Open-DS: An Open-Set Modulation Recognition Method for Low Probability of Interception Radar Emitters Based on Dictionary Similarity

Abstract

The recognition of radar emitters modulation in an open-set scenario presents a challenging task, particularly when identifying unknown modulation. This paper proposes a dictionary similarity based method for low intercept probability radar signal open-set modulation recognition (OMR), designed to address the unknown modulation in open-set scenarios. First, deep features of the input 1-D signal are extracted, and a random Fourier transform is applied to map the signal into a high-dimensional space, thereby converting the nonlinear feature optimisation problem into a linear optimisation problem. Next, an inter-class discreteness (ICD) module and an intra-class similarity (ICS) module are designed. Based on the Hilbert-Smith independence criterion, the correlation between features is quantified, and the quantitative values of ICD and ICS are used as loss functions to constrain the network's learning process. This approach effectively enhanced the representational power of the class dictionaries and significantly improved the model’s overall performance. Experimental results demonstrate that the proposed strategy successfully extracts high-dimensional feature prototypes, achieving high accuracy in closed-set recognition while effectively performing open-set recognition tasks.