Applying Machine Learning to Enhance Runway Safety Through Runway Excursion Risk Mitigation

Abstract

Risk of runway excursion caused by pilots continuing an unstable approach to landing has been identified by aviation accident investigators as a primary contributing factor in airline landing accidents. The purpose of this research was to develop and test predictive models for unstable approach risk misperception in the National Airspace System using machine learning. The research applied machine learning algorithms to flight recorder data gathered from a fleet of commercial transport aircraft and made available by NASA. Once evidence of unstable approaches was identified and extracted from the flight recorder data, a determination was made whether a rejected landing or continuance to landing was made. Federal Aviation Administration unstable approach criteria were used in the identification of unstable approaches based on flight data variables evaluated at 500 feet above the ground on approach to landing. Six machine learning algorithms were used, including logistic regression, decision tree, gradient boosting, random forest, and support vector machine. The results indicated that the decision tree with three branches produced the best predictive model. This model was able to predict the pilot error of continuing an unstable approach to landing with an accuracy of 98%. Glideslope deviation, selected airspeed, localizer deviation, and flaps not extended were the most important influential predictors of pilot error leading to an unstable approach.