Data-Driven Inference of ATCO Separation Intent Using Flight Plans, Radar Trajectories and Neural Networks

Air Traffic Control Officers (ATCOs) are responsible for controlling air traffic and ensuring the safety of the aircraft. Capacity, understood as the maximum number of aircraft that can be safely managed for one hour, is calculated based on the workload of ATCOs. This calculation normally is based on a manual and tedious data collection process that demands a high consumption of human resources. To improve and relieve human re-sources, automation tools that automatically generate a preliminary annotation of Air Traffic Control (ATC) activity have been developed. This paper focuses on the feasibility of employing data-driven approaches using neural networks to classify ATC events, as well as if it is possible to improve the performance of these ATC-activity tools. Particularly, this approach seeks to infer ATC intent for separation actions, which are the most critical in terms of ATC workload. A modular methodology has been developed to include information from different sources: flight plans, radar trajectories, trajectory prediction, conflict detection and rule-based knowledge. Different experiments are evaluated based on the different input’s combination, as well as three neural networks (Multilayer Perceptron, Convolutional Neural Network and TabNet). Results show that TabNet is the best neural network option, reaching a similar performance in task classification than current ATC tools and improving classification metrics around 4% by employing the outputs of ATC tool metrics as inputs.