A machine learning-based methodology for automated classification of risks in an oil refinery

Abstract

Oil refineries process hazardous substances at extreme operational conditions to produce valuable products. The necessary and required risk assessment is generally rather time-consuming and involves a multidisciplinary group of experts to identify potential accidental hypotheses, and compute their frequency and severity. With respect to this context, in this work, we present a machine learning method to mine out useful knowledge and information from available data of past risk assessments. The aim is at automatically classifying possible accident scenarios that may occur in oil refinery processing units by using SVM (support vector machines). Data from a previous qualitative risk assessment of an ADU (atmospheric distillation unit) of a real oil refinery is used to demonstrate the applicability of the SVM-based approach. The test classification was made with an F1 score of 89.95%. In this way, the results obtained showed that the proposed method is promising for efficiently performing automated risk assessment of oil refineries.