Improving decision making in fault detection and isolation using model validity

Fault detection and isolation commonly relies on the computation of numerical residuals deduced from the analytical model of the process to be diagnosed. This model can be more or less valid, due to the difficulty of getting good data sets in industry and to the process ageing. If this validity is not taken into account during the evaluation of the residuals, false symptoms may be produced and consequently a false diagnosis. The aim of this paper is to explicitly take into account the validity of the model to make the diagnostic decision. The decision formulation is inspired by the logical approach to consistency-based diagnosis proposed by the artificial intelligence community. However, in this paper fuzzy aggregation is used because it allows gradual diagnostic decision making. A fuzzy indicator of the model validity is first evaluated a priori during the model identification and further modified on-line to take into account actual operating conditions. The numerical residuals are fuzzified, taking into consideration their dynamical evolution. Finally, the model validity is aggregated on-line with the fuzzy residuals in order to provide fuzzy symptoms and a diagnostic decision that reflects both the residual value and the model validity. Academic examples illustrate the benefits of the proposed method.