Open AccessMachine Learning classification techniques applied to static air traffic conflict detection
Author(s) -
Javier A. Pérez–Castán,
Luis Pérez-Sanz,
J. Bowen-Varela,
Lidia Serrano-Mira,
Tomislav Radišić,
Thomas Feuerle
Publication year - 2022
Publication title -
iop conference series. materials science and engineering
Language(s) - English
Resource type - Journals
eISSN - 1757-899X
pISSN - 1757-8981
DOI - 10.1088/1757-899x/1226/1/012019
Subject(s) - separation (statistics) , recall rate , air traffic control , computer science , precision and recall , artificial intelligence , simulation , data mining , machine learning , engineering , aerospace engineering
This article evaluates Machine Learning (ML) classification techniques applied to air-traffic conflict detection. The methodology develops a static approach in which the conflict prediction is performed when an aircraft pierces into the airspace. Conflict detection does not evaluate separation infringements but a Situation of Interest (SI). An aircraft pair constitutes a SI when it is expected to get with a horizontal separation between both aircraft closer than 10 Nautical Miles (NM) and a vertical separation closer than 1000 feet (ft). Therefore, the ML predictor classifies aircraft pairs between SI or No SI pairs. Air traffic information is extracted from The OpenSky Network that provides ADS-B trajectories. ADS-B trajectories do not offer enough SI samples to be evaluated. Hence, the authors performed simulations varying the entry time of the trajectories to the airspace within the same time period. The methodology was applied to a portion of Switzerland airspace, and simulations reached a 5% rate of SI samples. Cost-sensitive techniques were used to handle the strong imbalance of the database. Two experiments were performed: the Pure model considered the whole database, and the Hybrid model considered aircraft pairs that intersect horizontally closer than 20 NM and vertically lower than 1000 ft. The Hybrid model provided the best results achieving 72% recall, representing the success rate of Missed alerts and 82% accuracy, which means the whole predictions’ success rate.