A Signal-Understanding Semi-Supervised Learning Framework for Signal Recognition

Abstract

Signal recognition plays a crucial role in wireless communications, with artificial neural network models being widely applied, and the success of these models largely depends on abundant labeled data. However, practical signal recognition scenarios often face a shortage of labeled samples and an abundance of unlabeled ones. Therefore, semi-supervised learning (SSL) methods have emerged as a solution. This letter proposes a novel signal-understanding semi-supervised learning (SUSSL) framework to enhance the performance of SSL further. SUSSL comprises a reconstruction and a metric module. The former module learns useful features by disrupting and reconstructing low-level features, and the latter utilizes similarity learning to process low-level features. A symmetric dual-branch neural network (SDNN) model is also designed to facilitate these two modules. Simulation experiments on the open-source datasets RadioML2016.10a and RadioML2016.10b demonstrate that the proposed method outperforms current SSL methods.