Detection of Unknown Signals at Low SNRs

The detection of unknown communication signals in low signal-to-noise ratio environments has utility in many wireless communications and security applications, such as opportunistic spectrum sharing, signal localization, and exploitation in low probability of detection environments. However, existing detection methods rely on digital signal processing techniques that perform poorly at low energies, or machine learning techniques that rely on well-structured, offline training data with known signal labels sufficient for a model to learn the features of a signal. This is impractical for environments where labeled training data is limited or difficult to obtain, such as for unknown signals that have not been previously observed. To overcome these challenges, this paper introduces a novel feedback architecture for pseudo-label generation in an online learning paradigm to detect wireless signals without a priori signal knowledge or model pre-training. The methodology significantly outperforms digital signal processing-based techniques and is accurate to within 3 dB of models trained with known signal labels, without similar limitations. The iterative architecture exhibits generalized learning as new, unknown signals are introduced to its online detection, and its practical design and implementation make it ready for adoption in realistic scenarios.