Machine learning based evaluation of concrete strength from saturated to dry by non-destructive methods

Abstract

Machine learning (ML) techniques have been increasingly applied in various scientific fields, including non-destructive testing (NDT), to enhance efficiency and speed up data analysis. In this study, an approach that aims to predict the compressive strength and quality of concrete with NDT and ML algorithms is presented to provide a great advantage in terms of both time and cost and to shorten the long laboratory periods. In the study, we aimed to set up a laboratory environment for determining the strengths of concrete samples by using Schmidt hardness values from NDT, curing times, water contents, and ultrasonic velocities. The data obtained in the laboratory environment was subjected to machine learning algorithms in the next process. In the laboratory environment, which is the first stage of the study, because concrete can be found at a variety of humidity levels depending on where it is used, 63 concrete samples were exposed to curing for 7, 28, and 90 days. Then these samples were put through a pressure test and subsequently exposed to various moisture conditions, from saturated to dry. Thus, the moisture state of concrete samples was evaluated based on their dryness or saturation in tests conducted on concrete samples. Afterwards, the ultrasonic velocities and Schmidt hardness values of these samples were measured. In the second stage, the strengths of concrete samples were classified with LR, MLP, SVM, and k-NN algorithms, and high R values were achieved. As a result of the studies carried out, the best R value (0.89) was achieved in the k-NN algorithm. This study has demonstrated that concrete strength values may be approximated with the k-NN method at high levels using sparse data collected in a laboratory setting.