TY  - JOUR
T1  - Multi-Modal Automatic Modulation Recognition Network Based on the Swin Transformer Architecture
AU  - Shin, Da-Min 
AU  - Jeon, Min-Wook 
AU  - Kim, Hyoung-Nam 
JO  - The Journal of Korean Institute of Communications and Information Sciences
PY  - 2025
DA  - 2025/1/1
DO  - 10.7840/kics.2025.50.10.1524
KW  - automatic modulation recognition
KW  - deep learning
KW  - 1D Swin Transformer
AB  - In modern electronic warfare, deep learning-based automatic modulation recognition (AMR) systems that provide high recognition accuracy across a wide range of communication signals have been actively studied.
This trend stems from the growing complexity of signal environments in modern battlefields, resulting from the rapid evolution of communication technologies, which has significantly increased the difficulty of analyzing received signals. In this paper, we propose a multi-modal AMR neural network designed based on the 1D Swin Transformer architecture. The 1D Swin Transformer effectively extracts multi-resolution features from input sequences through a hierarchical feature representation and a shifted window-based mechanism. The proposed model employs a multi-modal structure that simultaneously utilizes IQ time-series signals in the time domain and spectrum information in the frequency domain. By extracting features from each modality and feeding them into the 1D Swin Transformer, the model performs modulation recognition by leveraging information from both the time and frequency domains. The proposed model achieved an average accuracy of 61.15% and outperformed conventional models. Furthermore, within the QAM family, the model achieved a high average recognition accuracy of 57.00% and demonstrated its ability to distinguish between structurally similar modulation schemes.