Detección de anomalías en trayectorias de vuelo utilizando autoencoders y segmentación del espacio aéreo basada en regiones de Voronoi

Abstract

Given the increasing global air traffic, this article compares two autoencoder approaches for anomaly detection in flight trajectories, using the DBSCAN algorithm as an initial reference. The first model utilizes normalized continuous features (latitude, longitude, speed, and heading), while the second incorporates a discrete segmentation of the airspace through Voronoi regions, alongside kinematic variables. The results indicate on average 96% accuracy for the continuous autoencoder and 97% for the Voronoi-based model, with the latter showing a greater ability to identify normal trajectories. Qualitative analysis revealed that autoencoders, by including additional variables, capture more complex anomalies than DBSCAN. The integration of Voronoi regions improved the model's explainability, facilitating the interpretation of anomalies within their geographic context.