A Novel Deep Learning Framework for Efficient Automatic Modulation Recognition of Sub-Nyquist Sampled Signals.

Automatic Modulation Recognition (AMR) is an exciting technology that empowers intelligent communication receivers to identify different signal modulation schemes. In recent times, deep learning (DL) research has led to significant advancements in high-performance DL-AMR techniques, particularly in the context of traditional Nyquist sampling. However, in practical applications, the cost, complexity, and power consumption of analog-to-digital converters (ADCs) pose significant challenges to the implementation of real-time wideband spectrum sensing. Therefore, a sub-Nyquist sampling mechanism has been proposed to solve the challenging problem of meeting the high Nyquist sampling rate required by 6G ultra-wideband technology. Nevertheless, sub-Nyquist sampling may lead to non-linear distortion of the original signal, which can greatly increase the difficulty of signal modulation recognition and result in a significant decrease in the recognition accuracy of traditional DL-based methods. In this paper, we propose a novel deep learning framework for sub-Nyquist signal modulation recognition, which consists of two parts: (1) signal reconstruction based on Simultaneous Orthogonal Matching Pursuit (SOMP), and (2) time-frequency feature-based modulation recognition network including shorttime discrete Fourier transform (STFT) and efficient convolutional neural networks (CNN), called STFT-AMCNet. The proposed approach's performance is thoroughly evaluated and compared with previous state-of-the-art methods on a publicly available dataset to showcase its exceptional efficiency.