A Hybrid Deep Neural Network Approach for Attribution Analysis and Risk Evolution of Aircraft Hard Landings

Hard landings represent a critical safety challenge in civil aviation, characterized by low probability, complex multi-factor coupling, and extreme data imbalance. This study proposes a physics-aware deep learning framework to analyze hard landing risks utilizing Quick Access Recorder (QAR) data. To standardize heterogeneous flight parameters, a distance-based spatial alignment method is implemented. Subsequently, a parallel hybrid network integrating CNN, LSTM, and Multi-Head Attention is constructed to synergistically capture both local pilot control inputs and global temporal dependencies of aircraft energy states. To mitigate the regression bias caused by the scarcity of high-load samples, a Weighted Focal Loss function is introduced, significantly optimizing performance on imbalanced data. Furthermore, the SHAP framework is employed for model interpretability, quantifying the attribution of risk to specific flight parameters. Results reveal a dynamic evolution of risk factors, transitioning from energy management parameters during the approach phase to instantaneous pilot control inputs–specifically pitch attitude and pitch rate–in the crucial moments preceding touchdown. This research enables the transformation of safety management from reactive investigation to proactive intervention, providing a data-driven basis for flight training and operational monitoring.