Deep Transfer Learning Method for Automatic Modulation Recognition
Wenlong Zeng, Hanmin Sheng, Xintao Xu, Xi Wang · 2024
In the field of non-cooperative automatic modulation recognition (AMR) for multi-device communications, acquiring a sufficient amount of labeled data is often challenging. This limitation leads to the failure of traditional machine learning and general deep learning methods due to their tendency to overfit. Recent studies have indicated that deep neural networks can learn transferable features, which can generalize well to new tasks in domain-adaptive settings. In this paper, we propose a Multi-layer Domain Adaptation Hybrid Network (MDAHN) tailored for modulation signal recognition scenarios. In MDAHN, the network's feature extraction layer combines Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) networks, enhancing the network's ability to learn modulation signal features. We adapt all task-specific layers using Maximum Mean Discrepancy (MMD), enabling MDAHN to learn transferable features with statistical guarantees. We validate the proposed method on public datasets and self-test datasets. Extensive experimental results demonstrate that the proposed network architecture significantly improves classification accuracy compared to state-of-the-art studies.