Accurate soil engineering classification is fundamental to developing reliable subsurface stratigraphy, which underpins the safe design of underground projects. Traditional approaches rely on laboratory testing of borehole samples, which lead to real-time decision making but require extensive time and human labor. The cone penetration test (CPT) has long been widely used in geotechnical investigations and offers continuous data for timely soil properties assessment. Traditional machine learning-based clustering algorithms that have gained traction in soil classification can be a good choice to address CPT data. However, previous applications can hardly deal with the nonlinearity of input features and often ignore the correspondence between physical and mechanical properties. This study proposes a semisupervised deep embedded clustering (SDEC) framework that integrates CPT data with some critical soil classification parameters to capture both composition and mechanical characteristics. Initially, a generative adversarial network is employed to synthesize and enrich CPT data, thus alleviating the class imbalance in CPT datasets. Based on clustering results, a novel classification index Is using the elliptical arc function is then proposed to develop new soil classification charts and optimize decision boundaries beyond the traditional soil behavior type chart, which are more intuitive and user-friendly. The SDEC model is examined on three open datasets and one private dataset, which proves robust across diverse field conditions. A benchmarking study shows that the resulting charts provide more accurate interpretation of soil, further demonstrating their applicability.
Zhou et al. (Sat,) studied this question.