Forecasting of pile plugging using machine learning

Plugging of open-ended pipe piles affects the drivability and the capacity of the pile. Depending on their geometrical and geotechnical properties, as well as the driving dynamics, piles can be Plugged, Unplugged, or Partially Plugged. However, forecasting the plugging condition confidently, ahead of driving, has not been yet established, often resulting in driving difficulties. In this study, ten Machine Learning (ML) algorithms, namely: (i) Support Vector Machine (SVM); (ii) Logistic Regression; (iii) k-Nearest Neighbor; (iv) Decision Tree; (v) Random Forest; (vi) AdaBoost; (vii) XGBoost; (viii) Multi-Layer Perceptron (MLP); (ix) Hard Voting Classifier; and (x) Soft Voting Classifier, were explored in an effort to forecast the plugging condition using a dataset of 144 open-ended pipe piles. The calculated capacities were determined using four methods. These capacities were compared to measured (interpreted) capacities from static load tests for the purpose of identification of plugging in load tests where the information was not recorded. Sixteen different combinations that encompass geometric properties (diameter (D), length (L), L/D ratio, thickness) and geotechnical properties (predominant soil type, soil at pile toe, number of soil layers in the profile) for the piles were used as inputs for the ten ML algorithms. The highest achieved accuracy in identifying presumed plugging was 74% using a Soft Voting Classifier with the diameter, L/D ratio, and thickness as inputs. The model was finally tested using 24 published load tests and achieved similar accuracy.