Georgia Institute of Technology
A Data-driven Methodology for Aircraft Trajectory Analysis to Improve Mid-air Conflict Detection in Terminal Airspace
Abstract
dc:description.abstractA mid-air collision occurs when two aircraft come into contact while airborne. It stands as one of the most devastating accidents in aviation history and remains a pressing safety issue in current flight operations. Presently, Air Traffic Control (ATC) serves as the primary means to ensure safe separation between aircraft and prevent mid-air collisions. This heavily relies on human operators, Air Traffic Controllers (ATCOs), who manage critical tasks under significant workload. The anticipated expansion of aviation, both in terms of traffic volume and diversity, particularly in terminal airspace, presents substantial challenges to the existing ATC system. The workload on ATCOs may exceed their capacity, potentially compromising safety. To address forthcoming aviation demands and maintain high safety standards, ATC is gradually integrating automated systems to assist ATCOs in transitioning from manual to supervisory roles. This dissertation is driven by the need for advanced analytics and automated decision supports concerning air traffic within terminal airspace. By leveraging Global Navigation Satellite System (GNSS) technologies, specifically Automatic Dependent Surveillance–Broadcast (ADS-B), ATC is able to access real-time and extensive historical operational data. Hence, this research presents a novel data-driven methodology to conduct thorough aircraft trajectory analysis, aiming to improve mid-air conflict detection within terminal airspace. The outlined methodology comprises three key steps: (1) traffic flow identification and recognition, (2) trajectory prediction, and (3) conflict detection. The traffic flow identification and recognition step entails two key requirements: (1) an effective method to identify air traffic flows in terminal airspace, and (2) a fast and accurate method to recognize the air traffic flow of individual flights. Achieving the first requirement demands a clustering approach capable of filtering out non-nominal trajectories commonly encountered in daily operations. While Density-Based Spatial Clustering of Applications with Noise (DBSCAN) may be applied, it can struggle with density variations in traffic flows observed in historical trajectories. Thus, Ordering Points to Identify the Clustering Structure (OPTICS) is proposed as an alternative clustering algorithm. Additionally, Weighted Euclidean Distance is suggested as a distance metric to account for the significance of different trajectory points. An experiment is designed to implement the OPTICS and DBSCAN algorithms using Weighted Euclidean Distance as the distance metric to identify air traffic flows in terminal airspace, and the results have demonstrated OPTIC's superior effectiveness in enhancing identification over DBSCAN. Addressing the second requirement involves employing a method capable of multi-class classification with rapid training and high accuracy. Ensemble models such as Random Forest and Extreme Gradient Boosting (XGBoost) provide a favorable balance between accuracy and efficiency, rendering them viable choices. Conversely, the Long Short-Term Memory (LSTM) model is anticipated to yield even higher accuracy, albeit with longer training time. An experiment is designed to implement Random Forest, XGBoost, and LSTM models for multi-class classification of aircraft trajectory segments, aiming to recognize air traffic flows of individual flights. Subsequently, their performance in terms of accuracy and training time is compared. The results of the experiment indicate that Random Forest achieves accuracy levels comparable to LSTM while significantly reducing training times. The trajectory prediction step necessitates a method for aircraft trajectory prediction. Existing methods typically employ an encoder-decoder architecture with LSTM trained on entire trajectory sets, leading to potential challenges: (1) difficulty in effectively learning hidden features due to significant differences in input trajectories, and (2) sequential nature of LSTM resulting in prolonged training durations. To overcome the first challenge, this study proposes to train multiple predictors on subsets with distinct traffic flows identified earlier, rather than a monolithic predictor on the entire dataset. An experiment is devised to implement the encoder-decoder architecture with LSTM to train a monolithic predictor and multiple predictors, on datasets containing all trajectories and subsets with distinct traffic flows respectively, then compare the accuracy and training time of the two approaches. The results have revealed that employing multiple predictors leads to increased accuracy and decreased training time compared to the single predictor approach. To address the second challenge, Transformer is proposed as an alternative to LSTM, benefiting from attention mechanisms to eliminate sequential operations and enable parallelization. An experiment is designed to train trajectory predictors for distinct traffic flows using the encoder-decoder architecture, first with LSTM and then with Transformer, followed by a comparison of the prediction accuracy and training time between the two approaches. The implementation results indicate a considerable reduction in training time and comparable accuracy achieved by Transformer compared to LSTM, particularly for extended prediction horizons. The conflict detection step requires an automated method to identify conflicts within terminal airspace, with a critical focus on addressing uncertainty. Utilizing historical trajectory data is crucial, especially in the context of aircraft position estimation, which conventionally relies solely on mathematical tools without leveraging real-world data. Kernel Density Estimation (KDE), a statistical technique for deriving Probability Density Functions (PDFs) from sampled data, emerges as a promising tool to enable robust estimation of aircraft positions based on historical trajectories. Furthermore, the intersection of PDFs from different flights serves as a means to identify potential conflicts. Hence, a novel Weighted KDE method is proposed to estimate aircraft positions by integrating outputs from traffic recognition and trajectory prediction, subsequently facilitating conflict detection in terminal airspace through the intersection of flight PDFs. To validate the proposed method, an experiment is designed to implement Weighted KDE to synthesize the outcomes of traffic flow recognition and trajectory prediction to estimate aircraft positions and then perform conflict detection by representing conflict with the intersection of aircraft position PDFs. The implementation results reveal that the conflict probabilities calculated by the Weighted KDE method show an inverse relationship with actual distances between aircraft, in both horizontal and vertical planes, thereby demonstrating the effectiveness the proposed conflict detection method. Several representative real flight scenarios serve as use cases to showcase the efficacy of the proposed data-driven methodology for analyzing aircraft trajectories to improve mid-air conflict detection in terminal airspace. The exploratory nature of this research suggests its potential evolution into a real-time decision support tool that offers conflict detection advisories for ATCOs. Transitioning from research to practical application may require real flight tests and incorporation of Real-time Assurance (RTA) mechanisms.
Degree
thesis:*- Level thesis:degree_level
- Doctoral
- Department dc:contributor.department
- Aerospace Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Zhang, Wenxin
- Advisor dc:contributor.advisor
-
- Mavris, Dimitri N.
- Committee members dc:contributor.committeemember
-
- Schrage, Daniel
- Chau, Polo
- Vamvoudakis, Kyriakos
- Puranik, Tejas
- Payan, Alexia
Subjects
dc:subject × 7Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/75259