Deep Contrastive Clustering for Signal Deinterleaving
Shuyuan Yang, Xinyi Zhao, Huiling Liu, Chen Yang, Tongqing Peng, Rundong Li, Feng Zhang
Abstract
In a complex electromagnetic environment, radar signal deinterleaving (RSD) is a challenging task. In this article, a deep contrastive clustering algorithm (DCCA) is advanced in a new self-supervised paradigm for the accurate RSD without any prior information about radar emitters. First, a contrastive self-supervised deep attention network (CSDAN) is constructed to learn signal representations by using self-defined pseudolabels of augmented signals as supervision. We use CSDAN to learn the differences between different radiation source data and generate deep features suitable for clustering. Three metrics are then used to automatically determine the number of clusters for the subsequent clustering. Extensive experiments are performed on several datasets containing different numbers of emitters. The results show that the proposed DCCA can accurately determine the number of emitters and deinterleave radar pulses. Furthermore, CSDAN can extract discriminative features of emitters with low intraclass similarity and high interclass similarity.