| Resumo : |
Classical methods of soil classification from cone penetration test (CPT) data propose a graphical approach. It requires representing data with just two parameters in order to create plane representations. To overcome this restriction, the fundamental objective of this work is to develop multi-dimensional analyses using different machine learning (ML) algorithms, including supervised and unsupervised techniques. A general methodology for the predictive supervised approach is proposed and applied up to five dimensions, including raw, non-normalized and normalized CPT data as continuous inputs and geological age as a discrete one. The dataset is composed by 179 CPT soundings from several world locations. Two popular chart-based classification methods are reproduced and studied in the predictive supervised approach: one influenced by soil granulometry (ISG) and one focused on soil behavior (FSB). Descriptive statistical procedures as well as ML techniques are used for data preprocessing, including data cleaning, balancing and transformation. Different ML techniques are calibrated, studied and compared with statistical tests and employing the 10-fold cross-validation procedure. Within the first studies, the input features relevance is evaluated with distance-based techniques. These first results have shown that using non-normalized CPT data with depth included is enough for the techniques to properly distinguish between different soil types with a satisfactory accuracy, around 90%, and that depth introduce relevant information for the task of soil classification from CPT data. It was also observed that the inclusion of the geological age as a discrete input can introduce relevant information. In the second part of the study, other ML techniques are tested and their particularities are explored. The symbolic techniques were those with the best performance and lowest sensitivity to methodology variations. The techniques are then combined into a multiple model predictor (MMP), whose performance overcame the one of the other ML techniques according to statistical tests. Finally, unsupervised techniques are employed to generate new classes from non-normalized inputs with two different approaches: a general one and a case-based one. So, an ANN is trained to create a mathematical model that allows the new method reproduction with simple tools as spreadsheets. The clustering results are then compared to the classical methods and to standard penetration test (SPT) data paired with CPT. |