Drone classification and improvement of elevation angle classification rate of drone using multi-polarization characteristics using convolutional neural network

Abstract

With the recent rapid development of the drone industry, positive aspects and side effects are occurring. In particular, with the commercialization of personal drones, technological advances are being made for the regulation of drones, and as one of them, drone detection technology through radar is being studied. The distinction between the two is important because drones have similar radar reflection areas to birds. The need to further expand and identify information such as the type and angle of drones is also emerging. Compared to birds, the biggest feature of the drone is that it transmits and receives radio waves for control, and the other is that the propeller works finely and quickly. In this paper, it was possible to check the existence of a drone by detecting a leak signal from a GPS system (GPS) generated from a drone, and the type of drone was identified using a convolutional neural network by extracting a fine Doppler signal from the received signal. The leakage wave of GPS generates a lot of noise due to its low power, and the identification rate of drones is improved by removing the noise using the Bivariate Empirical Mode Decomposition (BEMD) algorithm. In addition, the elevation angle of a single drone was identified. The drone's propeller was an experiment based on research showing that different micro-Dopplers occur depending on the angle. The drone's elevation angle was measured at 10-degree intervals and applied to Visual Geometry Group-11 (VGG11), one of the types of convolutional neural networks, to record an average identification rate of 84.7%. Additionally, focusing on the fact that the characteristics of the fine Doppler generated from the propeller of a drone change according to the polarization, which is physically independent information, four pieces of information were simultaneously received through a receiver having four different polarizations. The received four polarization information was combined with each other and applied to the same model, VGG11, recording a high identification rate of 97.9%.