Abstract

Automatic modulation recognition (AMR) is an important signal classification technology in cognitive radio. As AMR advances, an increasing number of artificial neural networks are being employed in the field to enhance its performance. In order to further improve its performance, a ConTransNet model based on I/Q-language mutual learning and supervised learning is proposed in this work. First, a ConTransNet model is introduced to handle modulation signals. The model consists of two branches: one is CNN, and the other is transformer. To facilitate information exchange between the two branches, an information interaction module is introduced, implemented with a bridge connection. To enhance the model's performance, a training algorithm called I/Q-language mutual learning and supervised learning is designed. This method utilizes mutual supervision between the output of one branch of the ConTransNet model and the output of a language feature extraction model, while the other branch adopts supervised learning. Finally, through experimental comparisons with five other algorithms (CE-FuFormer, ConvLSTMAE, DAE, FEAT , and MCLDNN), the effectiveness of the proposed method is validated.