Bayesian Solutions to Control Loop Diagnosis

• Author / Creator

  Gonzalez, Ruben T

• While there has been much literature in the area of system monitoring and diagnosis, most of these techniques have a relatively small scope in terms of the faults and performance issues that they are built to detect. When implementing several monitors simultaneously on a single process, a single problem can result in multiple alarms, making it difficult to single out the underlying cause. Recent work has been done on incorporating information from multiple monitoring systems by means of Bayesian diagnosis; however, work so far is still in its infancy. This thesis focuses on a number of techniques that can be used to improve performance of previously proposed Bayesian diagnosis techniques. Previous work improved Bayesian diagnosis by accounting for incomplete evidence (monitor readings). Evidence is often presented in a multivariate vector, thus evidence with missing elements is incomplete. Missing elements can also appear in the mode (or set of problem sources). Many times, the mode information can also be incomplete within the historical data, such modes are ambiguous. This thesis develops two approaches for handling ambiguous modes. One technique is derived using Bayesian methods, while another technique is a modification on Dempster-Shafer Theory. Evidence in previous work was considered to be a vector of discrete variables, and the resulting probability estimates consisted of discrete categorical distributions. However, most monitors have continuous outputs that are only discretized for the sake of alarms. Discretization results in information loss, so it is desirable to use a technique that can easily estimate likelihoods for continuous evidence. Kernel density estimation is a popular technique for the non-parametric estimation of probability densities. Non-paramteric methods enjoy the advantage of not requiring assumptions on the nature of the distribution, so that they naturally fit the shape of the data's distribution (which is the main motivation for discretization). Kernel density estimation enables the construction of non-parametric estimates for continuous densities, allowing us to circumvent discretization procedures. Bootstrapping was a topic of interest for generating additional data if the data was sparse; however, it is also likely that modes will be sparse, that is, the history will often not contain all modes of interest. This thesis presents a two-pronged approach: Frst, to break down the problem into analysing components and properly selecting monitors; second, to generate additional modes by incorporating gray-box models and bootstrapping. Finally, incorporating ambiguous modes will affect the autocorrelated mode solution, while incorporating continuous evidence through kernel density estimation will affect the autocorrelated evidence solution. This thesis lays down a framework for dynamic implementation of the newly proposed ambiguous mode and continuous evidence techniques.

• Subjects / Keywords

  • Diagnosis
  • Bayesian
  • Fault Detection

• Graduation date

  Fall 2014

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/R36094

• License

  This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.