maximum likelihood (ML) estimation of the line spectrum parameters. We also design a method of merging and pruning the hidden-layer nodes of the Quasi-NN for model-order selection, i.e., to estimate the number of sinusoids. Numerical simulations verify the effectiveness of the proposed method.