Volume 43, N. 4, October-December 2020 | PDF(39 downloads)
Most methods available in the literature for soil classification from cone penetration test (CPT) data define soil classes using laboratory tests. One disadvantage of this approach is that field soil conditions are difficult to replicate in a lab. The alternative adopted in this work is trying to define soil classes only by the similarity of the CPT measurements, using clustering. This study is the first, to the best knowledge of the authors, to cluster soil classes in a four-dimensional input feature space using measurements directly taken from the CPT experiment. Nine soil classes are produced from a general dataset containing 179 CPT soundings and, in a complementary study, four more specialized classes are obtained from 5 CPT soundings. Artificial neural networks (ANN) are used to produce simple models capable of reproducing both class groups, which are compared with classical soil classifications from the literature and with standard penetration test (SPT) samples. Results show that both general and specialized class groups can be reproduced by ANN although accuracy is better for the latter, reaching a 97.04 % accuracy with a standard deviation of 1.24 %. Furthermore, it is shown that accuracies above 80 % are obtained even if incomplete data is used. This shows that the here proposed soil classes can become an interesting alternative in engineering practice.