Identification analysis of defect topologies by self-attention based machine learning (Effect of the number of training data on identification accuracy)

Abstract

In this study, a method was developed for estimating defects in concrete from test data generated by hammering a concrete plate using machine learning. A neural network was constructed based on a self-attention network to estimate the three-dimensional position and size of the defects placed within the concrete plate. The scalograms generated from the acceleration responses were used as the input. Identification was also conducted using data augmentation, in which we evaluated the effect of the number of training data items on identification accuracy.