Deep Learning-Aided Acoustic Source Localization in Thin-Walled Waveguides

The timely detection of fault locations represents a critical task in Structural Health Monitoring (SHM) of thinwalled elements. In particular, the localization of acoustic emission sources is particularly important for the identification of damages caused by stress and can be achieved by estimating the Difference in Time of Arrival (DToA) between the waves captured by a sparse sensor array. In this work, a novel method for DToA extraction suitable for isotropic structures is proposed. Our approach is based on the combination of Convolutional Neural Networks (CNNs) and a dispersion compensation operator, the Warped Frequency Transform (WFT). CNNs are deployed to enhance the localization process against the detrimental losses caused by non-ideal conditions, such as the presence of reflections and multiple propagation modes. The results show that such method yields localization errors of a few centimetres, with an average below 2 cm, when only relying on hundreds of real data for training.