Pile Damage Detection Using Machine Learning with the Multipoint Traveling Wave Decomposition Method

Abstract

The in-hole multipoint traveling wave decomposition (MPTWD) method is developed for detecting and characterizing the damage of cast in situ reinforced concrete (RC) piles. Compared with the results of MPTWD, the results of the in-hole MPTWD reconstruction technique are found ideal for evaluating the lower-part pile integrity and are further utilized to establish a data-driven machine-learning framework to detect and quantify the degree of damage. Considering the relatively small number of field test samples of the in-hole MPTWD method at this stage, an analytical solution is employed to generate sufficient samples to verify the feasibility and optimize the performance of the machine learning modeling framework. Two types of features extracted by the distributed sampling and statistical and signal processing techniques are applied to three machine-learning classifiers, i.e., logistic regression (LR), extreme gradient boosting (XGBoost) and multilayer perceptron (MLP). The performance of the data-driven machine-learning framework is then evaluated through a specific case study. The results demonstrate that all three classifiers perform better when employing the statistical and signal processing techniques, and the total of 24 extracted features are sufficient for the machine-learning algorithms.