Data-Driven Process Monitoring, Fault Detection And Fault Diagnosis

• Author / Creator

  Bahador Rashidi

• As the process control industry and production lines become highly complex and significantly invested with high-dimensional variables, process health monitoring attracts more attention from the domain experts and process operators. Since data is ubiquitous nowadays thanks to the advanced computer and communication technology, data-driven approaches are frequently used to ensure the safety and process quality performance. The proposed research in this thesis is mainly focused on two aspects: fault detection in non-stationary processes and root-cause fault diagnosis using causality analysis.

  In Chapter 2, fault detection is investigated from an unsupervised perspective for processes with non-stationary time-series measurements, especially those subject to time-varying mean changes. To this end, a moving-mean principal component analysis (MM-PCA) approach is proposed, in which the mean values of process measurements are updated using a moving-mean algorithm based on the upper bound of expected range of variations. The proposed MM-PCA does not require a heavy online calculation in comparison with the existing adaptive solutions and it can successfully compartmentalize the faults from healthy variations. Applying the concept of MM-PCA, three monitoring feature indices are proposed to monitor the statistical behavior of the process measurements. Moreover, an overall health index is suggested based on the proposed features using kernel density estimators (KDEs) which is considered as a process condition indicator.

  In Chapter 3, quality output-related fault detection based on non-stationary process measurement is studied. A cascaded modeling framework based on partial least-squares (PLS) approach is introduced, which entails a complete orthogonal projection of the process variables onto quality output-related and quality output-unrelated subspaces. The principal manifold is defined to represent the underlying auto-regressive model of the time-series, and such a relationship remains unchanged during the normal time-varying operations. Consequently, proper quality output-related and unrelated indices are derived.

  In a majority of multivariate processes, the propagating nature of abnormalities makes root-cause fault diagnosis a challenging task. As the second main focus of this research, we endeavor to develop a root-cause fault diagnosis framework based on causality analysis using transfer entropy (TE). With this aim, in Chapter 4, a novel data-driven strategy is proposed for real-time root-cause fault diagnosis in (non-)linear processes by estimating the causal strength between measured process variables and variations of a residual signal (e.g. square prediction error derived by PCA or kernel PCA) using normalized transfer entropy (NTE). A novel approach for TE estimation, i.e. the so-called symbolic dynamic-based normalized transfer entropy (SDNTE) is proposed, which has achieved a faster computation speed and less complexity than the conventional KDE method. For this purpose, a new definition of joint xD-Markov machine is given to capture dynamic interactions between two time-series. The concept of SDNTE is built upon principles of time-series symbolization, joint xD-Markov machine, and joint-Shannon entropies. Not only the SDNTE has less calculation complexity in comparison to KDE approach, but also the proposed general framework in this chapter can effectively identify the source of the process fault among certain potential candidates. The proposed root-cause fault diagnosis framework is applied to the Tennessee Eastman Process (TEP) benchmark and its computational advantages are clearly demonstrated.

  Finally in Chapter 5, a complete autonomous framework is proposed for conducting root-cause fault diagnosis which requires a minimum \textit{a priori} process knowledge and intervention of a human operator. Upon the presence of a fault, potential process variables are identified using a contribution score algorithm and SDNTE is used for generating the directed graph which presents the causal inference among the candidates. Then, Direct transfer entropy (DTE) is utilized to prune the indirect and spurious edges. To this aim, the application of symbolic dynamic filtering (SDF) is extended to the propose symbolic dynamic normalized direct transfer entropy (SDNDTE). Accordingly, concepts of immediate and source intermediate variables are defined and autonomous algorithms are developed to efficiently find them using the initial causal graph. In the end, a depth-first search (DFS)-based algorithm is developed and deployed on the pruned graph to locate the root-cause variable(s).

• Subjects / Keywords

  • Fault Detection
  • Root-CAuse fault Diagnosis
  • Symbolic Dynamic Filtering

• Graduation date

  Spring 2020

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/r3-cddc-sv93

• License

  Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.