Neural-network-aided automatic modulation classification

Abstract

Automatic modulation classification (AMC) is a pattern matching problem which significantly impacts divers telecommunication systems, with significant applications in military and civilian contexts alike. Although its appearance in the literature is far from novel, recent developments in machine learning technologies have triggered an increased interest in this area of research. In the first part of this thesis, an AMC system is studied where, in addition to the typical point-to-point setup of one receiver and one transmitter, a second transmitter is also present, which is considered an interfering device. A convolutional neural network (CNN) is used for classification. In addition to studying the effect of interference strength, we propose a modification attempting to leverage some of the debilitating results of interference, and also study the effect of signal quantisation upon classification performance. Consequently, we assess a cooperative setting of AMC, namely one where the receiver features multiple antennas, and receives different versions of the same signal from the single-antenna transmitter. Through the combination of data from different antennas, it is evidenced that this cooperative approach leads to notable performance improvements over the established baseline. Finally, the cooperative scenario is expanded to a more complicated setting, where a realistic geographic distribution of four receiving nodes is modelled, and furthermore, the decision-making mechanism with regard to the identity of a signal resides in a fusion centre independent of the receivers, connected to them over finite-bandwidth backhaul links. In addition to the common concerns over classification accuracy and inference time, data reduction methods of various types (including “trained” lossy compression) are implemented with the objective of minimising the data load placed upon the backhaul links.