A Novel Approach for Robust Automatic Modulation Recognition Based on Reversible Column Networks

Abstract

Automatic Modulation Recognition (AMR) technology, as a key component of intelligent wireless communication, has significant military and civilian value, and there is an urgent need to research relevant algorithms to quickly and effectively identify the modulation type of signals. However, existing models often suffer from issues such as neglecting the correlation between IQ components of signals, poor feature extraction capability, and difficulty in achieving an effective balance between detection performance and computational resource utilization. To address these issues, this article proposes an automatic modulation classification method based on convolutional neural networks (CNNs)—OD_SERCNET. To prevent feature loss or useful features from being compressed, a reversible column network (REVCOL) is used as the backbone network to ensure that the overall information remains unchanged when features are decoupled. At the same time, a novel IQ channel fusion network is designed to preprocess the input signal, fully exploring the correlation between IQ components of the same signal and improving the network’s feature extraction ability. In addition, to improve the network’s ability to capture global information, we have improved the original reversible fusion module by introducing an effective attention mechanism. Finally, the effectiveness of this method is validated using various datasets, and the simulation results show that the average accuracy of OD_SRCNET improves by 1–10% compared to other SOTA models, and we explore the optimal number of subnetworks, achieving a better balance between accuracy and computational resource utilization.