Only One Behavior at the Cost of Specifying Exact Equational Forms for Faults and Requiring Parameter Fault Detection and Diagnosis Using Qualitative Modelling and Inter- Pretation. in Proc. Ifac Symposium on On-line Fault Detection and Supervision in the Chemical

estimation. APE's solution to the parameter estimation problem is to estimate only those parameters associated with the proposed fault, thereby reducing the computation and focussing fault diagnosis. Within APE, Qmimic and QMI fault diagnosis can be accomplished in several ways, and we have examined two methods: deening all fault models in advance (QMI), or creating new models based on the faults (Qmimic). In both cases, diagnosis would be achieved by selecting the models which best t the data. In the rst method, one must identify all possible fault modes of the plant and supply full parametric models to the fault diagnosis phase. This type of algorithm will work only when all faults are known beforehand. The second method hypothesizes new models, based on the fault and a ow sheet of the plant, which allows for the possibility of unknown or unexpected faults. APE currently uses predeened fault models of the plant. Work is underway to allow it to dynamically generate fault models, as Qmimic does. APE is a blend of statistical model selection with model-based diagnosis techniques, providing an integrated chemical plant monitoring and diagnosis system. APE performs as well as Qmimic and QMI in terms of accuracy and is much faster at detecting faults. Future work with APE includes monitoring during process changeover and examining larger chemical plants. APE Qmimic QMI Models used to Quantitative Qualitative with Purely predict behavior quantitative ranges qualitative Fault detection Student's t Student's t on Fuzzy logic method both ranges and membership slopes slopes Fault diagnosis Multiple Multiple cascading Single faults capability cascading faults deened a priori Table 1: Main diierences between APE, Qmimic and QMI In this example, the temperature sensor is the rst to signal that a change has occurred at 1.05 hours. From the possible faults, two are selected as the most likely: temperature controller failed closed and feed temperature disturbance. Both are estimated to have started at 1.01 hours and the temperature disturbance is estimated to be a 0.2% increase. As can be seen in Figure 3, the temperature disturbance is ruled out at t=1.235 hours because the observed change in temperature is not reproduced by that model. The model continues to track the readings fairly well, although there is some bias due to mismatch between the model and the plant. This does not cause the model to be rejected because the allowable error increases in proportion to the deviation …