Research on automatic modulation recognition based on CNN

Automatic modulation recognition (AMR) is a technology that determines the modulation method of a received signal without prior knowledge of the signal. Traditional AMR methods often face issues such as high computational complexity and high requirements for channel conditions. In recent years, AMR based on deep learning (DL) has gained significant attention, showing remarkable progress. AMR based on DL can better extract signal features and achieve higher recognition accuracy compared with traditional AMR methods. Convolutional neural networks (CNNs), which have seen achievement in a variety of DL applications, are widely used in AMR as well. This article proposes a CNN algorithm that improves classification accuracy by introducing residual skip connections and parallel placement of multiple convolutional layers, and compares it with two existing CNN algorithms (CNN and MCNet). On the open-source dataset DeepSig RadioML2018.01a, this algorithm performed excellently. At the SNR of 18dB the classification accuracy of this model reaches 93.6%, far surpassing the 61.8% and 76.5% of the two existing models.