Deep Autoencoder for Anomaly Detection in Terminal Airspace Operations

Abstract

View Video Presentation: https://doi.org/10.2514/6.2021-2405.vid In recent years, the aviation industry has seen a large increase in the volume of operations. The maintenance and improvement of safety at acceptable levels is one of the most important concerns in civil aviation operations. Reactive methods to aviation safety improvement are being augmented with proactive and predictive approaches that leverage large amounts of routinely collected aviation data. Due to the increased availability of airborne sensor data and improvements in computing power, application of machine learning methods to various aviation safety problems for identifying, isolating, and reducing risk has gained momentum. Previous work in this domain has focused on identifying anomalies or abnormal operations as a first step towards identification of potentially risky situations using aircraft sensor data. However, most existing methods rely only on the aircraft data and do not take into consideration the environment and context in which it is operating. In this paper, a novel framework based on deep learning methods using autoencoders is proposed to identify anomalies in terminal airspace operations. Data from multiple sources (aircraft trajectory, weather, traffic/congestion) is fused and utilized in the model development process which has not been attempted in prior work. The framework is proposed with the central idea of using historical aircraft trajectory data fused with weather and traffic metrics to build an anomaly detection model to identify trajectories that deviation from the norm, given a specific context. The framework is demonstrated on six months of arriving flight data collected for San Francisco International Airport as a case study. The developed framework has the potential to aid air traffic controllers in identifying high risk situations from a holistic perspective and applying appropriate mitigation strategies.

Topics & Concepts