Domain Adaptive Hybrid Network for Automatic Modulation Recognition

Abstract

Translate Article

Automatic modulation recognition (AMR) is an essential technology for spectrum sensing and management. Machine learning-based AMR techniques have garnered significant research interest. However, in practical applications, the effectiveness of AMR is hindered by variations in data distribution due to changes in carrier frequency, sample rate, and channel model. Existing deep learning methods struggle to adapt to these diverse distributions, and training separate datasets for each scenario is often impractical. To address these challenges, we propose a domain adaptive hybrid network (DAHNet) that integrates convolutional neural networks (CNNs) and long short-term memory (LSTM) networks. This architecture leverages the phase and amplitude of constellation points, as well as time-series data from I/Q sequences. Additionally, by explicitly minimizing the maximum mean discrepancy (MMD) between features from the source and target domains across multiple neural layers, we tackle the model generalization difficulties posed by varying propagation conditions and noise in complex electromagnetic environments. We validate our method using deep transfer learning on real-world data, with training results derived from the publicly available RadioML2016.10a and RadioML2018.01a datasets. Experimental results demonstrate the proposed method’s superior generalization ability across different spectrum scenarios.