Machine Learning of Ultrasonic Data for Expansion Prediction of Concrete with Alkali-Silica Reaction

Abstract

Ultrasonic nondestructive testing is a promising method for performing damage assessments on concrete subjected to alkali-silica reactions (ASRs). Previous research incorporated only some ultrasonic wave parameters, and the other information from the ultrasonic signals was discarded. In this work, 13 features, including wave velocity and wavelet features, were extracted from the ultrasonic signals. A curve-fitting method was used to fit a polynomial relationship between the wave velocity and expansion of one concrete sample subjected to ASR to predict the expansion of another concrete sample subjected to ASR. Support vector regression (SVR), a machine learning model, was trained using all 13 features derived from the ultrasonic data obtained from the ASR samples. The SVR was then tested using the datasets from the ASR-2D sample. The performance showed that the curve-fitting method and the SVR had poor prediction results on the expansion of the ASR-2D sample. With feature selection, the performance of the SVR model using six selected features was significantly improved.