Identifying and Analyzing Atypical Flights by Using Supervised and Unsupervised Approaches

Abstract

Effective analysis of flight data is a prominent topic in the aviation industry, with flight data recorders containing vast amounts of information on flight safety violations and maintenance issues. A demand exists for innovative techniques to analyze flight data because predefined threshold criteria, such as exceedances, are often quite rigid and cannot encompass all types of safety issues. Further, the accumulation of flight data presents a big data problem, so that it is infeasible for humans to analyze the data manually in a reasonable time. Statistics show that many accidents and incidents in aviation are recurrent. Therefore, to mitigate potential safety issues, machine-learning algorithms can be trained to identify accident precursors. This research had two purposes. The first was to perform cluster analysis by using Kohonen self-organizing maps (SOMs) to identify unstable approaches that transpired at the Grand Forks, North Dakota, International Airport. The Cessna 172 model aircraft was used, and unsafe practices were identified without specifying predefined thresholds. The second purpose was to employ the analytical techniques asynchronously to address the big data problem. The validated results indicated that SOMs identified hard landings and unsafe low-level maneuvers and that some approaches that were high, fast, and steep would be harder to detect by using traditional flight safety analysis techniques.