Identification of Radar Signals from the Spectrogram with Open-Set Deep Learning

This paper presents a novel approach for radar emitter identification, that uses the spectrogram of the full received signal. It does not rely on pulse descriptor words and is therefore also suitable for non-pulsed signals. A convolutional neural network is used for identification, that is able to efficiently classify 50 different radar emitters with an accuracy of 97%. In addition, two open-set techniques are investigated, that are able to detect "unknown"emitters, that have not been available for training. This is an important feature when classifiers are deployed in real-world applications.