Diagnosis of faults in component networks using constraint based models

Abstract

In this paper we propose a practical method for the diagnosis of faults in networks made up of interconnected components which may be modelled by means of a set of constraints. The method focuses on an efficient method to produce all sets of components which conflict with a given set of observational data. Reiter [1] has shown that if all the 'conflict sets' have been found, they may be used to derive all possible diagnoses. However, examining all subsets of components directly for conflict by use of a model leads to a large computational overhead. The method proposed here aims to reduce the computation for a known set of observations, such as for example readings from a known set of sensors. It is shown that in this case, we can precompute a set of tests which will give a subset of conflict sets which is sufficient to produce all possible diagnoses. The computational overhead of this method is substantially less than the direct application of the model to test all subsets.