COST-EFFECTIVE DIAGNOSING BY SEQUENTIAL PROBING USING A BAYESIAN SYSTEM MODEL

  • Peter Vansweevelt

Student thesis: Master's Thesis

Abstract

The systems used in high-tech industries are becoming more complex,more varied and are developing rapidly. Down-time of a machine must be reduced as much as possible. Detecting the reason of a failure in these complex systems is then an important task, saving time andmoney if it can be done efficiently.
To isolate the most probable diagnosis, the diagnosing task can be supported by a tool, based on a model of the system, fed by data and expert knowledge. The tool can reduce the cost and augment the effectiveness of diagnosing by weighing an increasing cost of probing with an increasing certainty about the diagnosis.
Having a Bayesian model of a system at hand, a diagnostic framework has been
built. This framework uses the existing Model Based Diagnosis and Bayesian troubleshooting tools as background. It offers concepts, definitions and formulas to use in the diagnostic process. Particularly, some information functions and utility functions combining cost and information, are developed.
The experiments done on different example systems with different cost distributions, show that it is not possible to deduce general rules about which functions to use and how to set the values of their constants. Therefore, a heuristic is proposed leading to a low expected cost of diagnosing with a limited computational effort. The heuristic is based on specific characteristics of the cost-information functions. The CEDUBAM app, developed for this study, implements the heuristics and, given an existing Bayesian SystemModel, generates a cost-effective probe scenario.
The framework developed in this thesis to solve diagnostic tasks, including costs,
has a solid theoretical background, is experimentally investigated and practically implemented, making it a foundation for further research.
Date of Award15 Mar 2024
Original languageEnglish
SupervisorArjen Hommersom (Examiner), Emile M.J.A.M. van Gerwen (Co-assessor) & Jesse Heyninck (Co-assessor)

Master's Degree

  • Master Computer Science

Cite this

'