CFCS: A Robust and Efficient Collaboration Framework for Automatic Modulation Recognition

Most of the existing automatic modulation recognition (AMR) studies focus on optimizing the network structure to improve performance, without fully considering cooperation among the basic networks to play their respective advantages. In this paper, we propose a robust and efficient collaboration framework based on the combination scheme (CFCS). This scheme effectively explores the spatial and temporal characteristics of complex signals by associating the advantages of convo- lutional neural network (CNN) and long and short-term memory (LSTM) network. In addition, the robustness of the CFCS is verified by transfer learning. Experiments demonstrate that the recognition rate of CFCS for highorder modulation signals such as 64QAM, 128QAM, and 256QAM is more than 90% at high signal-to-noise ratios (SNRs), and 24 modulation types are effectively identified. Moreover, CFCS was transferred from RML2018.01a to RML2016.10b using transfer learning, which can still be deployed efficiently while reducing the training time by 20%. The CFCS has strong generalization ability and excellent recognition performance.