A Geometric Feature Space Learning Method for Automatic Modulation Open-Set Recognition

Abstract

Automatic modulation open-set recognition (AMOSR) aims to correctly classify known signals while reliably rejecting unknown signals. However, existing methods often suffer from insufficient geometric constraints in the feature space and inadequate modeling of unknown regions, leading to degraded classification accuracy for known signals and limited rejection capability for unknown signals. In this letter, we propose a geometric feature space learning method for AMOSR. Specifically, we introduce a multi-view expert representation module to enrich discriminative information for representation learning. Subsequently, a repulsive and attractive points collaborative learning module is developed to impose geometric constraints on intra-class and inter-class distributions while reserving open space for unknown signals. Furthermore, we design a pseudo-unknown manifold construction strategy to generate synthetic samples through manifold mixup to shape the unknown region without real unknown supervision. Experimental results on the RadioML 2016.10A dataset demonstrate that the proposed method outperforms representative baselines in several aspects.