Self-Supervised Contrastive Learning for Extracting Radar Word in the Hierarchical Model of Multifunction Radar

The analysis of intercepted multifunction radar (MFR) signals has attracted considerable attention in the field of cognitive electronic reconnaissance. The agility of pulse parameters makes it difficult to recognize their behavior states. Currently, most of the existing approaches for recognizing MFR behaviors heavily depend on prior information, which can hardly be obtained in a noncooperative way. This study develops a contrastive self-supervised method to adaptively segment and cluster MFR pulse sequences into radar words with little prior information. First, a convolutional neural network (CNN) is trained in a self-supervised manner to differentiate between adjacent and nonadjacent pulse group embeddings to learn robust feature representations from sequences of pulse descriptor words (PDWs). Next, based on the learned embeddings, test statistics are calculated for each PDW sequence to detect change points, which are, in turn, used to segment the sequence into multiple subsegments, each containing only one radar word. Finally, feature vectors are produced for clustering the subsegments into groups of radar words. Simulation results show that without using any labeled data, the proposed method can effectively extract the radar words of a hypothetical MFR under corrupted and overlapped pulse parameters and performs only slightly worse than the state-of-the-art fully supervised method.