Process Safety Assessment Considering Multivariate Non-linear Dependence Among Process Variables

• Published in: Process safety and environmental protection 2020-03, Vol.135, p.70-80
• Main Authors: Ghosh, Arko, Ahmed, Salim, Khan, Faisal, Rusli, Risza
• Format: Article
• Language: English
• Subjects: Bayesian analysis; Causation; copula function; Data collection; Failure analysis; Mathematical models; Multivariate analysis; multivariate process system; nonlinear dependency; Process safety analysis; Process variables; Risk management; Safety; Stochastic processes; Stochasticity; Time dependence
• ISSN: 0957-5820; 1744-3598
• DOI: 10.1016/j.psep.2019.12.006

Nonlinear dependencies among highly correlated variables of a multifaceted process system pose significant challenges for process safety assessment. The copula function is a flexible statistical tool to capture complex dependencies and interactions among process variables in the causation of process faults. An integration of the copula function with the Bayesian network provides a framework to deal with such complex dependence. This study attempts to compare the performance of the copula-based Bayesian network with that of the traditional Bayesian network in predicting failure of a multivariate time dependent process system. Normal and abnormal process data from a small-scale pilot unit were collected to test and verify performances of failure models. Results from analysis of the collected data establish that the performance of copula-based Bayesian network is robust and superior to the performance of traditional Bayesian network. The structural flexibility, consideration of non-linear dependence among variables, uncertainty and stochastic nature of the process model provide the copula-based Bayesian network distinct advantages. This approach can be further tested and implemented as an online process monitoring and risk management tool.