Speclearn: Spectrum Learning in Shared Band Under Extreme Noise Conditions

Published in IEEE International Symposium on Dynamic Spectrum Access Networks, 2024

This paper focuses on the detection of radar signals within the shared spectrum, such as the Citizen Broadband Radio Service band, employing machine learning methodologies. The study investigates the influence of various types of synthetic noise on the spectrum, thereby affecting the accuracy of signal detection in shared regions. Our approach involves the utilization of a YOLOv5-based object detection method, where a trained model is generated using clean spectrograms as input. The trained model is then evaluated for their detection performance under diverse noise conditions. The analysis reveals that an extremely noisy environment leads to a detection failure of 0%, while low noise conditions remain tolerable without any noticeable performance degradation. This study provides valuable insights into the robustness of machine learning-based radar signal detection in real-world, noisy scenarios within shared spectrum environments.