Improving Process Operations Using Support Vector Machines and Decision Trees

Abstract

Statistical pattern-recognition methods are now widely applied in the analysis of process systems to achieve predictable and stable operating conditions. For example, multivariate statistical process control (MSPC) techniques use historical operating data to detect abnormal events, and assist engineers to focus their troubleshooting efforts to reduced subsets of variables in an otherwise broad operational space. Through an iterative process, it is hoped that the system variability remains bounded. Usually only a few samples collected under a state of statistical control are of interest, whereas the rest, which may be used to uncover potential improvement opportunities, are ignored. Beyond statistical control, an additional step is required to reduce the dispersion of process quality variables attributed to common causes. To achieve this goal, common and sustained causes not identified by MSPC must be interrogated. In this paper, a methodology based on kernel-based machine learning concepts is proposed to identify decision boundaries. A sparse set of instances or exemplars is identified that define a linear decision boundary in a feature space, which is equivalent to defining a nonlinear decision function in the associated input space. This is extended to defining operating strategies by integrating inductive learning into a decision support framework. Such an extension is founded on the fact that the success or failure of state-of-the-art approaches are invariably linked to the presence or absence of useful knowledge embedded in the system.