3.1 Modeling for Diagnosis

 

Generating successful models for diagnosis introduces a unique set of requirements.

• The models should describe normal and faulty system behavior. The former provides the reference variable values for the monitoring task, and the latter forms the core for the prediction algorithm.
• The model should incorporate sufficient behavioral detail so deviations in observed variables can be mapped onto system components and parameters.
• The model should generate dynamic behavior, especially when faults cause transients that take the system away from normal steady state operation. Faults cause changes in system parameters, therefore, the assumption of constant parameters does not hold and their temporal effects have to be included.
• When faults occur, the system may undergo a structural change. This may cause a change in model configuration which has to be explicitly modeled as a failure mode.

In addition, to constrain the inherently exponential search space for diagnosis, it is important that the model impose all relevant physical constraints on the search process. Also, given the limits of purely qualitative and purely quantitative schemes that have been discussed elsewhere [6, 7, 20], models that generate and use both qualitative and quantitative information are preferred. This prevents loss of a priori information that may be useful for generating and further refining candidate sets.