Discovering probabilistic models of pilot behavior from aircraft telemetry data

Abstract

A mobile ground-based sense-and-avoid system for Unmanned Aircraft System (UAS) operations was developed by the University of North Dakota. This system detected proximate aircraft with various sensor systems, including a 2D radar and an Automatic Dependent Surveillance-Broadcast (ADS-B) receiver. Information about those aircraft was then displayed to UAS operators with customized visualization software. Its risk mitigation subsystem was designed to estimate the current risk of midair collision for UAS operations below 18,000 feet MSL. However, accurate probabilistic models for the behavior of pilots of manned aircraft flying at these altitudes were needed before this subsystem could be implemented. In this paper the authors present the results of data mining a Flight Data Monitoring (FDM) data set from a consecutive 9 month period in 2011. Arbitrarily complex subpaths were discovered from the data set using an ant colony algorithm. Then, probabilistic models were data mined from those subpaths using extensions of the Genetic K-Means and Expectation-Maximization algorithms.