Toward safer aviation: Application of GA-XGBoost-SHAP for incident cognition and model explainability

Abstract

Flight incidents are characterized by complex mechanisms, leading to poor prediction model robustness and explainability. Based on the full-dimensional description of flight incidents, the explainable module is added to the prediction model to achieve its accuracy, stability, and explainability. Firstly, imbalance processing is performed employing the sampling method, and a genetic algorithm (GA) is applied for feature selection; these results are then considered as prediction model input. Secondly, an extreme gradient boosting algorithm (XGBoost)-based incident severity prediction model is established with five categories of none, minor, serious, fatal, and total as prediction labels; real data is used for validation, and the model shows good robustness and superiority. Finally, the SHapley Additive exPlanation (SHAP) is introduced to explain the correlation between incidents severity and input features and to measure feature importance. The results show that the proposed method has higher prediction accuracy and robustness. Which can provide some decision-making reference for aviation operation management departments to emergencies, learn the deep-seated law of incidents, and promote the paradigm of active safety management.