Classification of Manifold Learning Based Flight Fingerprints of UAVs in Air Traffic

Abstract

As the number of UAVs (Unmanned Aerial Vehicles) and the market size have been expanding rapidly in recent years, projects such as NextGen and SESAR aim to include UAVs in air traffic. Therefore, different perspectives on understanding flight patterns can contribute to more effective management of future air traffic. Analysis of flight data offers an important insight into the operations of a UAV. In this study, it is aimed to extract a flight fingerprint using different machine learning techniques by means of a public dataset and the data obtained from our experimental flights. To get the individual flight pattern, multidimensional UAV sensor data has been reduced using manifold learning methods. By comparison, the most proper manifold method that allows highest classification accuracy (CA) has been investigated. Their performances are compared using both different manifold types and different classification methods. Then, the obtained manifold is used as flight fingerprints and validated by classification techniques. Various unsupervised manifold learning techniques such as t-Distributed Stochastic Neighbor Embedding (t-SNE), Locally Linear Embedding (LLE), Isometric Feature Mapping (ISOMAP) were tried for dimension reduction. For flight fingerprint classification, supervised machine learning techniques such as k-Nearest Neighbors (k-NN), Adaboost, Neural Network, Bayes, etc., were tested. It has been observed that the highest classification accuracy is achieved with the t-SNE manifold and k-NN classification pair. The extracted fingerprint can find many application areas such as performance tests in production lines, air traffic control, risk analysis, anomaly detection, observing pilot performance, drone efficiency over time.