Model-based reasoning (MBR) uses computational models to support automated inference about mechanisms. It is known to out-perform older, associative methods in differentiating sensor failure from mechanism failure. Heretofore, MBR has used constraint-set models. Recently, there has been interest in taking the simulation models that are built as part of the design process and re-using them in diagnosis and sensor placement. This paper examines the requirements for such a re-use of these models, particularly those built for modeling process-flow systems.
The switch from constraint-set to simulation models will force new requirements on the inference theory (and its implementations,) on the specification languages for simulations, and on the simulation code itself.
The paper identifies areas in which the theory of MBR is not yet completely adequate for using the information that simulations can yield, and reviews recent work in these areas. In particular, it is argued that using MBR along with simulations forces the use of specific fault models. Fault models are used so that a particular fault can be instantiated into the model and run. This in turn implies that the component specification language needs to be capable of encoding any fault that might need to be sensed or diagnosed. It also means that the simulation code must anticipate all these faults at the component level.
This also may force changes in the practice of how simulation models are specified. Commonly, model specifications omit distinctions which are not crucial to the simulation of the mechanism's nominal behavior. Such omission is normally not a characteristic modeling language. Rather, it is something modelers do to simplify their model specifications, when there is no reason not to. However, during diagnosis, these details may be critical performing the simulation of a particular fault mode, and must be retained.