A Supervised Learning Approach for Safety Event Precursor Identification in Commercial Aviation
Jamey L. Ackley, Tejas G. Puranik, Dimitri N. Mavris
Abstract
Growth in the aviation industry coupled with advancements in technology and data science techniques have led commercial airlines to more data-driven analysis for both operational efficiency and improved safety practices. Multiple sources of safety-related information are now available to airline safety teams for evaluation post-flight. Various techniques currently exist that leverage large data sets to identify anomalous flight behavior. While many of these techniques provide valuable insight into safety-related events as triggered by various parameter exceedances, few are capable of leveraging available data to identify precursors to triggered events. Event precursors are often difficult to detect due to the high dimensionality of data sets and the heterogeneous nature of the information captured. Therefore a robust and repeatable methodology is needed to identify flight parameter subsets from the available data sources that can accurately predict safety related events prior to occurrence. These parameters can provide insights on precursors to events and enable proactively avoiding them. In this paper, a methodology is developed using supervised machine learning techniques to assess and track critical parameters leading up to safety events in the approach and landing phase. Time-series data obtained from commercial aviation operations is collected, analyzed, and pre-processed. This data is then used in conjunction with safety event definitions to train a model for classifying the data into event and non-event flights. The parameter significance from these trained models is used to understand and isolate parameters that lead to the clear separation between event and non-event flights. Preliminary results from implementation of this methodology on commercial airline flight data identify critical parameters at multiple altitudes and their progression for unstable approach events, which can be used in defining precursors.