Liquefaction Susceptibility of Soil Based on In-Situ Tests Using Multi Objective Feature Selection (MOFS)

Abstract

Soil liquefaction is a phenomenon when there is a sudden loss of strength and stiffness in saturated or partially saturated soils in response to cyclic loads (usually earthquakes). Earthquakes such as Niigata, 1964; Alaska, 1964; Loma Prieta, 1989; Great Hanshin, 1995; etc made the whole world to sit and take notice of it. Liquefaction potential of a soil subjected to a given seismic loading is an important first step towards mitigating liquefaction induced damage. Field techniques commonly used to assess triggering of seismic soil liquefaction are standard penetration test (SPT), cone penetration test (CPT), shear wave velocity (Vs), Becker hammer test (BHT). Liquefaction assessment correlations based on in situ penetration index tests are more widely used in engineerin practice to estimate the potential for triggering or initiation of seismically-induced soil liquefaction.

Class imbalance is important part of model building that plays a significant impact on the predictive capacity of classification models. In this research, liquefaction potential of soil is evaluated using multi objective feature selection (MOFS) algorithm based on three types of in-situ tests standard penetration test (SPT), cone penetration test (CPT) and shear wave velocity (Vs) test.

Feature subset selection involves minimisation of both prediction error and the number of features which are usually mutually conflicting objectives. In this research multi-objective optimisation technique, non-dominated sorting genetic algorithm (NSGA II) and multiobjective symbiotic optimisation search algorithms (MOSOS) are combined with learning algorithms, artificial neural network (ANN) and multi variate adaptive regression spline (MARS) to get a Pareto optimal set. In this thesis it is shown how NSGA II and MOSOS can be effectively deployed for selection of optimal parameters and simultaneously minimizing the error rate.