Skip to the content.

Data-Driven Model Reduction, Scientific Frontiers, and Applications

Chiranjivi Botre, Artie McFerrin Department of Chemical Engineering
Modified Nonlinear Partial Least Square (PLS) as a Model Reduction Technique with Applications to Fault Detection

Authors

  • Chiranjivi Botre
  • M. Nazmul Karim

Abstract

Multivariate statistical techniques are powerful tools that utilize data based model reduction methods capable of efficiently handling process noise and correlated data sets. In literature, multivariate statistical methods are widely discussed for applications in process monitoring and fault detection. Moreover, data based process monitoring techniques have been successfully applied to applications where accurate process model are not available. Process monitoring can be carried out in two phases; Data based model reduction and statistical fault detection.

The partial least square (PLS) method is an input-output model that has been effectively applied to linear processes and is most commonly used for fault detection applications. A kernel extension of PLS has been proposed to provides an effective technique for modeling nonlinear industrial processes; however, the selection of an appropriate kernel function and its associated parameter values have shown significant impacts on the performance of the KPLS algorithm. Therefore, new methods that rely on using a multi-kernel adaptation of the KPLS model have been developed to enhance the fault detection performance through the use of a stochastic multi-objective optimization approach. Moreover, a multiscale representation of the incoming process data using a wavelet function based approach to effectively separate the deterministic and stochastic features of the data is applied to further enhance the KPLS method.

This holistic approach, termed the multi-scale kernel partial least square (MS-KPLS) method, has the ability to handle process noise, non-normal data distribution, and auto-correlated data sets to provide an effective model reduction technique that can be applied to nonlinear industrial process data. In this talk we will discuss the applicability of the PLS method to chemical process systems as a model-reduction technique and examine the need for new methods to handle noisy and nonlinear process data. Current state-of-the-art modifications to the PLS algorithm for addressing these issues will be discussed, with emphasis on the MS-KPLS algorithm. Finally, the performance of MS-KPLS methodology is illustrated through a process monitoring example using the benchmark Tennessee Eastman process problem (TEP).

References

  1. C. Botre, M. Mansouri, M. Nounou, H. Nounou, and M.N. Karim, Kernel PLS-Based GLRT Method for Fault Detection of Chemical Processes, Journal of Loss Prevention in the Process Industries 43 (), 212–224.
  2. J.F. MacGregor and T. Kourti, Statistical Process Control of Multivariate Processes, Control Engineering Practice 3:3 (), 403–414.
  3. P. Teppola and P. Minkkinen, Wavelet–PLS Regression Models for Both Exploratory Data Analysis and Process Monitoring, Journal of Chemometrics 14:5–6 (), 383–399.