Mapping real flight data to weighted directed networks: An investigation on precursor detection for runway excursion incidents (REI)

Runway excursion incidents (REI) during landing remain a major operational risk at plateau airports. This study presents an integrated framework that combines association rule mining, a Lift-weighted directed interaction network and AMR CN based machine learning early warning to identify precursors of long landing and lateral veer. Quick Access Recorder (QAR) data from 2,342 landings are converted into exceedance events and mined with the Apriori algorithm. The final ruleset contains 57 association rules with moderate to high Support and Confidence. Several rules show strong coupling; for example, an exceedance in the landing roll distance increases the probability of total landing distance exceedance by about 13.6 times. The rules form a 20-node Lift weighted interaction network that highlights engine performance and pilot control variables as central hubs. Network metrics and exceedance counts are used as inputs to AMR CN enhanced XGBoost and Random Forest classifiers for early warning of E2 long landing tendency and E9 lateral veer tendency. For E2, the XGBoost model reaches macro F1 0.930, with macro precision and macro recall near 0.93. For E9, the Random Forest model reaches macro F1 0.914, with macro precision and macro recall near 0.91. These values clearly exceed the artificial neural network (ANN) baseline, which remains near macro F1 0.76 for both tasks. Confusion matrix analysis confirms that AMR CN improves recall for Mild risk levels (from 83.3 %/75 % to 88.9 %/83.3 %) while keeping false Mild alerts for normal flights rare (0.2 %). The framework provides interpretable and quantitatively validated early warning support for runway excursion prevention.