Combining natural language processing and bayesian networks for the probabilistic estimation of the severity of process safety events in hydrocarbon production assets

This work investigates the possibility of using the information contained in reports describing Process Safety Events (PSEs) occurred in hydrocarbon production assets to support Quantitative Risk Assessment (QRA). Specifically, a novel methodology combining Natural Language Processing (NLP) and Bayesian Networks (BNs) is proposed to estimate the probabilities of having PSEs of various classes of severity and identifying the factors that have mostly influenced their variation along the monitored period. A repository of reports of PSEs of hydrocarbons plants is considered to show the potentialities of the developed methodology. An application to a repository of reports of PSEs of hydrocarbons plants is considered to show the potentialities of the developed methodology. The results obtained in the application show that the proposed methodology allows identifying the critical factors for the severity of the consequences of PSEs. These results show that the framework can be used to inform and guide decisions about possible improvements of the system safety by mitigative and preventive barriers.