Generative Adversarial Network Enables Incremental Learning for Shortwave Signal Open-Set Recognition Through Prototype-Based Algorithms

Signal recognition is one of the key technologies for spectrum monitoring, signal demodulation, and signal decoding in shortwave communication systems. In recent years, deep learning has made significant progress in the field of wireless communication signal classification. However, open-set recognition(OSR) schemes aimed at identifying known signal types while also recognizing unknown signals have not been thoroughly studied. Therefore, this paper proposes a novel Generative Adversarial Network with Prototype Learning (GANPL) framework to improve the OSR performance of shortwave signals. First, leveraging knowledge from the wireless signal domain, a series of signal-related features are extracted to generate a time-frequency diagram dataset of shortwave signals. Second, the basic prototype algorithm achieves certain results in simultaneously identifying known and unknown class. Finally, by introducing a Generative Adversarial Network (GAN) to generate samples that are similar to known samples but belong to unknown class, these samples are integrated into the basic prototype algorithm for joint training, enabling the CNN to understand unknown class. The results show that the proposed Generative Adversarial Network can improve the performance of the basic prototype in shortwave signal openset recognition.