Prediction of Cone Penetration Test Data from Standard Penetration Test Attributes using Support Vector Regression Neural Networks

Abstract

Although the cone penetration test (CPT) is essential for geotechnical design, it is often not economically feasible and time-consuming to perform the tests at all sites. This work used data from Port Sudan City to create a model to predict CPT values from Standard Penetration Test (SPT) attributes using a Support Vector Regression (SVR) algorithm. This approach converted the SPT into attribute combinations through a stepwise regression analysis. Subsequently, the attributes were converted to CPT by training the SVR neural network with the available CPT data. Cross-validation was performed to assess the credibility of the SPT parameters in the CPT transformation. In this procedure, a quarter of the data set is omitted from the training set each time, and the transformation is recalculated. Next, the accuracy of the transformation in estimating CPT from the deleted data is evaluated; this procedure is applied to all data in the training set. The results show that comparing the actual and predicted CPT gives a good agreement, as indicated by high correlation coefficient values for the training and test data sets. This approach can be used for automated estimation of CPT from SPT parameters and can serve as an extension of conventional techniques.