PATH-SAMPLING AND MACHINE LEARNING FOR RARE ABNORMAL SAFETY AND RELIABILITY EVENTS
Degree type
Graduate group
Discipline
Data Science
Engineering
Subject
Industry 4.0
Machine Learning
Multivariate Alarm Systems
Rare Abnormal Events
Funder
Grant number
License
Copyright date
Distributor
Related resources
Author
Contributor
Abstract
It is crucial for chemical and manufacturing industries to ensure safe and reliable operation of their plants and processes, by mitigating safety issues (e.g., extreme operating conditions) and reliability issues (e.g., production losses). But, a significant challenge faced by these industries is that such events are rare and undesirable, with little occurrence data available from process historians. Extensive control and alarm systems, with Safety Instrumented Systems (SIS) and reliability risk assessment methods, are often successful in mitigating postulated abnormal events anticipated in HAZOPs. However, it is very challenging to consider the effects of highly infrequent unpostulated abnormal events (i.e., non-specific, randomly-occurring events), which cannot be anticipated in process design, and lead to severe consequences. Hence, in this dissertation, novel, improved multivariate alarm systems are developed using path-sampling and machine learning, for handling rare unpostulated abnormal events resulting from random perturbations in one or more process variables. As a first application of path-sampling to analyze rare abnormal events for chemical process safety, Moskowitz (2016) introduced transition path-sampling (TPS) to locate rare safety pathways for an exothermic continuous stirred-tank reactor (CSTR) and an air separation unit (ASU). In this dissertation, to circumvent the computational limitations posed by TPS, forward-flux sampling (FFS) is introduced. It simulates rare unpostulated abnormal events more-efficiently in a piecewise manner, moving from desirable to undesirable operating regions, with valuable key process-variable data stored during the simulations, followed by calculations of committer probabilities to reach undesirable regions (i.e., pB). Given the process variable-pB data, accurate predictive models are developed using machine learning (ML) – one of the cornerstones of Industry 4.0’s vision for improved automation through digital transformation. Using predictions provided by the ML-based models, initial multivariate alarm systems are developed, which are improved significantly by introducing an alarm rationalization - dynamic risk analyses (DRAn) integrated framework. Such improved systems, when implemented alongside widely-used HAZOP studies, aid operators in handling both postulated and unpostulated abnormal events to improve overall safety and reliability.
Advisor
Patel, Amish, J.