Model-Based Data Interpretation and Diagnosis Robustness

Model-based data-interpretation techniques are widely used to identify the behavior of structures. These techniques exploit information provided by in-situ measurements to make diagnosis related to structural performance. In the context of model-based infrastructure diagnoses, measuring and modeling uncertainties are generally estimated using engineering heuristics. The accuracy of diagnosis is related to the accuracy of uncertainty estimation. For civil infrastructure diagnoses, the estimation of modeling uncertainties is non-trivial, because special attention is required to avoid diagnosis errors. For data interpretation methodologies that generate multiple candidate models, a diagnosis error occurs when \emph{incorrect models are accepted, while the correct model is rejected}. The probability of diagnosis error is sensitive to two factors: (1) misevaluation of uncertainties and (2) the number of measurements used for data interpretation. In this context, the robustness is defined as the ability of providing the right diagnosis in presence of misevaluation of uncertainties. This paper presents a preliminary study that quantifies the sensitivity of diagnosis to errors with respect to misevaluation of uncertainties and the number of measurements used. The study of a beam example shows that when the mean of uncertainty is misestimated, the probability of diagnosis error increases with the number of measurements. Inversely, when the uncertainty standard deviation is underestimated, the probability of diagnosis error decreases with the number of measurements. For the case where both uncertainty mean and standard deviation are misevaluated, it is possible to find a minimum number of measurements that assures the diagnosis robustness.