Joint Semantics Embedding and Self-Supervised Denoising Contrastive Learning for Multifunction Radar Work Modes Recognition

The pulse sequences-based work mode recognition of multifunction radars (WMR-MFRs) is crucial for the electronic support measures (ESMs) to analyze the intentions of radar platform and carry out the jamming countermeasures. However, the measurement errors, missing and spurious pulses, and lack of prior information can still pose a significant challenge. This article proposes a WMR-MFRs method through joint radar semantics embedding with selfsupervised denoising contrastive learning. First, the radar functional-level semantics is embedded into the input parameter sequences via radar experts domain knowledge. Then, self-supervised denoising contrastive learning is utilized to effectively extract distinctive features between different pulse groups and similar features within the intrapulse groups. More specifically, denoising contrastive loss and Gaussian noise regularization loss are designed to reduce the sensitivity to the undesirable noise and outliers during training process. Finally, a feature difference-based label mapping (FDLM) method is proposed to improve the recognition robustness. Simulation results show that the proposed method can achieve over 96% accuracy for the WMR-MFRs task in the case of interclass parameters overlap, and more than 90% accuracy in the scenario with 40% nonideal pulses. Compared to the baseline methods, the proposed method can achieve better and more robust performances in WMR-MFRs task.