Improvement of impact-echo method by applying image recognition of sound spectrogram

Abstract

The authors are investigating a method for lifetime prediction of concrete structures by a comprehensive judgment of survey results and information on design and environment. In this paper, we tried to apply impact-echo acoustic spectrogram to image recognition by machine learning in order to have a better estimation of defects in concrete. impact-echo sound sampling was conducted on specimens containing variated voids in depth and diameter. Each sound spectrogram was transformed into 28 by 28 pixels grayscale picture. As a result of the study, it was confirmed that the estimation by machine learning showed competitive accuracy as estimation by a human expert. Also, it was shown that the depth and diameter of the void are predictable under certain conditions. Furthermore, the authors showed the possibility of this method in actual concrete structure.