Radio Frequency Signal Recognition of Unmanned Aerial Vehicle Based on Complex-Valued Convolutional Neural Network

The rapid development of the low-altitude economy is driving significant societal and industrial transformation. Unmanned aerial vehicles (UAVs), as key enablers of this emerging domain, offer substantial benefits in many applications. However, their unauthorized or malicious use poses serious security, safety, and privacy risks, underscoring the critical need for reliable UAV detection technologies. Among existing approaches, such as radar, acoustic, and vision-based methods, radio frequency (RF)-based UAV detection has gained prominence due to its long detection range, robustness to lighting and weather conditions, and capability to identify RF-emitting UAVs even when visually obscured. Nevertheless, conventional RF-based approaches often suffer from limited feature representation and poor generalization. In the past few years, convolutional neural networks (CNNs) have become the mainstream solution for RF signal recognition. However, most real-valued CNNs (RV-CNNs) process only the magnitude component of RF signals, discarding the phase information that carries valuable discriminative characteristics, which may degrade recognition performance. To address this limitation, this paper proposes a complex-valued CNN (CV-CNN) for UAV RF signal recognition, which exploits the full complex-domain structure of RF signals to enhance recognition accuracy and robustness. The proposed CV-CNN accounts for both the magnitude and phase components of RF signals from UAVs, thereby enabling true complex-valued convolutional operations without loss of phase information. The effectiveness of this approach is validated on the DroneRFa dataset, which encompasses RF signals from 25 distinct UAV categories. The impact of model hyperparameters, including network depth, convolutional kernel size, and dropout strategy on recognition performance is investigated through a series of ablation experiments. Comparisons are also conducted between the performance of CV-CNN with identical parameters and RV-CNN, both in noise-free and noisy conditions. The experimental results demonstrate that the CV-CNN exhibits superior robustness and interference resistance in comparison to its real-valued counterpart, maintaining high recognition accuracy even under low signal-to-noise ratio (SNR) conditions.