Onset time detection of acoustic emission signals for structural monitoring with deep learning

The Acoustic Emission (AE) technique is widely employed in structural engineering as a non-destructive testing method for detecting defects and monitoring structural health. AE captures signals emitted by structures due to damage, providing valuable insights into crack propagation and durability. This study focuses on the identification of the onset time of AE signals, a critical parameter for structural monitoring. Traditional signal analysis methods are often inadequate due to background noise similarities, necessitating novel approaches. In this study, the signal onset time identification is tackled by performing a one-dimensional segmentation task on the AE time series. To do so, a U-net is implemented, which is an artificial Neural Network (NN) specific for segmentation. The dataset comprises signals obtained from a Pencil Lead Break (PLB) test, closely resembling AE signals. The U-net assigns probabilities to each signal point, enabling accurate differentiation between the AE signal and background noise. Averaging probabilities with neighboring points mitigates false positives. The presented methodology offers real-time, continuous monitoring without the need for extensive preprocessing or temporal window separation, making it practical for AE-based structural health monitoring. The obtained results demonstrate that trained models achieve such precision in identifying the onset time that it enables locating the crack source with centimetric accuracy.