A Machine Learning-Based Framework for Aircraft Maneuver Detection and Classification

The increasing availability of historical air traffic data (e.g., Automatic Dependent Surveillance-Broadcast (ADS-B) data) has enabled more advanced post-analysis of traffic scenarios, which leads to a better understanding of decision-making in air traffic control. Such kind of analysis is often complex and requires a careful design of analysis tools. Advanced machine learning techniques are shown to be very effective in dealing with the complexity of air traffic data analysis. This paper presents a machine learning-based framework to detect aircraft maneuvers in past traffic data and classify the maneuver into three key air traffic maneuvers. Aircraft maneuvers are identified in the ADS-B data using Isolation Forest algorithm, followed by maneuver clustering using K-means algorithm. Three time-dependent contextual features are proposed for dynamic traffic scenario representation and shown to be effective for maneuver clustering. Each maneuver cluster is associated with a label provided by Air Traffic Controllers (ATCOs), indicating the reason for such maneuver which took place in the past. Experiments were conducted on the framework using a dataset of 2793 arrival trajectories over 30 days in two Singapore Flight Information Region sectors. The results show that the framework efficiently allows post-analysis of air traffic scenarios, by which one can gain better insights into the decision-making patterns of ATCOs in response to various air traffic scenarios.