Data-driven models for safety management of complex systems

Alauddin, Md. (2021) Data-driven models for safety management of complex systems. Doctoral (PhD) thesis, Memorial University of Newfoundland.

Full text not available from this repository.

Abstract

The advancement in big data and computing has prompted many industries including process industries to re-examine their traditional roles in design, control, and maintenance. The data-based models have the potential to contribute significantly to making the process industry more efficient, much safer, and environmentally friendlier. However, poor quality data, process uncertainty, random and spurious errors are major challenges to be handled by the data-driven methodologies. The objective of this research is to develop data-driven models for the safety management of complex process systems. This work also extends process safety principles to pandemic risk management to safeguard human health. Even though the Coronavirus Disease of 2019 pandemic has not resulted from process operations, energy industries were one the hardest hit sectors due to the pandemic. The COVID-19 pandemic has disrupted processing operations resulting in the historic collapse in demand and price crash leading to unprecedented scenarios. Nonetheless, the present pandemic provides numerous possibilities to strengthen engineering risk management approaches for benefiting all enterprises, especially, the oil and gas processing sectors. The reasons for studying pandemic risk management using the process safety framework are three-folds; to evince the multi-disciplinary nature of process safety principles, to demonstrate similarities between system safety and epidemiological risk management, and to manage pandemic risk using process safety principles. This work presents robust data-based efficient methodologies for fault detection, fault characterization, and mitigation for ensuring safety. The robustness has been inculcated by explicitly addressing the data quality issues, reconciling data-driven models with mechanistic models, integrating meta-learning, and incorporating prior knowledge and expert opinions. This thesis addresses the data quality issue by developing a robust model based on harnessing data quality features and assigning a lower weight to the low-quality data. The proposed method demonstrated improved results in detecting abnormalities in two case studies; a continuous stirred tank heater problem and the Tennessee Eastman (TE) benchmark chemical process. The percentage improvements in accuracy in detecting the step fault (IDV-1 fault) of the TE process were 1.0 %, and 4.5 % on 1%, and 10% mislabeled data respectively. A process dynamics-guided neural network model has been proposed to improve generalization. This has been implemented by adding an additional layer to the deep neural network architecture to incorporate process dynamics such as material and energy balance equations, universal laws, standard correlations, and field knowledge. The proposal has been evaluated on regression and classification tasks representing transient and steady-state operations of chemical processing systems. It resulted in significant gains in predicting dependencies in steady-state operations and forecasting transient conditions due to its improved generalization ability. The fault detection of a neural network model has been improved by incorporating meta-learning as well. The thesis presents a data-driven fault detection model using an artificial neural network and variable mosquito flying optimization technique for parameter tuning by maximizing fault detection rate while minimizing the false alarm rate. The proposed fault detection method has been applied for detecting faults in the TE benchmark process. The thesis also presents a hybrid formalism in pandemic risk management where the performance of the mechanistic models has been improved with advanced data-driven approaches. Thus, an artificial neural network-based susceptible, exposed, infected, quarantined, recovered, deceased (SEIQRD) model has been devised to effectively capture the temporal variability of a disease spread. It also addresses inconsistencies in reporting of infected cases by assigning a higher weight to the mortality data which is a more credible indicator of the disease progression than the reported infected cases. This yielded satisfactory results in forecasting infection cases of the COVID-19 in Ontario, British Columbia, Italy, and Germany. Uncertainty is another critical factor of processing systems and epidemiological modeling. This thesis quantifies the risk with randomness in the model parameters such as incubation, infection, and recovery periods under distinct measures of lockdown, schools and business closures including no measure. The uncertainty caused due to government interventions, changes in individual behavior, and the advent of multiple waves of an outbreak has been accounted by the parameter sharing feature of a hierarchical Bayesian network. Markov chain Monte Carlo simulation has been used to study the variability in the hierarchical Bayesian network. Finally, the thesis presents pandemic risk management frameworks using engineering safety principles such as precautionary, as low as reasonably practicable, and the layer of protection analysis approaches. An event tree model of pandemic risk management for distinct risk-reducing strategies realized due to natural evolution, government interventions, societal responses, and individual practices has been proposed. The proposed framework also investigates the impacts of distinct interventions on the survivability of an infected individual under existing healthcare facilities. This thesis can help in ensuring safety of complex systems by detecting abnormalities using robust data-driven and semi-mechanistic models. The thesis explores the synergy between process safety and epidemiology to better understand, analyze, and manage the risk. This is a first step towards the interdisciplinary study of this sort and paves the way for more interdisciplinary studies, enriching both disciplines.

Item Type: Thesis (Doctoral (PhD))
URI: http://research.library.mun.ca/id/eprint/15697
Item ID: 15697
Additional Information: Includes bibliographical references -- Restricted until December 21, 2024
Keywords: data-driven, neural network, process safety, pandemic risk management, Bayesian network, precautionary principle, risk assessment
Department(s): Engineering and Applied Science, Faculty of
Date: October 2021
Date Type: Submission
Digital Object Identifier (DOI): https://doi.org/10.48336/B03P-ET93
Library of Congress Subject Heading: Neural networks (Computer science); Bayesian statistical decision theory; Industrial safety

Actions (login required)

View Item View Item