Towards online prediction of safety-critical landing metrics in aviation using supervised machine learning. (November 2020)

• Record Type: Journal Article
• Authors: Puranik, Tejas G.; Rodriguez, Nicolas; Mavris, Dimitri N.
• Is Part Of: Transportation research. Volume 120(2020)
• Publication Date: 2020-11
• Subjects: Air transportation -- Risk -- Safety -- Supervised learning -- Random forest -- Landing performance; Transportation -- Periodicals; Transportation -- Technological innovations -- Periodicals; 388.011
• DOI: 10.1016/j.trc.2020.102819
• Languages: English
• ISSNs: 0968-090X

Abstract

Highlights: Provides a novel online predictive model of aircraft landing performance using data collected on-board an aircraft during the approach phase. Demonstrates accurate prediction of the critical metrics further ahead in time than existing approaches and at a higher accuracy. Introduces innovations in generating feature vectors and target labels using a flexible approach that can be easily replicated for other metrics of interest. Significantly improves over previous similar work in literature both in terms of accuracy and applicability in an online setting. Abstract: In recent years, due to the increased availability of data and improvements in computing power, application of machine learning techniques to various aviation safety problems for identifying, isolating, and reducing risk has gained momentum. Data collected from on-board recorders in commercial aircraft contain thousands of parameters in the form of multivariate time-series (continuous, discrete, categorical, etc.) which are used to train the machine learning models. Among the phases of flight, approach and landing phases result in the most accidents and incidents. The performance and trajectory of the aircraft during the approach phase is an indicator of its landing performance which, in turn, affects incident or accident probability such as runway excursions. Landing performance is commonly measured using metrics such as landing airspeed, vertical speed, location of touchdown point on runway, etc. While current applications of machine learning to aviation focus on retrospective insights to implement corrective measures, they offer limited value for real-time risk identification or decision-making as they are inherently reactive in nature. In this work, a novel offline-online framework is developed for building a global predictive model offline to predict landing performance metrics online. The framework leverages flight data from the approach phase between certain approach altitudes (also called gates ) in order to train the offline model to predict the landing true airspeed and ground speed using a Random Forest regression algorithm. Permutation importance is used to identify and retain important features among those available and the training data is balanced with respect to flight safety events to ensure a good performing model. The developed global model is robust and can predict landing true airspeed and ground speed with root mean square errors of 2.62 and 2.98 knots respectively over six different airframes operating at over seventy airports. The model operates fast in prediction mode which, coupled with the ability to provide the prediction at an altitude where a go-around decision may be made, makes it particularly suitable for online application. The framework is demonstrated using data obtained from commercial airline operations that contains thousands of flight records and performs better than existing techniques in literature at predicting true airspeed and ground speed at touchdown.