Application of Image Recognition Technology to Sound Spectrogram of Impact-Echo Method

Abstract

In this paper, as a part of the comprehensive deterioration evaluation of the concrete structure, a method of evaluating the impact sound by machine learning was studied. The recorded impact sound data are converted into spectrograms, and the images are recognized by supervised learning, whereby the state of the defect in concrete. With concrete test specimens having voids in different depth (from 35 to 80 mm) and diameter (from 80 to 350 mm), the spectrograms of the Impact-Echo sound are taken and converted into grayscale pictures in different frequency-time range and resolutions. Neural networks were created with the grayscale images with supervised learning on the specification of voids, and the estimation results are evaluated. It was confirmed that the image recognition of spectrogram of Impact-Echo sound by machine learning could possibly estimate the state of the defects in concrete. Also, adequate specification of frequency-time range and image resolution were examined and discussed.