Contribution
Quantitative Acoustic Resonance Testing based on Synthetic Training Data
* Presenting author
Abstract:
Acoustic resonance testing (ART) is a nondestructive testing method based on the analysis of intentionally excited natural vibrations of the parts to be inspected. The method utilizes the fact that structural anomalies or macroscopic defects are reflected in the natural vibration characteristics of defective test objects. The corresponding test decisions are usually made with the help of a specific classifier, which has previously been derived on the basis of training data generated by measurements on representative parts. However, the experimental generation of suitable training data is often very time-consuming and, due to practical limitations, the amount of data sets may be too small, and the distribution of component properties may be poorly conditioned. This contribution addresses this problem with a concept based on synthetic training data that is primarily calculated using simulation software and subsequently transformed into the “measurement world” by incorporating some experimental reference data. The concept is demonstrated with the help of a fictional inspection task involving machine-made parts with artificial defects. It is shown that synthetic ART training data can not only result in reliable good/bad classifications, but also enable quantitative conclusions about the anomalies of faulty parts, which exceeds the usual performance spectrum of ART.