We describe how machine learning and decision theory is combined in an appl
ication that supports control room operators of a combined heating and powe
r plant to cope with the overwhelming complexity of situations when severe
plant disturbancies occur. The application is designed as an assistant, rat
her than as an automatic system that intervenes directly in the operator/pl
ant loop. The application is required to handle vague and numerically impre
cise background information in the construction of classifier committees. A
classifier committee (or ensemble) is a classifier created by combining th
e predictions of multiple sub-classifiers. The presented method combines cl
assifiers into a committee by using computational methods for decision anal
ysis that are designed to work when the information at hand is imprecise. T
he application evaluates and make priorities between classified alarms acco
rding to credibilities that depend on the current context. Machine learning
techniques are used to construct classifiers that recognize various malfun
ctions in a process, determine whether a situation is normal or not, and ma
ke priorities among alarms.