RF-Based Drone Classification Under Complex Electromagnetic Environments Using Deep Learning

Recent studies have demonstrated that using deep-learning (DL) methods to classify drones based on the radio-frequency (RF) signal is effective. As known, the rich and diverse data is an important guarantee for good identification performance. In reality, due to the complexity and high dynamic of wireless environments, the costs of signal collection and labeling with sufficient diversity are often unacceptable. In this work, we propose a low-cost data augmentation (DA) method to improve the robustness of the signal identification neural network (NN). It generates extra training data by mixing pure drone signals with practical background signals and trains the NN to distinguish drone signals from interferences. Moreover, to further improve performance, we design a spectrogram segmentation (SS) method that directly splits the entire spectrogram into several subspectrograms to separate the interference signals working outside of the drone bandwidth. Finally, using five representative drone signals collected in multiple scenarios, we examine the identification accuracy and generalization capability of the proposed algorithm. Experimental results reveal that the proposed algorithm performs better than the NN trained without SS or DA.