OSMR: Open-Set Modulation Recognition Based on Information Enhancement

Modulation recognition is a vital component of communication systems. With the emergence of neural networks, many researches have used IQ signal as the input for various learning algorithms in hopes of improving the recognition per-formances. However, it has been neglected that IQ signal contains the information of modulation characteristics, source charac-teristics, and other influencing factors. Besides, the modulation characteristics are scattered across various dimensions where modulation occurs, such as frequency domain, phase domain, and amplitude domain. Therefore, the modulation characteristic information of the raw IQ signal is difficult to be fully extracted. To solve this problem, it is necessary to enhance the modulation characteristic information and effectively extract its specific features. In this paper, we propose an open-set modulation recognition (OSMR) framework, which exploits the joint enhancement of modulation characteristic information, residual network, spatial attention mechanism and openmax layer. Specifically, we enhance the modulation characteristic information by calculating instantaneous features of frequency, phase and amplitude, and utilize spatial attention mechanism to further extract the specific features. OSMR can improve the performance of inter-class and intra-class recognition. Simulation results of the proposed method demonstrate an open-set recognition accuracy of 81% on the RadioML2018.01A dataset.