Effects of diagram plane on neural network based modulation recognition

Abstract

Modulation recognition using deep learning presents challenges in effectively distinguishing high-order modulation schemes while maintaining a balance between complexity and recognition accuracy. In this study, we curate a comprehensive dataset in the rθ plane, encompassing eight distinct modulation schemes. Leveraging hyperparameter optimization and transfer learning, we explore the capabilities of various CNN-based architectures, including MobileNetV2, ResNet50V2, ResNet101V2, InceptionV3, ResNet152V2, Xception, and InceptionResNetV2, for the classification of modulation schemes. The simulation results demonstrate that with signal-to-noise ratio (SNR) values exceeding 5 dB, all models exhibit classification accuracies surpassing 50% and approach near-perfect accuracy at an SNR value of 20 dB. However, under low SNR conditions, such as 5 dB, the recognition accuracies of all models, except for ResNet152V2 and InceptionV3, show minimal variation. As the SNR increases by 5 dB from −5 dB to 20 dB, ResNet152V2 and InceptionV3 demonstrate remarkable classification accuracy improvements, exceeding 40%, 30%, 30%, 10%, and 15%, respectively. In contrast, the other models do not exhibit such robust responsiveness in accuracy enhancements. The remarkable performance improvements are achieved by fine-tuning pre-trained models through these processes.