MILAFormer: A Multi-Stage Hybrid Deep Learning Architecture for Robust Radio Signal Modulation Recognition
IEEE Transactions on Cognitive Communications and Networking · 2026 · 12 · 7574-7588
DOI: 10.1109/tccn.2026.3683207 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Cognitive Communications and Networking, 2026. Verfasst von Jun Zhao und 6 weiteren Autor:innen. Die Arbeit wurde bisher 3-mal zitiert (Crossref) und nennt 53 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Zhejiang Provincial Natural Science Foundation of China, Fundamental Research Funds for the Provincial Universities of Zhejiang.
Abstract (Originalsprache)
Automatic Modulation Recognition (AMR) plays a crucial role in non-cooperative wireless communication systems, driving the advancement of future wireless networks. Recently, deep learning–based AMR (DL-AMR) techniques have achieved remarkable progress, showing superior accuracy and adaptability. However, under low signal-to-noise ratio (SNR) conditions, existing networks still face challenges in fully exploiting the diverse features of radio signals and achieving robust performance. To address these limitations, we propose a hybrid deep learning framework named MILAFormer, which integrates a Mamba-inspired linear attention (MILA) mechanism with a Transformer module. First, MILAFormer utilizes a data preprocessing module to strengthen the feature representation of the input signals. Subsequently, a Feature Enhancement Embedding (FEE) module is designed to improve signal quality and optimize feature extraction. Furthermore, a hierarchical Transformer module is introduced to learn local and global features from embedded tokens through convolution and a Mamba-inspired linear attention (MILA) mechanism. These components together facilitate the full exploitation of the diverse characteristics of radio signals and contribute to robust overall performance. Experimental results on RadioML2016.10a, RML22 and HisarMod2019.1 datasets demonstrate that the proposed MILAFormer model significantly outperforms existing state-of-the-art DL-AMR methods in radio signal AMR. The code implementation of MILAFormer is available at https://github.com/stu-cjlu-sp/rsrc-for-pub/tree/main/MILAFormer
Förderung
- National Natural Science Foundation of China
- Zhejiang Provincial Natural Science Foundation of China
- Fundamental Research Funds for the Provincial Universities of Zhejiang
Zitieren
APA
Zhao, J., Sun, Y., Chen, Y., Dong, X., Song, G., Jin, N., & Quan, D. (2026). MILAFormer: A Multi-Stage Hybrid Deep Learning Architecture for Robust Radio Signal Modulation Recognition. IEEE Transactions on Cognitive Communications and Networking, 12, 7574-7588. https://doi.org/10.1109/tccn.2026.3683207
Vancouver
Zhao J, Sun Y, Chen Y, Dong X, Song G, Jin N, et al. MILAFormer: A Multi-Stage Hybrid Deep Learning Architecture for Robust Radio Signal Modulation Recognition. IEEE Transactions on Cognitive Communications and Networking. 2026;12:7574-7588. doi:10.1109/tccn.2026.3683207
BibTeX
@article{Zhao2026,
author = {Zhao, Jun and Sun, Yulin and Chen, Yi and Dong, Xudong and Song, Guangxiao and Jin, Ning and Quan, Daying},
title = {MILAFormer: A Multi-Stage Hybrid Deep Learning Architecture for Robust Radio Signal Modulation Recognition},
journal = {IEEE Transactions on Cognitive Communications and Networking},
year = {2026},
volume = {12},
pages = {7574--7588},
doi = {10.1109/tccn.2026.3683207}
}
EUVIMED