Data-Driven Long-Landing Event Detection and Interpretability Analysis in Civil Aviation

Long-landing events (LLEs) can lead to reduced available runway length and increased operating costs, which are primarily caused by the improper operation of aircraft, i.e., human error. The pilot’s data comprehensively recorded in quick access recorder (QAR) during aircraft takeoff landing, analyzing QAR determine causes LLEs. Traditionally, domain experts inspect LLEs manually setting thresholds on uni-dimensional data. However, this approach cannot detect defects maneuvering technique because potential mutual information between different features large amount is not considered. This paper proposes a data-driven LLE detection causation analysis workflow, automatically mine analyze information, overcome existing problems. Firstly, dataset established based extracted from 2002 flights, considering landing phase aircraft. Subsequently, hybrid-feature-selection (HFS) method implemented select that highly relevant using both unsupervised supervised algorithms. A categorical Light Gradient Boosting Machine (LGBM) with Bayesian optimization (LGBMBO) model used performance improvement. Furthermore, visualized marginal effect key parameters for SHapley Additive exPlanations (SHAP). experimental results demonstrate proposed HFS-LGBMBO reduces computational cost achieves better performance. Additionally, they be avoided maintaining appropriate descent speed, altitude, proper angle.