Credal networks for fault diagnosis: What we know about the unknowns

Supervisor: Daira Pinto Prieto (d.pintoprieto@vu.nl)

Background

We talk about fault diagnosis in the context of cyber-physical systems. When a cyber-physical systems goes down, it is a big challenge to identify the root-cause of what is failing. Typically, companies offer diagnosis applications that assist the service engineers with identifying such root-cause.

In this context, some sort of reasoning must be encoded into the diagnosis application, aiming to aggregating the available information about the system and its symptoms and providing with a solution. This problem has been tackled in the literature using probabilities and Bayesian theory. This approach impose strong assumptions that can be relaxed by using imprecise probabilities instead.

Description

This project aims to explore the application of credal networks to the domain of fault diagnosis. A natural consequence of applying imprecise probabilities is that the application’s conclusions are less precise (it can express “I cannot discern further given the available information”). Therefore, we propose to investigate how to explain this doubt and guide the gather new clarifying information. Some milestones of this project are:

• Find/simulate a fault diagnosis dataset.
• Define and deploy a Bayesian network as a baseline model.
• Learn about credal networks and applying them to the problem.
• Proposing original solutions to explain the unknows retrieved by the network and strategies towards information gain.

There might be the opportunity to develop this project within an internship.

Literature

Fabio Gagliardi Cozman. Graphical models for imprecise probabilities. International Journal of Approximate Reasoning. Volume 39, Issues 2–3, 2005, Pages 167-184, ISSN 0888-613X, [PDF].

Alberto Piatti, Alessandro Antonucci, and Marco Zaffalon. Building Knowledge-based Expert Systems: A Tutorial. In Advances in Mathematics Research, Jan 2010.