Abstract

Conventional geotechnical soil classifications aim to classify soils into families with geotechnical characteristics and therefore similar behaviour; however, they require core samples and laboratory identification testing. Several empirical systems for estimating the nature of the soils have been developed on the basis of several in situ geotechnical tests. However, at present these systems remain empirical and they are often only used on an indicative basis. The objective of this article is, based on the analysis of dynamic penetrometric signals, to develop a methodology able to provide an estimate of the nature of the soil crossed. The methodology developed provide an automatic classification based on artificial neuron networks (ANNs) tools. Two types of ANN architectures were considered: multi-layer feedforward perceptron (MFP) and probabilistic neural network (PNN). The learning of these two tools was achieved through a base carried out in the laboratory and in situ. Both classification models were then tested in blind conditions and showed a good efficiency for calibrated soils and promising results for in situ soils.