Bayesian Inference of Piping Model Uncertainties Based on Field Observations

Abstract

This paper presents a Bayesian model to determine the model uncertainty of a critical horizontal gradient model for piping for dikes, such a Lane and Bligh. A Bayesian model is needed for two reasons. First, there is a large overlap in cases that failed and survived. Second, the evidence of the failed cases is limited .The model consists of a non-informative prior that is combined with likelihood functions for failed and survived cases. This involves modeling the mean and standard deviation of the model uncertainty as random variables. For survived cases we know the limit state function was larger than 0 for the observed water level. For failed cases we know the limit state function was smaller than 0; or Z = 0; which is a less conservative assumption. This information is used to determine the likelihood functions for failed and survived cases. The prior and likelihoods are combined to find the posterior distributions of the mean and standard distribution of the model uncertainty. Using integration, this finally results in the (lognormal) distribution of the model uncertainty. The model is applied to the data of Bligh and Lane and shows both a high mean and high standard deviation of the model uncertainty, where the model of Lane performs better than Bligh. It is recommended to tailor the proposed model to dikes by making a different distinction between horizontal and vertical erosion. Furthermore, it is recommended to apply the model to more dike specific data since the Bligh data mainly consists of dams instead of dikes.