Prediction of load-carrying capacity of piles using a support vector machine and improved data collection

Abstract

Model development for the prediction of the axial load carrying capacity of piles, at least at the model verification stage, relies on the measured data at full scale. Artificial intelligence and machine learning approaches use data in the whole process of model development and verification, making it necessary to incorporate reliable and diverse data. This study aims to develop a more accurate model for predicting pile capacity based on cone penetration test and full scale static pile load test data, employing a support vector machine (SVM) technique. Furthermore, it draws on the concept of support vectors to make suggestions for compiling additional data that are more representative of the problem, leading to enhancing the accuracy of the future models. In fact, in models developed using the SVM technique, those samples within a dataset for which a model shows the greatest uncertainties are detected as support vectors and are the only data that contribute to model development. In previous studies an examination of the distribution of input parameter values and the concentration of support vectors in input intervals were considered as guidelines for collecting further data. Geotechnical problems, however, are more complicated in that the importance of each input in model development and also the interaction of those parameters should be taken into account. Therefore, this study employs a sensitivity analysis of the model input parameters and other statistical analyses to improve the existing support-vector based approaches to data collection.