Modeling and testing of temporal and non-linear dependence in a multivariate process system

Ghosh, Arko (2019) Modeling and testing of temporal and non-linear dependence in a multivariate process system. Masters thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (2MB)

Abstract

This thesis presents the modeling and testing of temporal and spatial non-linear dependence among the process components in a process system. Due to interconnectivity among process units, the variables are highly correlated and dynamic. Accident models should capture these complex and dynamic behaviours of the components to predict accidents early. Fault tree, Dynamic fault tree, Bayesian network, Dynamic Bayesian network and Copula-based Bayesian network models have been selected to model these characteristics of the variables and develop the early prediction of accidents. At first, temporal dependency has been modeled and experimentally validated. The performances of dependence models are illustrated for accident analysis using Fault tree, Dynamic fault tree and Bayesian network models. Process datasets from a lab-scale pilot plant introducing faults into the system have been used for this purpose. The analysis shows that the inherent properties to capture different spatial (indirect dependencies) and temporal dependencies among process variables make the Bayesian network superior to Dynamic fault tree and the traditional fault tree models. Secondly, non-linear spatial dependence (modeled as covariate direct dependence) along with temporal dependence have been modeled to investigate accidents. A copula-based Bayesian network and traditional Bayesian network have been used to model direct dependence and the performances of the models are validated experimentally. A pilot plant has been used to perform experiments and collect process data sets. The results illustrate that, the copula function can capture the non-linear dependence among process variables. The integration of the copula function and Bayesian network can predict accident probability more efficiently than the traditional Bayesian network. The successful validation of the accident models confirms the evolving nature of the models capturing spatial and temporal dependence to address operational safety challenges in the process industries.

Actions (login required)

View Item