Using candidate space structure to propose the next measurement in model based diagnosis

Abstract

Model-based diagnostic algorithms consist of two steps in a loop: (1) measurement of data on the device to be diagnosed and (2) evaluation of discrepancies between the data a nd predictions provided by the model. The next measurement is selected in accordance with an entropy-base d criterion. In order to simplify its evaluation additional assumptions have been imposed. They result in easy to evaluate, but only suboptimal solutions. In the paper we show how to reduce unnecessary comput ation so that the original criterion can be calculated efficiently. Properties of diagnoses with re spect to their capacity to predict values are analysed. All diagnoses which do not predict any value for a v ariable can be represented by the set of the smallest diagnoses with this property and the rest of the diagnostic space can be organised so that it is searched efficiently. An algorithm is outlined which makes use of these results.

This research is a part of the Vital project which is partially funded by the ESPRIT Program of the Commision of the European Communities, as project 5365. The partners in the Vital project are: Syseca Temps Reel (F), Bull Cediag (F), Onera (F), The Open University (UK), University of Nottingham (UK), Royal PTT Netherland (NL), Nokia (SF), University of Helsinki (SF), and Andersen Consulting (E).