A Transformer-Based Self-Supervised Learning Framework for Robust Time-Frequency Localization in Concurrent Cognitive Scenario

Time-frequency localization (TFL) based intelligent wideband spectrum sensing is capable of achieving precise dynamic spectrum management. Recent studies demonstrate that object detectors can achieve excellent TFL performance in simple electromagnetic scenarios when trained with massive and labeled datasets. However, in real-world concurrent cognitive scenarios that allow users to reuse the same frequency band under a certain interference constraint, the phenomenon of signal overlapping in the time-frequency domain will seriously degrade the performance of object detector. To the best of our knowledge, no comprehensive analysis has been conducted to assess the impact of overlapping in TFL. To fill this research gap, we analyze the impact of overlapping and identify three challenges: variety of overlapping, hard to label, and feature destruction. To enhance the robustness of the detector, we first adopt a self-supervised learning (SSL) framework based on a masked autoencoder. This framework aims to pre-train a backbone with excellent feature extraction ability using unlabeled dataset to overcome variety of overlapping and labeling difficulties. Subsequently, we develop a transformer based robust TFL (TRTFL) detector. This detector is designed to leverage both time-frequency correlation and fine-grained features, effectively addressing issues related to feature destruction. Finally, simulation results demonstrate the superiority of the proposed method and the effectiveness of SSL framework and TRTFL. Compared to existing detectors, the TRTFL achieves superior feature extraction, yielding a mean average precision (mAP) of 90.70% in overlapping signal scenarios. Moreover, the TRTFL with SSL can achieve an mAP of up to 95.08% outperforming the state-of-the-art.