A hybrid data-driven approach to analyze aviation incident reports

In this paper, we construct a hybrid data-driven approach to connect a variety of elements related to the occurrence of the incidents reported in the Aviation Safety Reporting System (ASRS) database with the incident consequence. The hybrid model aims to predict the risk associated with the consequence of each hazardous incident from the contextual information (e.g., flight phase, weather, visibility, and aircraft information, etc.) as well as the text-based description of the incident. The developed approach is illustrated with a four-step procedure. In the first step, all the event outcomes are grouped into five risk categories ranging from high risk to low risk by expert opinion. In the second step, a support vector machine model is trained for making classification based on the text description of the incident (event synopsis). Meanwhile, an ensemble of deep neural networks are trained to learn the intricate associations between the event contextual features and the severity of event outcome. In the third step, a probabilistic fusion rule is developed to blend the two model predictions together, thereby improving the hybrid model prediction performance. Finally, the risk-level prediction is further extended to the event-level outcome prediction using a probabilistic tree. Computational results are given to demonstrate the effectiveness of the hybrid model in quantifying the levels of risk related to the consequences of hazardous causes.