Two-Layer Fusion Multi-Step Flight Safety Parameter Prediction and Early Warning Strategy Based on QAR Parameters

Flight safety, as the core of civil aviation transportation industry, has attracted much attention from society in recent years. Based on the concept of aviation safety and data-driven, this paper innovatively proposes a direct multi-step prediction model and early warning strategy based on QAR real-time flight parameter data. First, based on the data pre-processing, COG NORM ACCEL parameter was used as the prediction index, Random Forest algorithm was used to analyze the importance of the remaining 29 flight parameters, and 6 important flight parameters were selected for real-time prediction. Then, based on a dynamic weight allocation mechanism, an end-to-end two-layer fusion prediction model was constructed. The model employed LSTM, Transformer, and TCN as parallel base models, where complementarity among multiple models was achieved through feature-level first-layer fusion and prediction-level second-layer fusion. Numerical experimental results demonstrated that the proposed fusion strategy not only achieved hierarchical improvement in prediction performance but also enhanced the interpretability of prediction results. Finally, an early warning strategy based on the real value and sliding window volatility was established, and its effectiveness is verified by the simulation of CAP CLM 1 POSN. In conclusion, this study provides a new data-driven approach for aviation safety management, which helps to identify and deal with potential flight risks in advance, and provides a strong reference for the automated early warning system of airlines.