Changepoint analysis using Gaussian process regression: a Bayesian statistical model for the identification of lithological strata in geotechnical engineering
File(s)
Author(s)
Bendel, Jens
Type
Thesis
Abstract
Around the Earth's surface it is common for the ground to be made up of layers of different soil stacked on top of each other. This composition of layers can for example be the result of marine deposition over the course of thousands or millions of years. The field of geotechnical engineering studies these lithological layers and their effect on engineering work such as tunnelling or the construction of buildings and bridges. A standard procedure to determine a segmentation of the ground into individual layers is to take a drilling core and examine slices sampled at various depths. Typical methods to distinguish soil layers are based on microfossil content or measurements of physical quantities such as the water content of the soil. Common practice is the inspection of these properties by eye using expert engineering judgement. Such approaches lack scientific rigour and fail to address the uncertainty that is inherent to any such analysis.
This thesis discusses statistical methodology (changepoint analysis) in order to propose a reproducible scientific approach for the identification of soil layers based on measurements of water content. With a focus on uncertainty quantification the proposed approach combines a Bayesian changepoint method with a Gaussian process regression model for each soil layer. This Gaussian process changepoint method is applied to data from the construction site of underground railway tunnels. The method correctly identifies well-established layers while suggesting that a previously proposed subdivision of a particular layer (A3) is not supported by the dataset.
Further results indicate that more research work is needed with regard to the collection of data as well as the development of statistical methodology. Overall, this thesis shows that the proposed focus on mathematical rigour and uncertainty quantification is very much needed.
This thesis discusses statistical methodology (changepoint analysis) in order to propose a reproducible scientific approach for the identification of soil layers based on measurements of water content. With a focus on uncertainty quantification the proposed approach combines a Bayesian changepoint method with a Gaussian process regression model for each soil layer. This Gaussian process changepoint method is applied to data from the construction site of underground railway tunnels. The method correctly identifies well-established layers while suggesting that a previously proposed subdivision of a particular layer (A3) is not supported by the dataset.
Further results indicate that more research work is needed with regard to the collection of data as well as the development of statistical methodology. Overall, this thesis shows that the proposed focus on mathematical rigour and uncertainty quantification is very much needed.
Version
Open Access
Date Issued
2019-11
Date Awarded
2020-04
Copyright Statement
Creative Commons Attribution NonCommercial NoDerivatives Licence
Advisor
Barnett, Ryan
Publisher Department
Mathematics
Publisher Institution
Imperial College London
Qualification Level
Masters
Qualification Name
Master of Philosophy (MPhil)