Robust Probabilistic Slow Feature Analysis for Soft Sensor Development and Model Quality Assessment

• Author / Creator

  Dyson, Cameron

• Model predictive control (MPC) is a popular advanced control technology. Unfortunately, over time the behaviour of the plant may deviate from its initial design conditions resulting in model-plant-mismatch. The detection and diagnosis of such mismatches is an important task to ensure that MPC systems are operating optimally, and any potential model re-identification is targeted to only necessary sub-models. Conventional mismatch detection methods directly use plant operating data for such purposes. The quality of assessment of these methods may suffer in the presence of significant disturbances. In Chapter 2, a linear slow feature analysis (SFA) data reconstruction is proposed to remove fast and typically irrelevant variations, extracting only those slow-varying and important components of the data to detect model-plant-mismatches. This preprocessing approach is shown to improve the performance of a conventional model-plant-mismatch detection method through both simulated and industrial case studies, and thus provide a more targeted selection of sub-models for re-identification.As SFA does not directly model process noise, and the conventional probabilistic SFA (PSFA) extension treats all noises as Gaussian, these algorithms are susceptible to the presence of outliers in the data. As industrial data often contains outliers there is motivation to remedy this issue. In Chapter 3, a robust PSFA (rPSFA) method with the measurement noises modeled as a scale mixture of Gaussians, switched according to a Bernoulli distribution, is considered for the modelling of systems where data contains outliers. To demonstrate the effectiveness of the proposed method over regular SFA, conventional PSFA and a previously developed Student-t robust PSFA, simulations are conducted through Tennessee-Eastman benchmark process data. The algorithm is then applied to an industrial zinc roaster process.The developed rPSFA models the switching between inliers and outliers according to a Bernoulli distribution which is completely random with respect to the previous outlier-inlier state. Many industrial systems exhibit correlated noise behaviour in which an outlier is more likely to occur after another. To account for this, Chapter 4 replaces the Bernoulli distribution with a Hidden Markov Model (HMM) to allow for the previous measurement noise mode to influence the prediction of future noise modes. Further, current literature lacks an outlier robust PSFA based method that is designed to capture the behaviour of multi-modal systems. To this end, the proposed HMM based robust PSFA is implemented in a mixture model fashion, where multiple independent process models are developed simultaneously, and their results are blended according to some weightings. The proposed model is verified in a soft-sensor task for a simulated system with a single operating mode but with outliers generated according to a HMM. Additionally, an industrial system which contains outliers and displays two distinct operating modes is used to demonstrate the development of a soft-sensor and a MPC model-plant-mismatch detection and diagnosis task.

• Subjects / Keywords

  • Model Predictive Control
  • Slow Feature Analysis
  • Model Plant Mismatch
  • Soft Sensor

• Graduation date

  Fall 2022

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-spnp-p645

• License

  This thesis is made available by the University of Alberta Library 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.