Trouble-Shooting by Graphical Models

Abstract

Warranty claims, call backs, accidents and troubles during production can give raise to massive error backtracking or diagnosis. One way to support diagnosis is to use Bayesian belief network models (BBNs) as a kind of graphical models. They allow to backtrack error symptoms to their causes or to predict effects of interactive engineering settings. Trouble shooting with such models is based on a two-step procedure. Firstly, the causes-effects structure is designed as a directed, acyclic graph. Secondly, (subjective) probabilities are allocated to each node. According to the inner degree of a node the probabilities are either of the marginal or conditional type. Due to the singularity of call back events no data-driven procedure can be applied. Instead, domain experts have to allocate subjective probabilities to the set of variables. Various sensitivity studies are used in order to highlight extreme scenarios. These scenarios give the management of Company X a fair chance to identify individual households who use risky ovens. HUGIN Light is used as an appropriate software tool.