Adaptive Frequency-Domain Filter Shrinkage Network for Automatic Modulation Recognition

With the explosion of modern wireless communication information and the need for high transmission rates, automatic modulation recognition (AMR) based on deep learning (DL) has become increasingly popular. However, DL-based AMR methods often have lower accuracy at high noise environments. In order to solve this problem, we propose a adaptive network based on frequency-domain filter shrinkage (FDFS) module which contains two blocks for frequency noise reduction effect on the feature transmission of signals. The network can efficiently fuse the time and frequency domain features of signals, enhancing the capture of effective information by the model, and greatly improve the identification accuracy. Experiments with on two existing well-known datasets RML2016.10a and RML2016.10b demonstrate that the model performs significantly in modulation identification.