Predictive probability of detection curves based on data from undamaged structures

Abstract

This paper develops a model-assisted approach for determining predictive probability of detection curves. The approach is “model-assisted,” as the damage-sensitive features are evaluated in combination with a numerical model of the examined structure. It is “predictive” in the sense that probability of detection (POD) curves can be constructed based on measurement records from the undamaged structure, avoiding any destructive tests. The approach can be applied to a wide range of damage-sensitive features in structural health monitoring and non-destructive testing, provided the statistical distribution of the features can be approximated by a normal distribution. In particular, it is suitable for global vibration-based features, such as modal parameters, and evaluates changes in local structural components, for example, changes in material properties, cross-sectional values, prestressing forces, and support conditions. The approach explicitly considers the statistical uncertainties of the features due to measurement noise, unknown excitation, or other noise sources. Moreover, through confidence intervals, it considers model-based uncertainties due to uncertain structural parameters and a possible mismatch between the modeled and the real structure. Experimental studies based on a laboratory beam structure demonstrate that the approach can predict the POD before damage occurs. Ultimately, several ways to utilize predictive POD curves are discussed, for example, for the evaluation of the most suitable measurement equipment, for quality control, for feature selection, or sensor placement optimization.