Identification of non-dimensional density distribution in concrete by convolutional neural network using impact response waveforms

Abstract

In this paper, we present a method for identification of defects in concrete using machine learning. The time history data of acceleration of surface vibration obtained by hammering test is employed as learning data, and a convolutional neural network, which is generally utilized in image recognition is applied to estimate defect shape. Since information of relative position can be held in the convolutional neural network, it appears that this method is suitable for the position estimation of defects in concrete using the information from multiple sensors as input. In addition, a concrete floor plate is represented by a non-dimensional density matrix, so that both defective and healthy areas are treated in a unified manner and can be easily handled by a machine learning model. By delimiting the non-dimensional density matrix to a certain size and estimating the columnar non-dimensional density distribution, the data set can be constructed using the same method from floor plates of different sizes in the x and y directions. We challenge the estimation of 3D topology (position and location) of the internal defects in concrete plates, and some numerical results and considerations are shown in this paper. We also perform preprocessing on the data set and discuss the change in accuracy with and without preprocessing.

Numerical studies for defect topology identification based on the adjoint variable and the finite element methods

“…al., 2019). In addition, in the paper (Shimada et. al., 2021) Kurahashi, Takeuchi, Koike, Kishida, Murakami and Ikeda, Mechanical Engineering Journal, Vol.10, No.3 (2023) [DOI: 10.1299/mej.23-00090] We first investigated the relationship between the regularization parameter 𝜏𝜏 and identified defect topology in a 3D computational model.…”

“…Since it is difficult to set the hammering point and the observation point at the same point, the observation point is set a little away from the hammering point (c.f. (Shimada et. al., 2021)).…”