Learning from major accidents: A machine learning approach
Nicola Tamascelli, Riccardo Solini, Nicola Paltrinieri, Valerio Cozzani
Abstract
Learning from past mistakes is crucial to prevent the reoccurrence of accidents involving dangerous substances. Nevertheless, historical accident data are rarely used by the industry, and their full potential is largely unexpressed. In this setting, this study set out to take advantage of improvements in data science and Machine Learning to exploit accident data and build a predictive model for severity prediction. The proposed method makes use of classification algorithms to map the features of an accident to the corresponding severity category (i.e., the number of people that are killed and injured). Data extracted from existing databases is used to train the model. The method has been applied to a case study, where three classification models – i.e., Wide, Deep Neural Network, and Wide&Deep – have been trained and evaluated on the Major Hazard Incident Data Service database (MHIDAS). The results indicate that the Wide&Deep model offers the best performance.