An interpretable and augmented machine-learning approach for causation analysis of major accident risk indicators in the offshore petroleum industry

Abstract

The offshore petroleum industry faces continuous stress and threats of major accidents. A complete and well-designed spectrum of indicators could help identify and improve the understanding of the status of all relevant hazards for managing various risks on offshore installations. The Petroleum Safety Authority (PSA) in Norway develops a project of extended indicators, i.e., “Trends in risk level” (RNNP), which uses an assortment of risk indicators to reflect the status and trend of the risk level. However, a simple statistical approach is insufficient to see into the causes behind the risk indicators. This paper proposes an interpretable and augmented machine-learning approach for causation analysis of major accident indicators. The proposed methodology integrates the K-means SMOTE technique for data augmentation to promote the performance of machine-learning algorithms. Armed with the analytical approach, the problem of lack of sample data is solved and the importance of the features of risk indicators is analyzed. Furthermore, a case study is presented to augment and analyze the two types of risk indicators associated with major accidents recorded from 2005 to 2021. The effectiveness of the proposed approach is demonstrated by the case study. It can provide an effective tool to help find out which type of features deserves more attention in risk management and target risk reduction measures that could be taken to strengthen safety on offshore installations.