Abstract:
Many empirical formulas and models for soil classification based on Cone Penetration Test (CPT) data have been developed. However, the conclusions from different scholars vary in terms of the classification model of soils and the distinguishment of soil types due to the differences in geographical locations, standards and research objectives. In this study, the CPT data of 60 boreholes obtained in the Yinggehai Basin in the South China Sea and the samples taken from the adjacent boreholes are studied. Firstly, based on the data of boreholes located nearby, the CPT data corresponding to five common types of soils (i.e. clay, silty clay, silt, silty sand and fine sand) are selected and then processed interpretively. Secondly, soil behavior classification models suitable for using in the Yinggehai Basin are built up by using the normalized frictional ratio and the normalized cone tip resistance as the classification indicators and applying various machine learning techniques including Random Forest, Artificial Neural Networks, Decision Trees, Bayesian Learning, Support Vector Machines and Linear Perceptions. Thirdly, the metrics such as confusion matrix, overall accuracy and Kappa coefficient are used to evaluate the classification accuracy of the models. Finally, the developed models are applied to automatically interpret CPT data in the surrounding and distant sea areas, and then validated using borehole samples adjacent to the CPT data. The results indicate that the soil behavior classification models developed in this study demonstrate high accuracies in interpreting CPT data in the northern part of the South China Sea, providing valuable insights for geotechnical engineering applications in offshore wind energy projects and other marine construction activities.