REPRESENTATION OF PROCESS TRENDS .4. INDUCTION OF REAL-TIME PATTERNS FROM OPERATING DATA FOR DIAGNOSIS AND SUPERVISORY CONTROL

A methodology for pattern-based supervisory control and fault diagonsi s is presented, based on the multi-scale extraction of trends from pro cess data described in Part III of this series (Bakshi and Stephanopou los, Computers Chem. Engng 17, 1993). An explicit mapping is learned b etween the features extracted at multiple scales, and the correspondin g process conditions, using the technique of induction by decision tre es. Simple rules may be derived from the induced decision tree, to rel ate the relevant qualitative or quantitative features in the measured process data to process conditions. These rules are often physically i nterpretable and provide physical insight into the process. Industrial case studies from fine chemicals manufacturing, reactive crystallizat ion and fed-batch fermentation are used to illustrate the characterist ics of the pattern-based learning methodology and its application to p rocess supervision and diagnosis.