Abstract:

To further enhance the risk management mechanism during flight operations, an early warning for pilot handling smoothness was developed by integrating flight big data. First, core parameters related to unstable approaches were filtered from QAR data. The XGBoost algorithm was then utilized for feature optimization to identify key risk early warning indicators. Subsequently, a dynamic risk identification architecture capable of effectively capturing spatio-temporal dependencies was constructed by incorporating the attention mechanism of Transformer networks.. Finally, the method's performance was validated using flight data from B737-800 aircraft operated by an airline in Shandong. The results indicate that this method can effectively predict in-flight risk events, particularly in providing high-accuracy risk warnings during critical phases before landing. Compared with traditional warning methods, the approach demonstrates significant advantages in identification accuracy, model generalization capability, and feature extraction efficiency.

Key words: pilot, risk early warning, eXtreme Gradient Boosting(XGBoost), Transformer, quick access recorder(QAR)