Automatic Modulation Classification in Impulsive Noise: Log-Domain 3-D Constellation Diagrams and Multiscale Dual-Convolutional 3DCNN

Abstract

Automatic modulation classification (AMC) is vital in cognitive communication systems. Existing AMC methods are mainly designed for Gaussian noise channels, but research shows that non-Gaussian impulsive noise in wireless communication systems cannot be ignored. This paper proposes a novel AMC method based on Log-domain 3D constellation diagrams and multi-scale dual-convolutional 3DCNN (3D Convolutional Neural Networks). Firstly, we adopt non-linear logarithmic function transformation to effectively suppress non-Gaussian impulsive noise and transform multi-type modulated signals from the signal time domain to the graph domain, constructing Log-domain constellations. Secondly, the 2D constellations are projected onto three-dimensional space to form logarithmic domain 3D constellation diagrams, aiming to provide more discriminative feature dimensions for deep learning networks. Then, a new multi-scale dual-convolutional 3DCNN (MDC-3DCNN) is proposed, where the number and scale of dual-convolutional structures are the hyperparameters of MDC-3DCNN, to achieve feature extractions on both spatial and planar dimensions simultaneously for modulation classification. Numerical simulation results and actual measurement experiments demonstrate that the proposed method can effectively perform high-precision and robust classification of modulated signals in non-Gaussian impulsive noise channels.