Abstract

In this thesis, the design of model-based fault diagnosis systems for Discrete Event Systems (DES) is studied. The correct performance of a model-based diagnoser depends on the accuracy of the plant DES model used in its design. The behavior of this nominal model may differ from the actual plant behavior (i.e., the true model) due to various reasons such as modeling errors, modeling simplifications and coding errors. The difference between the nominal model and the true model may result in observations (i.e., sensor readings) that are unexpected by the diagnoser, resulting in "discrepancy."
In the literature, "learning diagnosers" have been proposed that in cases of discrepancy add transitions to the nominal DES model to account for the unexpected behavior. It is assumed that the nominal and true models have the same number of states, and their difference is in the transitions.
In general, every discrepancy can be explained by different sets of additional transitions, each representing a hypothesis. To narrow down the list of hypotheses, the principle of parsimony is used - giving preference to less complex hypotheses. After a set of hypotheses is generated, using future observations, this set is narrowed down. In some cases, the set may have to be expanded.
In this thesis, a new approach to learning diagnoser is introduced that takes advantage of the structure of the plant to generate hypotheses. This is meant to reduce the number of hypotheses and to generate hypotheses that are more likely to correctly explain the discrepancies. Specifically, the proposed method takes advantage of the fact that models are built incrementally from component models and their interactions. Hence, in the learning process, new transitions (to explain discrepancies) are added to the component models and their interactions (rather than to the complete flat model of the plant).
This results in a compositional approach to learning. In this thesis, a framework for the compositional approach is presented and the learning rules and the corresponding algorithms are developed. A case study from process control is used to illustrate the proposed approach.