This paper describes a novel approach to analysing trends of a performance signal indicator from an industrial metallurgical reactor over a number of years of operation. Using a minimum message length algorithm, a detailed ontology of the signal behaviours or modalities was established. An abstraction of these yielded a number of related super states that in turn provided an insightful correspondence for the domain experts. Further detailed identification of the likely composition and causal influences contributing to each mode was subsequently induced with supervised learning.
Available at: http://works.bepress.com/dstirling/5/