Cone penetration data classification with Bayesian Mixture Analysis

This paper presents an application of the Bayesian Mixture Analysis (BMA) to deal with the classification of spatially variable soil data from the cone penetration test. The cone penetration data classification postulates a problem where a set of cone penetration measurements is used to identify "hidden or unobserved" soil classes. The problem is formulated as an incomplete-data Gaussian mixture where the observed data are defined by the natural logarithm-transformed values of the normalized friction and the normalized cone resistance, while the soil classes to be identified are considered as hidden data. The solution for the incomplete-data problem which consists of class-dependent mixture probabilities and Gaussian distribution parameters is defined in a Bayesian framework. The implementation of conjugate priors for the Gaussian mixtures enables an efficient sampling of the posterior parameters by the Gibbs algorithm of the Markov Chain Monte Carlo method. When compared to the well-established Robertson classification charts, the BMA formulation has an advantage due to the Bayesian framework which enables the definition of soil classes through mixture priors, class-dependent posterior parameter estimates, and a probabilistic soil classification. The presented approach is applied to the cone penetration data from the Sheringham Shoal Offshore Wind Farm site.