Modulation Pattern Recognition Based on TR-Unet

• Published in: IEEE International Conference on Power, Intelligent Computing and Systems (Online) pp. 1081 - 1087
• Main Author: Chai, Yanzhe
• Format: Conference Proceeding
• Language: English
• Published: IEEE 26.07.2024
• Subjects: Accuracy; AMR; Computational modeling; Feature extraction; Modulation; Pattern recognition; ResNet; RML2016.10b; Signal to noise ratio; Testing; Training; Transformer; Transformers; Unet; Visualization
• ISSN: 2834-8567
• DOI: 10.1109/ICPICS62053.2024.10796271

Under the complex electromagnetic environment, Automatic Modulation Recognition (AMR) needs to improve the recognition accuracy. To address such problems, we design a Transformer ResNet-Unet (TR-Unet) model, which combines the Transformer and the Unet model with identity shortcut connections. Identity shortcut connections are from the residual network (ResNet). The original in-phase and quadrature (IQ) signals are input directly into the network. We design multi-layer encoder and decoder to extract features and introduce identity shortcut connection to preserve feature details. We utilize the Transformer instead of the convolution in the Unet model. This is to provide more comprehensive feature information extraction capability for the model and to reduce misjudgment and omission of the modulated signal. The training process is conducted with RML2016.10b, while the testing process is conducted with RML2016.10a. The results of simulation experiments show the classification accuracy of our proposed model can reach 92% when the signal-to-noise Ratio (SNR) is 12dB. This proves the effectiveness and reliability of TR-Unet.