ABSTRACT
Vertical and horizontal scales of fluctuation are measures of spatial correlation and variability in soils, and as such are extensively used in the modelling of soils in reliability methods, such as the Random Finite Element Method (RFEM). These parameters are conventionally estimated from site surveys, commonly using Cone Penetration Test (CPT) data. By fitting theoretical correlation functions to the site data, the horizontal and vertical scales of fluctuation can be estimated. Presented is a new approach that trains a Convolution Neural Network (CNN) with pseudo CPT data taken from generated Random Field (RF) data with known scales of fluctuation. Once trained the network can predict these measures of spatial variability from real CPT data. This paper presents the results of a study training a network to predict vertical scales of fluctuation.
