Learning Methods and Predictive Modeling to Identify Failure by Human Factors in the Aviation Industry

Abstract

This paper proposes a model capable of predicting fatal occurrences in aviation events such as accidents and incidents, using as inputs the human factors that contributed to each incident, together with information about the flight. This is important because aviation demands have increased over the years; while safety standards are very rigorous, managing risk and preventing failures due to human factors, thereby further increasing safety, requires models capable of predicting potential failures or risky situations. The database for this paper’s model was provided by the Aviation Safety Network (ASN). Correlations between leading causes of incident and the human element are proposed, using the Human Factors Analysis Classification System (HFACS). A classification model system is proposed, with the database preprocessed for the use of machine learning techniques. For modeling, two supervised learning algorithms, Random Forest (RF) and Artificial Neural Networks (ANN), and the semi-supervised Active Learning (AL) are considered. Their respective structures are optimized applying hyperparameter analysis to improve the model. The best predictive model, obtained with RF, was able to achieve an accuracy of 90%, macro F1 of 87%, and a recall of 86%, outperforming ANN models, with a lower ability to predict fatal accidents. These performances are expected to assist decision makers in planning actions to avoid human factors that may cause aviation incidents, and to direct efforts to the more important areas.