Classification of Radar Pulses Modulation Laws via Convolutional Neural Networks

This article takes place in the context of Electronic Warfare Support (EWS). The intended objective is to establish a superheterodyne receiver in order to detect potential threats. Because a superheterodyne receiver cannot listen to a wide frequency bandwidth, we suggest establishing a scan strategy that focuses more frequently on unknown radar signals. For this purpose, we propose a method to classify the radar signals in order to know if we already received the signal of a given source. We focus on pulse feature extraction and more specifically on classification of the modulation law. The idea is to be robust to other parameters like pulse length or Signal to Noise Ratio (SNR). We select signal representations which conserve the time dimension like raw signal and time-frequency representations and apply Convolutional Neural Network (CNN) to extract automatically the important features. Based on simulations carried out, we finally conclude that both representations offer similar results for SNR between 0 and 5 dB. We also notice that phase modulated pulses are the most difficult to distinguish. They are confused with each other. The amplitude modulated pulses can also be confused with simple pulses. The chosen frequency modulated signals are the easiest to classify.