PCSP: A Novel Class Discovery Algorithm for Radio Signal Classification

Novel class discovery (NCD) aims to cluster novel signals in an unlabeled set by using prior knowledge of the labeled signal sets. In order to improve the recognition ability of radio signals, in this paper, we propose a prototype classifier and similarity prediction method (PCSP) for NCD of radio signals. Firstly, focusing on the different recognition abilities of the network for labeled and unlabeled signals, PCSP adopts different similarity prediction strategies and comparative learning, so that the network can obtain a better cluster representation. Secondly, in view of the limitation of the closed set assumption of the traditional deep learning network on the expression ability of unlabeled data, PCSP uses prototype encoding for each type of signal to obtain a prototype classifier, so that the network has a more compact intra-class feature representation. Finally, simulation experiments are carried out on the interference signal dataset and the modulation signal dataset respectively. The results show that the performance of PCSP is much better than the existing algorithms.