- 185 views
- 245 downloads
Hierarchical Monitoring and Probabilistic Graphical Model Based Fault Detection and Diagnosis
-
- Author / Creator
- Fang, Mengqi
-
As the rapid development of modern industry, data based fault detection and diagnosis for industrial processes have become increasingly critical to ensure process safety and product quality. To effectively make use of underlying features of process data, multiple data based fault detection and diagnosis algorithms have been developed, among which the multivariate statistical process monitoring (MSPM) algorithms and the probabilistic graphical model based algorithms have been widely used. Through unsupervised training, the conventional MSPM algorithms have the advantage of simplicity but do not use the labeled fault information in the training phase. On the other hand, the probabilistic discriminative classifiers are supervised models and trained with label information. This thesis starts from solving a practical industrial fault detection and diagnosis problem based on the unsupervised MSPM approaches. Then to fully make use of both process observations and fault information, a supervised probabilistic discriminative framework, namely conditional random field (CRF) model, is introduced and then extended to deal with various practical scenarios and challenges.
Specifically, as a practical study on real-time fault detection and diagnosis, an early flare event prediction for a refinery process is first considered. Different operating conditions and production requirements from different process units result in hybrid data characteristics, therefore a single fault detection and diagnosis algorithm is not sufficient to deal with the problem. In this sense, a hierarchically distributed framework is designed to solve this problem, with two integrated and interactive monitoring layers to detect faults and track the root causes. Based on this layout, the majority of flare events can be successfully predicted with limited false positives.
Additionally, when fault label is available, supervised probabilistic classifiers are further explored. As a discriminative counterpart of the widely used hidden Markov models (HMMs), the linear-chain CRF (LCCRF) is introduced with demonstrated superior fault diagnosis performance to the HMMs. Then three practical challenges are addressed by extending the conventional LCCRF frameworks to variants of CRFs. First, to deal with the missing data problem, a marginalized CRF model is developed with a proposed efficient inference strategy. Second, to solve the feature selection and online adaption problem for operating mode diagnosis, a two-stage hidden CRF (HCRF) structure is proposed by combining the max-margin trained HCRF and LCCRF into a hierarchical framework. Third, to address the fault detection and diagnosis problem for processes with multiple operating conditions, a switching CRF model is proposed to deal with the variations of the process conditions, by extending unitary LCCRF to multiple LCCRFs.
This thesis aims to provide improved solutions to the fault detection and diagnosis problems in practical processes. As shown through multiple case studies of different chapters, the effectiveness of the proposed algorithms is demonstrated.
-
- Subjects / Keywords
-
- Graduation date
- Fall 2020
-
- Type of Item
- Thesis
-
- Degree
- Doctor of Philosophy
-
- 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.