The design and implementation of a probabilistic model-based fault diagnosi
s expert system is described in this paper. Possible cause and effect graph
(PCEG) methodology, an enhanced signed-directed graph (SDG) approach, was
used for qualitative modeling. A rule-based approach is proposed to transfo
rm the Bayesian belief network into an acyclic network dynamically during t
he diagnosis phase to allow simple on-line probability calculation in a bel
ief network with causality loops. A dynamic time-delay methodology is also
proposed to manage the possibility of phantom alarms, which are the consequ
ences of process time-delays. The application to a pilot scale distillation
column with communication with DCS environment is presented. Two sample ru
ns were included to demonstrate the concept of dynamic causal network modif
ication and time-delay management. (C) 2000 Elsevier Science Ltd. All right
s reserved.