A circumscribed diagnosis engine

Abstract

A difficult diagnosis task is to decide whether to incriminate or to exonerate the components of a system. A parsimonious theory of diagnosis requires doing exoneration. A robust theory of diagnosis must not presume the ways components fail. In order to build both, a parsimonious and robust theory of diagnosis, all the proofs which lead to the conclusion that components are not abnormal must be defeasible. This is a basic motivation to apply non monotonic reasoning to diagnosis problems. A basic issue becomes the choice of an exoneration criterion. Most model-based diagnosis engines exonerate components when there is a lack of evidence that they fail.

This article starts by introducing an alternative exoneration criterion based on the evidence that components possess the desired features defined by their model of correct behavior. Then this article shows how circumscription can be used to formalize this criterion. Next the properties of a Circumscribed Diagnosis Engine, CDE, are explained. As a result this article shows how a Circumscribed Diagnosis Engine enhances the localization of failures.