Wireless Signal Recognition Based on ResNet-Transformer
Jianyong Chen, Rui Zhao, Chengyu Yang, Yaoyao Dai · 2023
In the face of the rapid development of wireless technology in the Internet of Things (IoT), the communication capacity that wireless spectrum can carry is increasing. This has led to wireless signal recognition becoming a key technology for intelligent spectrum management. This paper proposes a deep learning model for wireless signal recognition, called Hybrid net. This model consists of a residual block that extracts local information from the sample and a Transformer block that extracts global information from the sample. These two blocks are added together through a residual unit to form a more complete representation of the feature space. It also avoids the problem of high model complexity and long training time caused by using data in frequency domain format. In order to better simulate actual communication environments, in-phase and quadrature (IQ) and IQ with random phase offset datasets were constructed for training and testing. Experimental results show that the proposed Hybrid net can achieve recognition accuracy of over 80% at low signal to noise ratios (SNRs grater than −6 dB). The Hybrid net even reach 100% at 4 dB. In addition, Hybrid net also demonstrated good generalization performance on different datasets.