Multivariate data-based safety analysis in digitalized process systems

Amin, Md. Tanjin (2022) Multivariate data-based safety analysis in digitalized process systems. Doctoral (PhD) 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 (3MB)

Abstract

Chemical process industries are vulnerable to accidents due to their inherent hazardous nature, complex operations, and growing size. Although the control system works as the first safety layer and is designed to maintain the setpoint within the safety limit, it cannot suppress all the deviations. Therefore, a warning system is used above the control layer to provide alarm(s) to the operators about unpermitted process deviations that cannot be negated by the controllers. Data-based process fault detection and diagnosis (FDD) and dynamic risk assessment (DRA) tools play a pivotal role to ensure that any significant process deviation is efficiently captured and necessary maintenance has been done to restore the process to normal operating mode. Besides, these tools can provide a detailed analysis of failure paths which are significant to prevent a fault from propagating into an accident. Conventional univariate monitoring is easier to implement and comes as standard with distributed control systems (DCS). This conventional approach is unsuitable in digitalized process systems due to increased close loop control, large process dimension, and complex interaction among variables. Modern process industries require techniques that can handle the complexity and scale of process plants. Timely detection of faults, diagnosis of root cause(s) of faults that affect multiple variables, and predicting a quantitative measure of consequence is vital to ensure process safety and reliability. The thesis deals with multivariate data-driven FDD and DRA for digitalized process systems. This research aims to reduce the technological gaps between the current methods and prerequisites of automated FDD and DRA tools for multivariate safety analysis. This thesis looks at improving all aspects of FDD and DRA methods, starting from data pre-processing to consequence analysis due to fault(s). First, the effect of data pre-processing is investigated in the context of multivariate FDD. Multivariate exponentially weighted moving average (MEWMA) is found to be an effective way of filtering process data without adversely affecting their correlation structure. The MEWMA is combined with PCA-BN, and a new method called MEWMA-PCA-BN is proposed. The developed framework can detect and diagnose the fault earlier than many contemporary multivariate process monitoring models. In this work, a novel methodology has been proposed to construct the BNs from historical fault symptoms. Second, the selection of the principal components (PCs) for the PCA-BN method is made automated; the correlation dimension (CD) is used in this regard. Also, a new methodology is proposed for developing BNs from continuous process data. Third, the prediction of the consequences of a fault has been adapted for multivariate process systems. A novel data-driven framework has been proposed for concurrent FDD and DRA using the naïve Bayes classifier (NBC), BN, and event tree analysis (ETA). This work utilizes a multivariate fault probability from NBC for dynamic failure prediction. It overcomes the limitation of using univariate probability in DRA. Finally, this thesis looks into improving the FDD performance by capturing the correlation structure of process variables and considering the consequence analysis. The R-vine copula is used to demystify the correlation structure accurately while the ETA predicts the consequences. Unacceptable deviation of risk is used as an indicator of a fault, and subsequently, root cause(s) diagnosis is performed using density quantile analysis (DQA). Industrial, experimental, and simulated datasets are used to test and validate the performance of the developed models. This thesis is an important step for multivariate data-driven FDD and DRA research.

Actions (login required)

View Item