Identification of Flight Maneuvers and Aircraft Types Utilizing Unsupervised Learning with Big Data

Security professionals are interested in knowing with a high level of confidence the details of each flight beyond what the self-broadcasted data provides. A program run by the Air Force is attempting to better characterize flights. From various data sources, this program attempts to determine aircraft identities. The purpose of this study is to test unsupervised machine learning methods to improve the Air Force's aircraft identification process. These alternative methods cluster unlabeled aircraft flight data from which we identify different aircraft types and maneuvers. The alternate methods include k-means and k-medoids. To improve the clustering, we first perform feature engineering. Some of the features built include acceleration rates, G forces, specific excess power, flight path angle, flight efficiency, and winding number. We then cluster the data using the above features and determine the ideal number of clusters via the average silhouette width. The initial clusters identify discrete flight maneuvers, such as takeoff and landing, and from these initial clusters we utilize sub-clustering to identify aircraft. To validate our aircraft identification process, we implement web scraping to develop a labeled dataset to compare the distribution of aircraft within the subcluster versus the initial cluster. The final model uses k-means, and 12 of the 14 sub-clusters generated by it have statistically different distributions of aircraft at alpha= 0.01 from the initial clusters. This indicates that we can better identify aircraft type given in which subcluster a data point resides. With some modifications, the method could be used by the Air Force to augment the current identification process.