Root Cause Analysis of Key Process Variable Deviation for Rare Events in the Chemical Process Industry

Abstract

Root cause analysis of rare but catastrophic events in the chemical process industry must deal with the challenges of data scarcity that may lead to inaccurate diagnosis. Previously, Bayesian models (BMs) have been applied with fault trees to account for data scarcity. However, the BM does not account for source-to-source variability in collected data. To deal with this limitation, this work proposes a new framework to simultaneously handle data scarcity and source-to-source variability. For the purpose of computational efficiency, it first identifies key process variables (KPVs) for rare events using a sequential combination of relative information gain and Pearson correlation coefficient. Then, it performs the root cause analysis of KPV deviations using the Hierarchical Bayesian Model with an informative prior constructed from process data to handle source-to-source variability. Finally, performance of the proposed framework is demonstrated through a case study of the Tennessee Eastman process.

Modeling the effect of emergency response on domino effects in the coal gasification process by fuzzy hierarchical analysis and Bayesian network

“…FAHP was applied to the safety analysis of offshore oil platforms and subways, which shows the superiority of this method in system assessment [ 20 – 22 ]. In recent years, BN is widely used in the representation and reasoning of uncertain knowledge in risk assessment [ 19 , 23 , 24 ]. In order to deal with unexpected accidents, it is very important to improve the emergency response ability of the petrochemical industry.…”

A Direct Transfer Entropy-Based Multiblock Bayesian Network for Root Cause Diagnosis of Process Faults

“…Hence, contribution plots identify the root cause of process faults, i.e., the faulty variables, by comparing contributions of all process variables to the monitoring statistics such as the T 2 and SPE statistics . Among other data-driven approaches for diagnosis in the CPI, methods based on the causal network such as Bayesian model are also widely used. − The Bayesian model is also a probabilistic model and therefore can handle the stochastic nature of most chemical processes.…”

A hierarchical Bayesian regression framework for enabling online reliability estimation and condition-based maintenance through accelerated testing

“…Meanwhile, Bayesian inference, such as Bayesian Network (BN) and Hierarchical Bayesian Modelling (HBM), has also become a popular tool for condition monitoring purposes, thanks to its features, among which the ability to deal with source-to-source variability (Kumari et al, 2020). Moreover, Bayesian Inference can provide better results than other parametric regression methods, such as Maximum Likelihood Estimation (MLE) (BahooToroody et al, 2020a).…”