Time–frequency analysis of ultrasonic signals for quality assessment of bonded concrete

Abstract

In accelerated bridge construction (ABC), 3D-printed concrete and structural retrofitting using concrete jacketing are used to reduce the onsite construction time. Both techniques involve bonded concrete, where a concrete layer is cast on the old substrate. As a result, the interfacial transition zone between these concrete layers is the weakest part of the structure. In this paper, a machine learning-based model is proposed to predict the bond strength between concrete layers from extracted ultrasonic features. Ultrasound tests provide the input parameters, and bi-surface shear test results are the targets for the machine learning model. To validate the method, laboratory testing is performed on 54 concrete specimens with different design mixtures and bond zone conditions. Variational Mode Decomposition (VMD) is applied to the ultrasonic signals, and three intrinsic mode functions (IMFs) are extracted from each signal leading to 21 independent features. Four machine learning algorithms are deployed, namely K-nearest-neighbors, support vector machine, random forest, and XGB regressors. XGB is found to be the most accurate model, and its hyperparameters are tuned to optimize the model performance. To improve the method, the features’ importance values are derived, and the most compelling features for training are identified, which are the third quartile and center frequency of the first IMF’s instantaneous frequency, and the center frequency and kurtosis of the second IMF’s instantaneous frequency. The resulting final model was a tuned XGBoost regression model that obtained an R-squared value of 95.5 percent taking the four features as input. For the four selected features, apparent clustering was found at a bond strength of 3 MPa, ensuring a reliable prediction of quality bond conditions. These results demonstrate the effectiveness of the proposed model in predicting the bi-surface shear strength of concrete layers based on ultrasonic testing.