The increasing demand and complexity of air traffic management (ATM) systems necessitate significant advancements in automation to ensure safety and efficiency. Artificial intelligence (AI) and machine learning (ML) are emerging as promising solutions to manage this growing compl
...
The increasing demand and complexity of air traffic management (ATM) systems necessitate significant advancements in automation to ensure safety and efficiency. Artificial intelligence (AI) and machine learning (ML) are emerging as promising solutions to manage this growing complexity, offering enhanced decision-making and predictive capabilities. However, the effectiveness of ML models in ATM heavily relies on the availability of extensive, high-quality data. In many cases, such data is scarce or incomplete, which presents a major challenge for training robust models. Synthetic data generation (SDG) represents a viable solution to address this data scarcity, enabling the creation of realistic, diverse datasets that can effectively fill data gaps and unlock the ML value proposition. The Terminal Maneuvering Area (TMA) is a crucial segment of airspace where aircraft perform complex maneuvers including rare go-arounds, and holding patterns. The TMA is characterized by high traffic density and diverse trajectory types, necessitating precise, granular data to model these dynamic scenarios accurately. The main research objective of this thesis project was to investigate the applicability of Generative Adversarial Networks (GANs) in generating synthetic 4-dimensional aircraft landing trajectories capable of capturing traffic patterns in this airspace. An unconditional recurrent-based network (TimeGAN) was applied and complemented with an investigation into the merit of conditioning to reap the rewards of transfer learning between classes. The resulting synthetic trajectories were evaluated in terms of data diversity, fidelity and usefulness. The main challenge identified during this research was the imbalance in data classes, which affected the models' ability to accurately capture data patterns, particularly in less frequent scenarios. Generating synthetic data based on separate groupings showed promise in addressing these imbalances, although this approach was sensitive to the number of clusters chosen. Smoothing filters frequently proved necessary to ensure realistic landings were devoid of unwanted noise. Overall, this work proves the capability of TimeGAN in generating diverse, realistic trajectories that are difficult to differentiate from real historical data.
