Process systems are often complicated and liable to experience faults and their effects. Faults can adversely affect the safety of the plant, its environmental impact and economic operation. As such, fault diagnosis in process systems is an active area of research and development in both academia and
industry.
The work reported in this thesis contributes to fault diagnosis by exploring the modelling and
analysis of fault propagation and detection in process systems. This is done by posing and answering three research questions. What are the necessary ingredients of a fault diagnosis model? What information should a fault diagnosis model yield? Finally, what types of model are appropriate to fault diagnosis?
To answer these questions , the assumption of the research is that the behaviour of a process system arises from the causal structure of the process system. On this basis, the research presented in this thesis develops a two-level approach to fault diagnosis based on detailed process information, and modelling and analysis techniques for representing causality.
In the first instance, a qualitative approach is developed called a level 1 fusion. The level 1 fusion models the detailed causality of the system using digraphs. The level 1 fusion is a causal map of the process. Such causal maps can be searched to discover and analyse fault propagation paths through the process.
By directly building on the level 1 fusion, a quantitative level 2 fusion is developed which uses a type of digraph called a Bayesian network. By associating process variables with fault variables, and using conditional probability theory, it is shown how measured effects can be used to calculate and
rank the probability of candidate causes.
The novel contributions are the development of a systematic approach to fault diagnosis based on modelling the chemistry, physics, and architecture of the process. It is also shown how the control and instrumentation system constrains the casualty of the process. By demonstrating how digraph models
can be reversed, it is shown how both cause-to-effect and effect-to-cause analysis can be carried out.
In answering the three research questions, this research shows that it is feasible to gain detailed insights into fault propagation by qualitatively modelling the physical causality of the process system. It is also shown that a qualitative fault diagnosis model can be used as the basis for a quantitative fault
diagnosis model