Enhancing air traffic conflict resolution through machine learning, conformal automation, and flow-centric paradigms

• Main Author: Guleria, Yash
• Other Authors: Sameer Alam
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Air traffic conflict resolution; ATCO-automation conformance; Machine learning; Supervised learning; Reinforcement learning; Flow centric operations; Air traffic controllers
• Online Access: https://hdl.handle.net/10356/177541
• Institution: Nanyang Technological University

Air traffic conflict resolution is a dynamic, time-sensitive, and safety-critical aspect of air traffic control, which involves a complex interaction of humans, machines, and procedures. In current sector-based operations where the airspace is subdivided into smaller geographical regions, ensuring safe separation between the aircraft by resolving potential conflicts and maintaining an efficient flow of traffic is the primary responsibility of air traffic controllers (ATCOs). The task being safety-critical, demands high cognitive effort. Researchers have proposed several computational and learning-based methods for air traffic conflict resolution. However, the translation of such methods to operations and their acceptance by ATCOs remains a challenge because of a mismatch between how the ATCOs perceive the conflicts and resolve them, and the resolutions proposed by the computational methods. In other words, the ATCOs’ conflict resolution maneuver preferences may differ from the optimized solutions generated by such computational and learning-based methods. In this regard, behavior cloning is an established methodology to encapsulate the preferences of a human in performing complex tasks. Moreover, sector based operations also present inherent scalability constraints that hinder meeting future demands. New concepts of operations such as flow-centric operations (FCOs), which involve managing air traffic from an aggregated perspective instead of individual flights (as in sector-based operations) based on the formation and evolution of major traffic flows, are being proposed to accommodate future traffic. This poses an additional challenge in ensuring safe separation between air traffic flows in flow-centric operations. Thus, there are two-fold challenges in meeting future traffic demands for sustained growth while ensuring safe operations. First, the development of methods that incorporate ATCOs’ conflict resolution preferences, and second, the development of novel conflict resolution techniques to accommodate future air traffic management concepts. Research approaches in the development of methods for air traffic conflict resolution have transitioned from mathematical models and optimization approaches to learning based methods. This shift has been triggered to better address the limitations of mathematical models in accommodating uncertainties and stochasticity associated with the environment, and generalizability to non-nominal scenarios with non-standard model inputs. Nonetheless, the proposed approaches have the following fundamental limitations. Firstly, air traffic conflict resolution is a sequential decision-making problem, where the ATCOs must make a series of sequential decisions to ensure safe separation between aircraft. Such sequential decisions are a result of the inherent preferences or strategies that the ATCOs develop over time and constitute an end-to-end conflict resolution maneuver of the aircraft. The machine learning methods proposed in the literature do not generate an end-to-end conflict resolution maneuver but rather, only the initial vectoring segment. ATCOs’ conflict resolution strategies are also not captured by such approaches. In this regard, behavior cloning is an established methodology to encapsulate the preferences of a human in performing complex tasks. Furthermore, in the design of such methods, researchers have employed simplistic simulation environments, which fail to emulate the complexity of an end-to-end conflict resolution maneuver. Secondly, the existing literature lacks analysis and discussion on how current conflict resolution methods can be adapted to function in flow-centric operations. In this context, the thesis addresses two key research questions. First, can the ATCOs’ conflict resolution strategies be learned through behavior cloning and incorporated into a learning-based model? This research question encompasses developing a methodology to identify the ATCOs’ conflict resolution strategies, and then, developing suitable machine learning models that can learn these strategies. Second, how can air traffic conflicts be resolved in a flow-centric paradigm where traffic is modeled as intersecting flows? The approaches proposed to address these research questions have resulted in three research contributions. The first contribution is based on developing a methodology to identify the conflict resolution strategies of the ATCOs. To investigate the ATCOs’ conflict resolution strategies and the factors affecting them, high-quality