Weak adhesion detection – Enhancing the analysis of vibroacoustic modulation by machine learning.

• Robust and reproducible artificial weak-bond creation in single-lap shear specimens. • Combination of vibroacoustic modulation analysis with machine learning to identify weak-bonds. • Improved classification of weak-bonds using a neural network resulted in 97% accuracy and prediction of shear strength by the neural network results is in general possible. • Neural network decisions are based on modulations traditionally not evaluated in previous approaches. Adhesive bonding is a well-established technique for composite materials. Despite advanced surface treatments and preparations, surface contamination and application errors still occur, resulting in localised areas with a reduced adhesion. The dramatic reduction of the bond strength limits the applicability of adhesive bonds and hampers further industrial adaptation. This study aims to detect weak-bonds due to manufacturing errors or contamination by analysing and interpreting the vibroacoustic modulation signals with the aid of machine learning. An ultrasonic signal is introduced into the specimen by a piezoceramic actuator and modulated through a low frequency vibration excited by a servo-hydraulic testing system. Tested samples are single-lap shear specimens, according to ASTM D5868-01, with artificial circular debonding areas introduced as PTFE-films or a release agent contamination. It is shown that an artificial neural network can identify various defects in the bonded joint robustly and is able to predict residual strengths and hence demonstrates great potential for non-destructive testing of adhesive joints. [ABSTRACT FROM AUTHOR]