Linear and Nonlinear Causal Analysis with Sparse Modeling

• Author / Creator

  Madhusoodanan, Nived

• Data has become an integral part of analysis and decision-making across various disciplines including health, economy, biology, process systems engineering, and sports. Despite the abundance of big data, the identification of the fundamental source of disruption in process systems engineering continues to pose a formidable challenge. Recognizing the significance of the rapidly evolving industrial landscape, in this thesis, we focus on various ways of using Causal Inference to determine the relationship among process variables to identify the root source of process abnormalities.
  The first contribution of the thesis considers the causal inference among nonlinear systems. The research investigates the applicability of the Gaussian Process in the Causality analysis among non-linear systems. Employing Gaussian processes facilitates the identification of causal relationships, thereby enabling a better understanding of complex non-linear systems. Since the traditional Gaussian Process-based causal inference tends to provide spurious causations, the study focuses on deriving sparse solutions to enhance the interpretability of the causal links identified. Sensitivity
  analysis is conducted to assess the robustness of the findings, providing insights into the reliability of the identified causal relationships within the intricate system dynamics.
  Given that conventional causal approaches rely solely on data, the integrity of the results is heavily contingent upon the quality of the data. Owing to the susceptibility of real-world industrial data to inaccuracies stemming from inadequate sensor maintenance, the effectiveness of complete data-based approaches is notably compromised. Hence, the second contribution of the thesis explores the possibility of amalgamating expert human knowledge with process data to reduce the over-dependence on data and produce more reliable and accurate modeling. In this case, it is achieved by using a constraint optimization technique where human knowledge is provided as constraints in the optimization. To verify the existence of the derived causal relationships
  and eliminate the possibility of chance findings, a novel surrogate generation algorithm specifically designed for oscillating data sets is also proposed.
  The third part of the thesis applies the previously developed methodology, integrating physics-based information with process data, to diagnose the challenge of flooding in separation columns using real industrial process data. Moreover, an advanced predictive model for forecasting instances of flooding is formulated and its results are analyzed. A user-friendly Graphic User Interface (GUI) toolbox is also
  developed to automate the process of combining process data and domain knowledge.
  The efficacy of all the contributions is verified through the numerical case study or industrial case study.

• Subjects / Keywords

  • Causal Analysis
  • Fault detection
  • Sparse Modeling
  • Gaussian Process
  • Physics-Informed

• Graduation date

  Spring 2024

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-mhyx-y162

• 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.