Robust isolable models for failure diagnosis

The paper presents a design procedure to generate parity equation models for analytical failure detection and isolation. The models are in “statistically‐isolable” structure that prevents incorrect isolation decisions under marginal failure size if the residuals are tested in parallel. Of the combinatorial multitude of structurally‐suitable models, a choice is made on the basis of sensitivities with respect to different failures and robustness relative to uncertainties in selected parameters of an underlying model. A complete set of parity equations fitting the isolable structure is first generated from a state‐space model. The equations unsuitable for sensitivity are eliminated, and a suboptimally‐robust diagnostic model is obtained in a constrained discrete search over the remaining set.