LAUSR

Dynamic spectrum access based wireless networks and next-generation cognitive electronic warfare systems demand rapid identification and labelling of high data rate radio frequency (RF) information. This requires receiver front-end designs to distinguish numerous kinds of wireless signals of different standards over a relatively wide spectrum. This paper proposes a novel attempt at large scale, blind identification of signals from 29 wireless standard technologies that occupy the modern day spectrum. A deep convolutional neural network model called ‘Wireless Standard Technology Identifier (WiST ID)’ is deployed, along with a pre-processing method based on the Stockwell transform time-frequency representation for highly accurate classification over relatively large number of signal classes. The model demonstrates enhanced learning of RF fingerprints from the pre-processed Stockwell images belonging to a variety of wireless technologies. Analyses of classification performance over synthetically generated samples with SNR scenarios varying from −10 dB to 10 dB reveal the robustness of the model under low and moderate SNR. At a modest SNR of 10 dB, the model achieves 100% classification accuracy over a small scale synthetic dataset (9 classes). Over a large scale dataset (29 classes) consisting of both synthetically generated and over-the-air captured samples, the classification accuracy achieved is 98.91%.