Environment Classification and Deinterleaving using Siamese Networks and Few-Shot Learning

In the age of digital communications, radar receivers prove to be essential for applications involving classification such as air traffic control towers, defense systems, and navigation systems. Detecting Emitters within a Radar Environment presents hurdles to the System Designers such as accounting for interference and trying to classify multiple emitters when they are stacked. This paper presents a few-shot machine learning model that utilizes Siamese networks with classification. Given a relatively small dataset, the Siamese network's task is to find the difference between stacked pulses and normal pulse trains, as well as classify the pulse-descriptor words (PDWs), of the signals in the environment. The PDWs will characterize various aspects of the signal with help from a dynamic-thresholding deinterleaving algorithm. The data for this experiment are laboratory generated signals that are transmitted and received using MATLAB, the Zynq Ultrascale+ MPSoC ZCU104 FPGA board, and the AD-FMCOMMS2-EBZ RF module.