Using flight data in Bayesian networks and other methods to quantify airline operational risks.

The risk assessment methods used in airline operations are usually qualitative rather than quantitative, despite the routine collection of vast amounts of safety data through programmes such as flight data monitoring (FDM). The overall objective of this research is to exploit airborne recorded flight data to provide enhanced operational safety knowledge and quantitative risk assessments. Runway veer-off at landing, accounting for over 10% of air transport incidents and accidents, is used as an example risk. Literature on FDM, risk assessment and veer-off accidents is reviewed, leading to the identification of three potential areas for further examination: variability in operational parameters as a measure of risk; measures of workload derived from flight data as a measure of risk; and Bayesian networks. Methods relating to variability and workload are briefly explored and preliminary results are presented, before the main methods of the thesis relating to Bayesian networks are introduced. The literature shows that Bayesian networks are a suitable method for quantifying risk and a causal network for lateral deviation at landing is developed based on accident investigation data. Flight data from over 300,000 flights is used to provide empirical probabilities for causal factors and data for some causal factors is modelled to estimate the probabilities of extreme events. As an alternative to predefining the Bayesian network structure from accident data, a series of networks are learnt from flight data and an assessment is made of the performance of different learning algorithms, such as Bayesian Search and Greedy Thick Thinning. Finally, a network with parameters and structure learnt from flight data is adapted to incorporate causal knowledge from accident data, and the performance of the resulting “combined” network is assessed. All three types of network were able to make use of flight data to calculate relative probabilities of a lateral deviation event, given different scenarios of causal factors present, and for different airports, however the “combined” approach is preferred due to the relative ease of running scenarios for different airports and the avoidance of the lengthy process of modelling data for causal factor nodes. The preferred method provides airlines with a practicable way to use their existing flight data to quantify operational risks. The resulting quantitative risk assessments could be used to provide pilots with enhanced pre-flight briefings and provide airlines with up-to-date risk information of operations to different airports, and enhanced safety oversight.