A Machine Learning-Based Method for Predicting End-Bearing Capacity of Rock-Socketed Shafts

Abstract

This paper presents a machine learning (ML)-based method for predicting the end-bearing capacity of rock-socketed shafts. For ML model training and testing, a database of 151 test shafts covering a wide range of rock types, shaft dimensions, and ground profiles has been developed from various sources. To properly take into account different factors, the rock property constant ${m}_{i}$, unconfined compressive strength of intact rock ${\sigma }_{c}$ (MPa), geological strength index GSI, length of the shaft within the soil layer ${H}_{s}$ (m), length of the shaft within the rock layer ${H}_{r}$ (m), and shaft diameter $B$ (m) were taken as the inputs and the ultimate bearing capacity factor ${N}_{\sigma }$, which is the ratio of ultimate end-bearing capacity to ${\sigma }_{c}$, was taken as the target output. Four commonly used ML algorithms, support vector machine (SVM), decision trees (DT), random forest (RF), and Gaussian process regression (GPR), were first utilized to train models, respectively. Then, the trained models with the four ML algorithms were fused together with an ensemble learning (EL) approach to further enhance the prediction accuracy. Comparisons with existing empirical equations show a much better performance of the ML-based method for predicting the end-bearing capacity of rock-socketed shafts. Parametric studies were also performed with the EL model to investigate the importance of the six input parameters and the results show that the most important parameter is ${\sigma }_{c}$, followed by B, GSI, ${H}_{r}$, ${H}_{s}$ and ${m}_{i}$ in the order of importance. For the convenient application of the ML-based method, a graphical user interface (GUI) app has been developed. Finally, two examples were analyzed to demonstrate the application of the GUI app with the implemented EL models. The results show that the GUI app can be used for quick and accurate prediction of the end-bearing capacity of rock-socketed shafts by considering the various parameters.

Highlights

• Four machine learning algorithms are fused together with an ensemble learning approach to predict the ultimate bearing capacity of rock socketed shafts.

• The proposed ensemble learning model outperforms other existing empirical methods.

• For the convenient application of the ensemble learning-based method, a graphical user interface app has been developed.