Dynamic process monitoring and fault diagnosis with qualitative models

18] P. M. Frank. Fault diagnosis in dynamic systems using analytical and knowledge-based redundancy { a survey and some new results. detection in an experimental heat exchanger process: a multiple model approach. Knowledge-based diagnosis systems for continuous process operations based upon the task framework. 26 models to the data in a Qmimic-style system is currently being studied 33]. A compiler from physical models into qualitative diierential equations. 25 QMI Qmimic Behavior generation Generate all trees Incremental Predictions Qualitative Qualitative and quantitative Observations Slope only Slope and average Conndence in diagnoses Fuzzy logic Statistics Types of faults Single faults Cascading faults possible Table 8: Diierences between QMI and Qmimic for exact quantitative models and values of parameters. This is particularly useful when the plant changes operating states or faults occur, as accurate quantitative fault models are rarely available. Simulation of qualitative models predicts all possible behaviors consistent with the model. However, with models on the scale of the controlled reactor used in the examples, the number of behaviors predicted becomes so large that it is impossible to actually produce all behaviors of some faults. Semi-quantitative simulation is one solution to these problems, as it eliminates branching based on what is quantitatively known about the plant, although it also required increased computations. There is ongoing work directed at further reducing the ambiguity of qualitative modeling 5]. Also, research is continuing in modeling more complex systems via qualitative reasoning, such as a nitric acid production plant 3] and a styrene polymerization system 20]. Although pure qualitative models and semi-quantitative models are quite similar, each has advantages and disadvantages. Purely qualitative models are easier to construct and simulate once constructed, although they provide only qualitative predictions of the physical system. Semi-quantitative models provide quantitative range predictions which allow statistical comparisons to readings. However, the addition of quantitative range information to qualitative models increases the time spent building the model, and greatly increases simulation time. Contrary to expectation, the addition of quantitative information does not greatly improve the detection and diagnostic capability of the qualitative system. Since it is easy to devise systems where the semi-quantitative models can distinguish faults which are qualitatively identical, it remains an empirical question when the additional information will be useful. Both the QMI and Qmimic algorithms are meant to suggest the structures of possible monitoring and diagnosis systems. The particular parts of the algorithms can be changed to test …