Unmanned Aerial Vehicle Flight Mode Classification using Convolutional Neural Network and Transfer Learning
Carolyn J. Swinney, John Woods
Abstract
Unmanned Aerial Vehicles (UAVs) are changing the way major industries conduct business in a globally friendly and efficient manner. Along with these economic benefits come major security challenges. Gatwick Airport is one example that cost £50 million. 1000 flights were cancelled over a 36hr period in 2018. Radio Frequency (RF) fingerprinting is an approach for UAV detection and classification associated with longer detection distances. Classification of a UAV flight mode would provide police commanders with information to assist risk assessment. For example, intelligence operations could be associated with a UAV flying while transmitting video. In this paper we introduce RF fingerprinting using Power Spectral Density (PSD) and spectrograms. This work utilises the open DroneRF dataset, made up of signals from the Parrot Bebop, Parrot AR and DJI Phantom 3 and further broken down into flight modes including switched on; hovering; flying with and without video transmission. Signal representations for each class are treated as images and transfer learning is implemented through the extraction of features using a VGG16 Convolutional Neural Network (CNN) pre-trained on the ImageNet database. Machine learning classifiers Logistic Regression (LR), Support Vector Machine and Random Forest are evaluated in these experiments for 2 classes (UAV present or not), 4 classes (UAV type) and 10 classifications (UAV flight modes). We show that PSD representation has higher accuracy than the use of time domain spectrograms, producing 100% accuracy for UAV detection and 88.6% accuracy for UAV type classification with LR. For 10 classification including flight modes, LR with PSD produced 87.3% accuracy which is a 40% increase on prior work in the field by A1-Sa'd et al.. Overall an approach has been introduced using transfer learning through CNN feature extraction and machine learning classification which performs with high accuracy compared with previous work on the same DroneRF dataset.