Data-driven Modelling for Process Identification with Flat-topped Gaussian Uncertainty

• Author / Creator

  Tan, Ruomu

• In data-driven modelling, model accuracy relies heavily on the data set collected from target process. However, various types of measurement noise exist extensively in industrial processes and the data obtained are usually contaminated. If the influence of measurement noise is neglected, both the quality of models trained from data and performance of further operations, such as control and optimization of objective variables, will be affected significantly. A good output noise model is essential in data-driven modelling if one wishes to attain a process model with satisfactory performance. Instead of the regular Gaussian distribution assumption for the noise, a novel type of the noise distribution is proposed and corresponding solutions to the process identification problems are established accordingly in this thesis. Specifically, a flat-topped Gaussian distribution, which combines the Gaussian and uniform distribution, is formulated to model a class of disturbances that often occur in practice. Moment fitting strategy is proposed as a general approach to approximating the distribution function of summed random variables with different distributions. The Flat-topped Gaussian distribution is then applied for identification of linear processes. As for more complicated nonlinear models, Flat-topped Gaussian distribution is considered for Gaussian Process modelling. Gibbs Sampling is incorporated and combined with the posterior distribution obtained from Gaussian Process in order to reconstruct the original output. Mixture Gaussian approximation is also used as an alternative of approximating the Flat-topped Gaussian distribution. The proposed algorithms are validated by numerical simulations and industrial applications. Soft sensors for estimation of emulsion flow rate in Steam Assisted Gravity Drainage (SAGD) process based on data-driven modelling are developed and relevant practical issues are discussed.

• Subjects / Keywords

  • Soft sensor
  • Process Identification
  • Flat-topped Gaussian Distribution
  • Gaussian Process

• Graduation date

  Spring 2016

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/R3JS9HG71

• 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.