Predicting the Occurrence of Incidents Based on Flight Operation Data

Abstract

This paper introduces a novel method to quantify the current probabilities of accidents or incidents for a specific airline, based on flight operation data. The method allows airline safety managers to quantify, monitor, and continuously influence airline safety as expressed by accident probabilities. The paper includes a general description of the method, followed by an application to determine the probability of a runway overrun. Due to the small number of incidents that happen during flight operation, the flight operation database from a single airline has previously been insufficient for making statistical predictions directly from the incident data. For example, many airlines never experience even a single runway excursion during operation, yet the probability for this specific incident is also not zero. Taking the average of incident/accident statistics, e.g. from the International Air Transport Association (IATA), is insufficient, because airlines have different safety cultures, flight procedures, types of aircraft, routes, training, etc. One of the aspects of the method in this paper is to identify and trace the measurable (physical) variables for an incident/accident. For each incident, the factors are modeled in a Contributing Factor Trees (CFT), which is similar to a fault tree. The top event of each CFT is the incident modeled by an equation. The other elements of the CFT are the top event’s contributing factors, whose combined probabilities are equal to the probability of the top event’s occurrence. The method is based on the hypothesis that even in cases where the incident/accident does not actually occur, the various contributing factors deviate significantly from their nominal values, or even exceed the airline-imposed limits on a much more frequent basis. If we combine the statistical distributions which contain the deviations for the various contributing factors, we are thus able to obtain the probability for the top event.