D-Vine-Based Correction of Physics-Based Model Output for the Identification of Risky Flights With Respect to Runway Overruns

• Published in: IEEE access Vol. 12; pp. 129173 - 129186
• Main Authors: Alnasser, Hassan, Pfahler, Marco, Hanebeck, Ariane, Beller, Lukas, Czado, Claudia
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accident prediction; Aerospace control; Aircraft; Atmospheric modeling; Calibration; Controllability; Data models; Error correction; Event detection; Flight; Mathematical models; Measurement; Model calibration; operational airline risk; Predictive control; Predictive models; rare events; Regression analysis; Regression models; Runway excursions; Statistical analysis; Statistical models; vine copulas
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3451719

Flight dynamic-based physical models are employed to assess the risk of runway overruns, predicting risk based on the metric given by the distance to a controllable speed of 80 knots after landing. The analysis incorporates Quick Access Recorder (QAR) data from 711 flights. Despite using observed flight data from flights without accidents as inputs, the physical model's risk predictions frequently exceed the observed risk metric, showing that the predicted risk metrics of the physical model are biased. To address this discrepancy, statistical models are designed to adjust the physical model's predictions by incorporating an additive error correction. Two correction approaches are evaluated: linear regression and D-vine copula regression. Linear regression fails to account for tail asymmetry, whereas D-vine copula regression does. Findings indicate that both correction methods enhance the physical model's accuracy relative to the observed risk metric, with D-vine copula regression achieving a closer match. To further refine the prediction of the risk metric, particularly in the tail of the distribution, numerous predictions are simulated using dependent input values generated using fitted R-vine copula models of the QAR data. Integrating these simulated dependent inputs and error corrections, the authors achieve a more accurate representation of the observed risk metric distribution and tail probabilities. This improvement is confirmed by comparing the Hellinger pairwise distances between the predicted and observed risk metric densities within the 2500 to 3000-meter range, where the D-vine copula regression error model with dependent inputs shows the smallest discrepancy. The results underscore the importance of considering asymmetric dependencies among input values to enhance the precision of runway overrun risk assessments.