| dc.description.abstract | In chemical process industries (CPIs), rare events are low-frequency high-consequence events caused by process disturbances (i.e., root causes). To alleviate the impact of rare events, it is crucial to understand their effects through consequence estimation and provide an efficient troubleshooting advice through root cause diagnosis. For these analyses, traditional data-driven methods cannot be used due to a lack of database for low-frequency rare events. This entails the use of a first-principle method or a Bayesian network (BN)-based probabilistic model. However, both of these models are computationally expensive due to solving coupled differential equations and the presence of a high number of process variables in CPIs. Additionally, although probabilistic models deal with data scarcity, they do not account for source-to-source variability in data and the presence of cyclic loops that are prevalent in CPIs because of various control loops and process variable couplings. Unaccountability of these factors results in inaccurate root cause diagnosis.
To handle these challenges, we first focus on developing computationally efficient models for consequence estimation of rare events. Specifically, we use reduced-order modeling techniques to construct a computationally efficient model for consequence estimation of rare events. Further, for computational efficiency in root cause diagnosis, we identify key process variables (KPVs) using a sequential combination of information gain and Pearson correlation coefficient. Additionally, we use the KPVs with a Hierarchical Bayesian model that considers rare events from different sources, and hence, accounts for source-to-source variability in data. After achieving computational efficiency, we focus on improving the diagnosis accuracy. Since existing BN-based probabilistic models cannot account for cyclic loops in CPIs due to the acyclic nature of BN, we design a modified BN which converts the weakest causal relation of a cyclic loop into a temporal relation, thereby decomposing the network into an acyclic one over time horizon. Next, to discover significant cyclic loops in BN, we develop a direct transfer entropy (DTE)-based methodology to learn BN. Since the key to discover cyclic loops is finding correct causality between process variables, DTE quantifies the causality effectively by accounting for the effects of their common source variables. | |
