An integrated risk prediction model for corrosion-induced pipeline incidents using artificial neural network and Bayesian analysis

Abstract

In onshore hazardous liquid transmission pipelines, corrosion-induced incidents are potentially significant hazard to people, property and environment. Therefore, several statistical data analysis methods have been developed to analyze their risks. These existing models utilize only a few attributes from a wide incident database and generally assume failure rates as homogeneous or constant over time which is not true. There is also a lack of frameworks to predict risk by utilizing the current condition of pipelines and historical incident data simultaneously. This article presents an integrated risk prediction model which leverages rich incident database of Pipeline Hazardous Material Safety Administration (PHMSA) and utilizes nonhomogeneous failure rates derived from data. From incident data for 2010–2019, 70 attributes are selected based on their significance to corrosion-induced pipeline incidents, and artificial neural network (ANN) models are developed for the prediction of causes and consequences of incidents. Next, the probability of incident is predicted using Bayesian analysis with nonhomogeneous failure rates. Here, ANN models employ current conditions of pipelines and Bayesian analysis utilizes historical incident data simultaneously. The predicted consequences and probability are multiplied to predict risk of corrosion-induced incidents. The effectiveness of the proposed framework is demonstrated using the PHMSA database.