data that is representative of the ATCOs’ behaviors is required. Therefore, a conflict resolution data collection experiment was designed with the objective of collecting ATCOs’ conflict resolution maneuvers for the corresponding conflict scenarios. In this experiment, conflict scenarios were simulated in a high-fidelity simulation environment that emulates an air traffic control radar interface. Eight ATCOs were involved in the experiments. The participants were shown real-time air traffic conflicts in the simulation environment representing a sector of the Singapore flight information region and their interactions with the simulation environment to resolve the conflicts were recorded. Individual analysis and comparison of each ATCO’s data demonstrated that ATCOs have distinct conflict resolution strategies, specifically in terms of maneuver direction and the magnitude of the heading deviations. The metrics used to identify these strategies were the selection of the aircraft to be maneuvered, maneuver initiation time, maneuver direction, magnitude of deviations, and preferred safe separations between the conflicting aircraft. It was also observed that the ATCOs’ conflict resolution strategies were influenced by factors such as the proximity of the conflict point from the sector boundary and which aircraft arrived first at the conflict point. The second contribution developed machine learning-based ATCO conformal conflict resolution models using behavior cloning. Here, ‘conformal’ implies resolutions that are similar to the ATCOs’ conflict resolution strategies. This work formulated conflict resolution as a sequential decision-making task and used a sequence of regressor and classifier-supervised machine learning models to map the environment states to low-level actions and clone the behavior of the ATCOs. Based on the ATCOs’ strategies identified previously, personalized (matching individual ATCO’s preference) and group conformal (matching the preference of a group of ATCOs) models were developed using data collected during the conflict resolution experiments with the ATCOs. The prediction results demonstrated the proposed models are able to generate ATCO conformal predictions on the test dataset. The trained models were also tested for added Gaussian noise to the data, to evaluate the models’ robustness. The results demonstrated that the models are robust up to 7.5% added noise, with low mean absolute errors and high accuracy for the predictions (for example, the classification accuracy was > 92.7% for predicting the maneuvering aircraft, MAE for maneuver initiation distance was < 5.3 NM and MAE for predicting the heading angle was < 5.3° for the prediction models). The work was further evaluated by human-in-loop experiments involving professional ATCOs, to evaluate their choices in the selection of their own conflict resolution strategies, and the conflict resolution strategies obtained from the prediction models. The results demonstrated that the ATCOs selected solutions matching their conflict resolution strategies for over 70% of the scenario, which reinstates that conformal conflict resolution advisories receive greater acceptance by the ATCOs. The third contribution focused on the emerging concepts of operations involving the flow-centric paradigm and the development of inter-flow (between the flows) and intraflow ( within each flow) conflict resolution models for flow-centric operations. Crossing conflict scenarios under uncertainty in a flow-centric paradigm were investigated in this work. The research was modeled as a sequential decision-making problem using the Markov Decision Process (MDP). This choice was made to align with the inherent sequential decision-making nature of the conflict resolution process. Each flow, which consists of a varying number of aircraft with different cruise speeds and the associated location uncertainties, was modeled as a self-stabilizing graph structure. The two-stage conflict resolution process involved training a conflict resolution policy to ensure inter-flow safe separation along with the use of a self-stabilizing graph structure to ensure intra-flow safe separation. The trained policy can ensure both intra and inter-flow safe separation between the aircraft for 100% of the scenarios. The performance was further analyzed in terms of maneuver efficiency and deviations from the flight plans. Inspite of the uncertainties and dynamics associated with the flows’ size, speed, and evolution over time, the absolute delays for the flow were 2.53 and 9.49 minutes respectively. In its entirety, this thesis addresses both immediate and future strategies for tackling the challenge of escalating air traffic by proposing innovative approaches for safe separation of air traffic. It presents methods for the creation of learning-based conformal models by cloning the behavior of ATCOs and the formulation of a conflict resolution model within a flow-centric airspace paradigm, which holds significance for the evolving concepts of operations.