A Local-Global Feature Space Constraint Network for Automatic Modulation Open Set Recognition

Automatic Modulation Recognition (AMR) aims to identify modulation types using specific technologies, which plays a crucial role in wireless communications and spectrum monitoring. Recent years have witnessed the rapid advancement in AMR due to the powerful feature extraction capabilities of deep learning methods. However, when dealing with open set environments, existing methods lack the ability of joint optimization for balancing the empirical risk of known signals and the open space risk associated with unknown signals. To address this issue, we propose a local-global feature space constraint network (LGFSCN) to tackle the AMR task in open set environments. The proposed approach begins with a local-global feature extractor that can complementarily exploit the long-distance dependencies of signals and focus on local features in depth. Then, we propose a feature space constraint by integrating class center vectors into the network to enhance intra-class compactness and inter-class separation, thereby ensuring sufficient feature space for unknown modulation types. Finally, a novel class-specific adaptive threshold rejection mechanism is proposed to discriminate between known and unknown samples. Extensive experiments conducted on RML2016.10a, RML2016.10b, RML2018.10a, and HisarMod 2019.1 datasets showcase that the proposed LGFSCN method achieves stable and accurate open set AMR performance and outperforms other competitive methods.