Non-Destructive Inspection of Tile Debonding by DWT and MFCC of Tile-Tapping Sound with Machine versus Deep Learning Models

Abstract

One of the essential processes of construction quality control is tile bonding inspection. Hollows beneath tile tessellation can lead to unbounded or completely broken tiles. An interior inspector typically used a hollowsounding technique. However, it relies on skill and judgment that greatly vary among individuals. Moreover, equipment and interpretation are difficult to calibrate and standardize. This paper addresses these issues by employing machine-learning strategies for tile-tapping sound classification. Provided that a tapping signal was digitally acquired, the proposed method was fully computerized. Firstly, the signal was analyzed and its wavelets and MFCC were extracted. The corresponding spectral features were then classified by SVM, k-NN, Naïve Bayes, and Logistic Regression algorithm, in turn. The results were subsequently compared against those from a previous works that employed a deep learning strategy. It was revealed that when the proposed method was properly configured, it required much less computing resources than the deep learning based one, while being able to distinguish dull from hollow sounding tiles with 93.67% accuracy.