DWT/ MFCC Feature Extraction for Tile Tapping Sound Classification

Abstract

Tile tapping sound inspection is a process of construction quality control. Hollow sound, for instance, indicate low quality tessellation and thus voids underneath that could lead to future broken tiles. Hollow-sounding inspection was often carried out by construction specialists, whose skills and judgment may vary across individual. This paper elevates this issue and presents a Deep Learning (DL) classification method for computerized sounding tile inspection. Unlike other existing works in the area, where structural details were assessed, this study acquired tapping sound signals and analyzed them in a spectral domain by using Discrete Wavelet Transform (DWT) and Mel-frequency Cepstral Coefficients (MFCC). The dull versus hollow sounding tile were then classified based on these features by means of a Convolutional Neural Network (CNN). The experiments carried out in a laboratory tessellation indicated that the proposed method could differentiate dull from hollow-sounding tiles with very high accuracy up to 93.67%. The developed prototype can be used as guideline for devising a tiling inspection standard.