Hammering Tests of Tiles Fixed with Organic Adhesives with the aid of Machine Learning

The rubbing sounds of the tiles fixed on concrete bases with organic adhesives are analyzed with the aid of machine learning. We demonstrate that the the loose tiles (adhesion rates: 0% or 60%) and the secure tiles (adhesion rates: 80% or 100%) are distinguishable with no less than 95% of accuracy by virtue of the machine learning models build from the sounds, even though this task has not been considered feasible so far. We examine the generalization performance of the models in two cases. Using two different inspection rods, one to build the learning models and the other to generate the test data, we find that the accuracy turned out to be 90% or better. Limiting the training data to 0% and 100% specimens, we examine if the model can classify 60% and 80% specimens to find that the accuracy mostly falls within the range of 70%–90%, although it drops down to 50% in some cases. Admittedly leaving challenges, these results show the possibility of the extensive generalization of our simple learning model.