Distributed Automatic Modulation Classification with Compressed Data

Abstract

This work presents an approach for performing automatic modulation classification (AMC) in a distributed environment using a novel multi-input averaging Convolutional Neural Network (CNN) which ingests one instance of raw received data, in Inphase/Quadrature (IQ) format, and compressed realizations of the same signal from neighboring nodes. Further, this work examines the use of undercomplete autoencoders (AEs) as a means to compress raw received IQ data for transmission to neighboring nodes while retaining the signal features most pertinent to performing AMC. The accuracy of the developed approach is evaluated using simulated BPSK, QPSK, and 16QAM signals, with a noise-only class, and the impact of the compression ratio, number of nodes, and SNR are considered. While results show that the implemented AE is not an effective means of compressing raw IQ data, results did show that by combining data realizations from neighboring nodes using the proposed approach, classification accuracy increases by as much as 7% per node.