Multi-Representation Signal Fusion for Robust RF-Based UAV Recognition via Structure-Guided Time–Frequency Modeling

Abstract

RF-based UAV recognition is important for low-altitude security because it enables passive sensing from electromagnetic emissions without requiring target visibility or illumination. Compared with active sensing, passive RF sensing offers a wider observation range and is less constrained by line-of-sight conditions. Although deep-learning methods have improved RF-based UAV recognition, most existing approaches focus primarily on label prediction and do not explicitly model time–frequency structural characteristics, which limits their robustness in practical RF environments where UAV signals are often weak and may overlap with co-channel interference. This paper presents a multi-representation signal fusion (MRSF) framework for robust RF-based UAV recognition through structure-guided time–frequency modeling. MRSF represents each RF signal using an original time–frequency map, a residual-enhanced map, and a binarized structural map to capture global spectral patterns, enhanced signal structures, and connected spectral morphology. These representations are fused through a heterogeneous three-branch stem for backbone classification, and the binarized structural map further supports connected-region analysis to estimate occupied bandwidth, temporal duration, and signal period in non-cooperative scenarios where protocol-level parameters are unavailable. On the RFUAV dataset, MRSF achieves an average accuracy of 96.05% across target SNR levels from −20 to 20 dB and maintains 88.29% accuracy at −20 dB, demonstrating strong robustness under severe low-SNR conditions. These results show that, for UAV RF emissions acquired within the effective coverage of the front-end receiver, MRSF improves robust UAV recognition while providing interpretable spectrum-structure descriptors beyond class labels.