Predictive probability of detection curves for ultrasonic testing

Abstract

The probability of detection (POD) is one of the most meaningful ways to quantify the detectability of damage, because it considers the statistical variability in the measurements. Predictive probability of detection (P-POD) curves are particularly efficient, as they generate POD curves based on a series of measurements from undamaged specimens, without having to run a series of destructive tests. P-POD methods are model-assisted, but instead of generating synthetic data, a sensitivity matrix is extracted from the model, and the measurement uncertainties are quantified based on experimental tests. However, so far, they have only been applied to global structural health monitoring applications. This paper sets out to apply the P-POD method to ultrasonic testing for the first time, and to experimentally validate the predictions for contact-based and air-coupled measurements. For that purpose, a contact ultrasound study is carried out, where changes of circular reflectors from their nominal values are evaluated in a polyamide cuboid based on the maximum reflected wave amplitude. Moreover, an air-coupled ultrasonic test is performed to determine thickness changes of a carbon fiber-reinforced polymer plate based on time-of-flight considerations. In both cases, the predicted POD is compared with empirical tests. The results show that the P-POD accurately predicts the POD with very small absolute deviations between predicted and actual POD. Since P-POD methods require analytical models that relate measured damage indicators to material changes, a separate study is added to demonstrate how the model-based uncertainty can be quantified using confidence intervals, and how it can be distinguished from data-driven uncertainties. As mentioned above, the POD can be predicted based on undamaged specimens, and that is why the P-POD method is an appropriate tool to optimize user-defined signal processing parameters before scanning the objects. To demonstrate this, different onset pickers are analyzed in a comparative study, where the one based on the Akaike information criterion led to a POD up to 38% higher than other pickers.