Qualitative interpretation of sensor patterns

Abstract
A framework that provides the ability to generate qualitative interpretations (QIs) from multisensor trend patterns for monitoring, control, and optimization of chemical plants is presented. QIs are transformations of sensor and quality product data into useful symbolic abstractions. The framework is founded on the principles of similarity-based pattern recognition. Although demonstrated for normality identification, the machine methodology is general-purpose and applicable to any context-dependent QI problem. The objective of this approach is to create a QI-map of known pattern classes that consists of spatially distinguishable regions of patterns in an n-dimensional representation space. Creation of a QI-map is a two-step process: unsupervised map generation followed by supervised labeling. The application of the ART2 neural network for clustering in the QI-map is described. The application of the ART2-based QI-map approach to process monitoring of a recycle reactor is also described.