Abstract: To solve the problem of open-set recognition of signal modulation, the reconstruction and discrimination networks (RDN) model for one-dimensional signal data is proposed based on the generative adversarial networks. The model is composed of two networks which are used to separately reconstruct and discriminate the input signals. During the adversarial training procedure, the two networks fully learn the data distribution form of the known modulation signal. The model enables the output of the reconstruction network to not only present more useful information of the known modulation signal, but also distort the unknown modulation signal, thereby enhancing the ability of the network to discriminate the type of modulation of the input signals. The simulation results show that the model can realize the open-set recognition of signal modulation. When the signal-to-noise ratio is greater than 0 dB, the recognition rates of both the known modulation signal and the unknown modulation signal are greater than 93%.

Key words: modulation recognition, open-set recognition, generative adversarial networks, reconstruction and discrimination networks (RDN)