A Deep-Learning-Based Open Set Automatic Modulation Classification Method Using Multiple Domain Representations and Group Constraint

Automatic modulation classification (AMC) plays a pivotal role in the radio monitoring systems for Internet of Things (IoT) and spectrum management. Many contemporary deep-learning-based AMC methods overlook the effects of signal frequency offset and signal sampling rate jitter caused by Doppler effect and signal bandwidth estimate deviation, respectively. Moreover, these approaches struggle with recognizing unknown classes effectively. In this article, we introduce an innovative signal data augmentation strategy during the training of deep neural networks. The approach involves utilizing multiple time domain and frequency domain signal representations as inputs to the neural networks. Additionally, it incorporates a group classifier and group constraint mechanism to enhance the unknown class recognition ability of deep neural network. The study investigates two prominent neural network architectures: 1) the convolutional neural network (CNN) and 2) attention mechanism-based transformer. Based on the experiment results, it indicates that: 1) data augmentation and multiple domain representations improve classification accuracy when frequency offset and sampling rate jitter existence; 2) the convergence speed of the transformer architecture-based neural network is faster than CNN-based neural network, but the former is easier to overfitting; and 3) the discrimination ability of unknown class was improved obviously when the neural network uses group classifier and training with group constraint. Experimental results also demonstrate that the proposed methods enhanced the ability of blind signal modulation recognition in radio monitoring systems.