Direct identification of nonlinear parity relations. A way around the robustness problem

Model-based diagnosis is naturally sensitive to modeling errors and the available robust techniques are subject to limitations. Therefore, it may be better to seek more accurate (nonlinear) models, in a diagnostic framework which does not imply algebraic manipulations. This is possible with structured parity relations whose underlying nonlinear models may be directly identified from plant data. The paper discusses ways of selecting the structure for such models, including a minimax optimization procedure, and addresses the issues of causality and stability