Rigorous Moment-Based Automatic Modulation Classification

Abstract

In this paper we develop the connection between the high-order moments, orthogonal polynomials, and probability densities representing signal constellations with AWGN in order to improve moment-based Automatic Modulation Classifi- cation (AMC). The result is that an approximate weighted L 2 distance between probability densities can be computed using a Euclidean distance on vectors consisting of series expansion coefficients. This analysis justifies the use of high-order moments in AMC. A discriminative Deep Neural Network (DNN) is trained to perform AMC, resulting in near-maximum likelihood performance at marginal SNR.