Bayesian Data Analysis of Imperfect Information for Decision-Theoretic Approaches to Structural Integrity Management

Engineers are responsible for the safe and economic operation of structures, equipped only with somewhat imprecise, unreliable and incomplete data, which is generally very expensive to obtain. They have access to multiple models in industrial standards, but none will be entirely representative of the structure they are assessing. In this thesis, the novel application of partial pooling (multi-level and non-parametric) Bayesian methods are proposed as an improved approach for quantifying multi-variate uncertainty in engineering models. Where possible, information-theoretic and cross-validation criteria have been used to demonstrate the improvements in approximate (out of sample) predictive performance.Risk based structural integrity management can be framed as an application of decision making under uncertainty. Information (from inspections, monitoring and testing) will benefit decision making, but is only available at a cost and identifying when this investment is expected to be worthwhile is not necessarily intuitive. Value of information analysis is an established solution to this challenge, allowing engineers to identify and pursue information consistent with a risk optimisation. However, the underlying methods of uncertainty quantification can make better use of available data by partially pooling estimates of parameters, and characterising the inter-dependencies between marginal probabilistic models.Detailed in this thesis is the application of partial pooling models in estimating degradation rates (for fatigue crack and corrosion growth) and limit states (for crack failure assessment) and in relating inspection measurements across multiple locations in a structure (or group of structures). A partial pooling model has also been integrated with a value of information analysis in a method that can demonstrate the expected utility of prospective novel methods of data collection and analysis. These more representative model structures can better characterise system effects and therefore lead to improved quantitative (decision-theoretic) risk based integrity management of ageing structures, particularly those associated with imperfect and incomplete information