Application of CNN deep learning model and CFAR filtering technique for RF-based drone signal classification in noisy conditions

Abstract

Currently, the use of small scale unmanned aerial vehicles, commonly known as drones, has increased due to the growing demand for remote interaction, contactless operations, and advanced technology. However, alongside the rising demand for drones across various sectors, their misuse for nefarious purposes has also escalated. Therefore, there is a need for drone monitoring systems to detect unauthorized drone usage. In this work, we propose a solution based on the radio frequency (RF) signature of drones. The data used in the research are control signals from 17 types of publicly disclosed drones. The proposed method addresses the issue of drone classification based on RF signatures; it maintains classification accuracy when the signal-to-noise ratio (SNR) decreases due to the presence of significant noise. In the experiments, we expanded the control signal data by adding white Gaussian noise to alter the SNR from −15 dB to 15 dB with 5 dB increments. Power spectral portraits with threshold values were applied to create training images for the convolutional neural network (CNN). The proposed model achieved 96% accuracy at an SNR of −15 dB and 99.82% accuracy in the high SNR region. From these results, we have confirmed that the proposed method not only has good classification capabilities but also high noise immunity.