Slope Stability Assessment with Bayesian Updating
Abstract
This thesis aims to explore the potential of implementing Bayesian updating in the probabilistic analysis of a geotechnical slope stability assessment. Bayesian updating is a well-defined method for combining supplementary information with already existing knowledge. In geotechnical slope stability assessment the previously existing knowledge would typically be a model of the slope, with material descriptions based on knowledge of the area and ground investigations. The supplementary information would typically be additional ground investigations, or observations of the performance of the slope. Including more information in the calculation model is expected to improve the accuracy of the predictions of slope behaviour. The model in this thesis is updated based on the observation that the slope is stable.
The Bayesian updating is implemented in two different analyses, where one models the spatial variability within each soil layer, while the other models each soil layer spatially homogeneous. The probabilistic analysis is implemented using Monte Carlo Simulations, and the slope stability assessment is conducted with an finite element model constructed in Plaxis 2D 2017.1. The shear strength of each material in the slope is modeled as a random variable. In the Single Random Variable (SRV) analysis these distributions are used to sample shear strength values that are applied to each layer. In the Conditional Random Field (CRF) analysis, these distributions form the basis for generating a random field where the shear strength of every element is modelled with a separate random variable.
Each of the analyses are implemented in three different simulations with different variance of the input shear strength. This is done to examine the effect of increased uncertainty in the input parameters on the estimated behaviour of the slope, and to investigate whether increasing the uncertainty leads to estimates that are more conservative. Each simulation provides distributions of both the Factor of Safety of the slope and the shear strength of the materials both prior and posterior to updating. These are compared to examine the effects of the updating, and of using various input and calculation models.
The results show that Bayesian updating is an efficient method for incorporating observations of the performance of slopes into a slope stability assessment. The probability of failure is reduced significantly, and the shear strength distributions are made more accurate by the updating. In the random field, the effects of the updating are mostly concentrated to zones close to the failure mechanisms, in layers with an inaccurate initial estimate of the shear strength. The SRV analysis predicts a response that deviates significantly from the response of the CRF analysis. This indicates that neglecting the spatial variability entails a simplification that reduces the quality of the results significantly. Lastly, the results show that increasing the Coefficient of Variation in the input parameters completely changes the estimated behaviour of the slope, rather than altering it slightly in a conservative direction. This is not a suitable method for obtaining conservative estimates.