A BAYESIAN APPROACH TO FAULT ISOLATION - STRUCTURE ESTIMATION AND INFERENCE

This paper considers a Bayesian inference method for fault isolation. Given a set of residuals, and a set of possible faults, the task is to calculate the probability distribution of the faults. The method requires the conditional probability distribution of how the residuals respond given the possible faults. Especially important is to know the structure of this conditional probability distribution since it facilitates the use of efficient Baysian network techniques for the inference. The conditional probability distribution, and in particular its structure, is estimated from training data using a Bayesian approach. The approach is evaluated on a simple but illustrative example, where it is shown that the estimated structure and the distributions capture the dependencies that are important to make the correct isolation decisions.