Magnitude-Phase Interaction Network for Wireless Interference Identification in Low-Altitude Networks

Abstract

Wireless spectrum awareness constitutes a critical research domain in wireless communications, particularly for the rapidly expanding low-altitude economy. Robust interference signal identification serves as a fundamental technique for ensuring the security of uncrewed aerial vehicles (UAVs). Leveraging the end-to-end learning framework of deep learning (DL), DL-assisted wireless interference sensing techniques have flourished, demonstrating significant potential and gradually replacing traditional feature extraction-based recognition methods. Current DL-based interference identification typically preprocesses signals using time-frequency transforms as network inputs, partially utilizing the network’s feature extraction while neglecting potentially superior signal representations. Therefore, this paper proposes a joint approach combining time-frequency transformation with two-dimensional fast Fourier transform (2D-FFT) for interference signal processing. The 2D-FFT operation enhances the time-frequency representation by revealing more discriminative features in the interference patterns, thereby significantly improving signal distinguishability. Furthermore, our methodology extends to embedding 2D-FFT operations directly into the network architecture. By applying 2D-FFT to intermediate features, we obtain their spectral representations and establish synergistic interactions between the derived magnitude-phase components and original spatial features. This integrated framework, named the magnitude-phase interaction network, enables comprehensive feature analysis in both spatial and frequency domains. Finally, we propose parallel stacked network, which has a frozen trained network and an adaptive network operate synergistically. The trained network’s fixed features provide a stable foundation that enriches the adaptive network’s learning space, while the adaptive network’s trainable parameters allow for task-specific refinement, resulting in enhanced feature extraction capabilities. The comprehensive simulation results demonstrate the effectiveness and superior performance of the proposed methodology, ensuring enhanced spectrum awareness for critical low-altitude conditions.