Efficient Detection of Defective Parts with Acoustic Resonance Testing Using Synthetic Training Data

Abstract

Analyzing eigenfrequencies by acoustic resonance testing enables a fast screening of components regarding structural defects. The eigenfrequencies of each specific part depend on the general geometric and material properties, including tolerable part-to-part variations, as well as on possible structural flaws. Separating good parts from defective ones is not straightforward and each application-specific sorting algorithm is usually found from experimental training data. However, there are limitations and training data collection may be intricate. We worked on this challenge focusing on machine-made model parts varying slightly in geometry. The application objective was the eigenfrequency-based detection of parts featuring a through-hole test defect drilled into some of the parts and enlarged stepwise. The eigenfrequencies were measured concomitantly. Unlike the industry standard, our approach is based on synthetic training data created mainly by simulation techniques, which resulted in a principally satisfactory classification of the good and defective parts. However, the parts with small defects were not identified from the eigenfrequencies alone, due to overlaying geometric variations. In order to counteract such noise and to improve defect detection based on synthetic training data, the specific actual part geometry was used, in the sense of additional a priori information. A multimodal data evaluation model showed a clearly enhanced sorting power.