In the heart of the complex and rapidly evolving world of air transportation, pioneering research continues to illuminate new paths towards improved operations, heightened security, and increasingly accurate predictions. The latest special issue of peer-reviewed journal Aerospace—entitled “Advances in Air Traffic and Airspace Control and Management”—features several interesting research papers that each tackle a unique facet of the aviation industry’s most pressing challenges. From an adaptive tracking solution for maneuvering aerial targets to a state-of-the-art model for entity recognition in cyber threat detection, and an innovative method for real-time flight arrival prediction, these new insights epitomize the remarkable progress being made in the quest for a safer, more efficient, and more resilient air travel experience.
Tracking Aerial Targets
The research paper titled “Adaptive IMM-UKF for Airborne Tracking” presents a novel tracking solution for maneuvering aerial targets. The authors introduce an adaptive interacting multiple model (AIMM) that works in combination with unscented Kalman filters (UKFs), labeled as AIMM-UKF. The newly proposed system is designed to yield more precise estimates, improve the consistency of the tracker, and enhance robust prediction during periods of sensor outages.
The framework is built around two modes: a uniform motion model and a maneuvering model. It rapidly alternates between these two models based on a distance function that adjusts the transition probabilities. To verify the proposed solution’s effectiveness, the authors performed Monte Carlo simulations and compared the AIMM-UKF with ACAS Xa, the upcoming generation of airborne collision avoidance systems, using hypothesis testing of root mean square errors, normalized estimation error squared (NEES), a new proposed noise reduction factor, and an estimated maximum error of the tracker during sensor dropouts.
The experimental results showed the superior performance of the AIMM-UKF in terms of tracking accuracy, consistency, and the expected maximum error, especially in situations involving sudden and abrupt maneuvers and during sensor outages. For uniform linear motion, the performance was consistent with the ACAS Xa. However, for curvilinear trajectories, the AIMM-UKF performed better.
The authors suggest that the findings of their research will benefit the design of target tracking systems, particularly in the fields of counter-UAV technologies and military applications. Future work includes creating a dataset of airspace encounters with ground truth data and observation data, and exploring the incorporation of modern artificial intelligence methods into the proposed framework.
Detecting Cyber Threats
A paper titled “TCFLTformer: TextCNN-Flat-Lattice Transformer for Entity Recognition of Air Traffic Management Cyber Threat Knowledge Graphs” presents a novel method for entity recognition in air traffic management (ATM) cyber threat detection using a model called TextCNN-Flat-Lattice Transformer (TCFLTformer). The researchers developed this model to improve upon traditional machine learning methods and more recent deep learning techniques, which were found to be lacking in recall and accuracy or struggled with capturing both global and local features. The TCFLTformer, with its CNN-Transformer hybrid architecture, first utilizes convolutional neural networks (CNN) to extract local features from the text and then uses a Flat-Lattice Transformer to learn temporal and relative positional characteristics of the text to achieve final annotation results. The model is also designed with a relative positional embedding (RPE) and a multibranch prediction head (MBPH) to enhance deep feature learning and encode position text content information.
The study introduces the ATM Cyber Threat Entity Recognition Datasets (ATMCTERD), containing 13,570 sentences, 497,970 words, and 15,720 token entities collected from international aviation authorities and cybersecurity companies. In tests using these datasets, the TCFLTformer achieved the highest accuracy and precision scores, at 93.31% and 74.29%, respectively, compared to six other Named Entity Recognition (NER) models. Additional experiments were conducted on the MSRA and Boson datasets for a more comprehensive evaluation of the model’s effectiveness.
The researchers conclude that the TCFLTformer shows promise for ATM cyber threat entity recognition, outperforming other popular methods in terms of accuracy and recall. However, they also note that the limited size and scope of the datasets used in this study constitute a potential shortcoming and suggest that future research could use larger datasets and consider other large-scale deep learning models, such as GPT and RWKV, for comparison and analysis.
Real-Time Flight Arrival Predictions
A paper titled “A Data-Light and Trajectory-Based Machine Learning Approach for the Online Prediction of Flight Time of Arrival” presents a new method for predicting flight arrival times in real time while a flight is airborne, specifically the Estimated Time of Arrival at Terminal Airspace Boundary (ETA_TAB) and Estimated Landing Time (ELDT). The method is data-light, meaning it requires minimal data inputs and is easy to implement, and is intended for use by stakeholders like airlines, airports, and air travel app developers who lack access to extensive real-time information.
The method makes use of machine learning techniques and uses only flight trajectory information, specifically latitude, longitude, and speed. The process includes four stages: reconstructing the sequence of trajectory points from the flown trajectory and identifying the most similar historical trajectory; predicting the remaining flight trajectory based on the flown path and the matched historical trajectory using a Long Short-Term Memory (LSTM) network; predicting the flight’s ground speed along its projected path using a Gradient Boosting Machine (GBM) model; and predicting ETA_TAB and ELDT using the trajectory and speed predictions.
The LSTM and GBM models used in the method can be trained offline, keeping online computational needs to a minimum. The approach was tested with real-world US flight data, and it was found to perform better than several alternative methods. The simplicity and effectiveness of the method make it attractive to potential users who need real-time ETA prediction but have limited access to data.
Despite the good performance of the approach, the researchers acknowledged that more sophisticated models with access to additional data like airspace congestion and en-route weather conditions could potentially improve the prediction accuracy. Future research could look into including more historical trajectories, incorporating altitude data into the trajectory prediction, and refining the prediction of flight terminal approach time.
Calculating Delays and Predicting Interruptions
In March, we covered a new artificial intelligence technology created as part of a project called Artimation at Mälardalens University (MDU) in Sweden. It aids air traffic controllers by calculating delay lengths and predicting interruptions. “The project results will improve the functionality, acceptance and the reliability of AI systems in general, but also meet global goals such as the improvement of industry, innovation and infrastructure in society,” according to Mobyen Uddin Ahmed, Professor of Artificial Intelligence at MDU.