Deep Learning-Based Convolutional Neural Networks for Improving Classification Performance of Categories in Quadrature Amplitude Modulation

Due to the increasingly widespread use of composite modulation signals in current communication applications, it is necessary to select the most suitable demodulation method for this type of composite modulation signal based on its specific type. However, traditional methods have low accuracy in classifying a variety of composite modulation signals. This article proposes a scheme based on deep learning to assist signal classification tasks, hoping to improve the accuracy of constellation classification and reduce the difficulty of selecting modulation methods for different signals through this approach. More specifically, this article uses classical neural network models to explore the accuracy of different models in constellation classification tasks, and focuses on the best performing model, MobileNet, to investigate the impact of different image resolutions on the final results under this model. This article utilizes basic communication knowledge to generate a large number of Quadrature Amplitude Modulation (QAM) constellation diagrams of different types and signal-to-noise ratios and uses these images to train and learn classical neural network models. The final results indicate that Mobilenet with size of 128×128 performs the best in handling constellation classification tasks and can provide reference value for more complex classification scenarios in the future.