Use this URL to cite or link to this record in EThOS:
Title: Towards Bayesian system identification : with application to SHM of offshore structures
Author: Rogers, Timothy J.
ISNI:       0000 0004 7651 7962
Awarding Body: University of Sheffield
Current Institution: University of Sheffield
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
Within the offshore industry Structural Health Monitoring remains a growing area of interest. The oil and gas sectors are faced with ageing infrastructure and are driven by the desire for reliable lifetime extension, whereas the wind energy sector is investing heavily in a large number of structures. This leads to a number of distinct challenges for Structural Health Monitoring which are brought together by one unifying theme --- uncertainty. The offshore environment is highly uncertain, existing structures have not been monitored from construction and the loading and operational conditions they have experienced (among other factors) are not known. For the wind energy sector, high numbers of structures make traditional inspection methods costly and in some cases dangerous due to the inaccessibility of many wind farms. Structural Health Monitoring attempts to address these issues by providing tools to allow automated online assessment of the condition of structures to aid decision making. The work of this thesis presents a number of Bayesian methods which allow system identification, for Structural Health Monitoring, under uncertainty. The Bayesian approach explicitly incorporates prior knowledge that is available and combines this with evidence from observed data to allow the formation of updated beliefs. This is a natural way to approach Structural Health Monitoring, or indeed, many engineering problems. It is reasonable to assume that there is some knowledge available to the engineer before attempting to detect, locate, classify, or model damage on a structure. Having a framework where this knowledge can be exploited, and the uncertainty in that knowledge can be handled rigorously, is a powerful methodology. The problem being that the actual computation of Bayesian results can pose a significant challenge both computationally and in terms of specifying appropriate models. This thesis aims to present a number of Bayesian tools, each of which leverages the power of the Bayesian paradigm to address a different Structural Health Monitoring challenge. Within this work the use of Gaussian Process models is presented as a flexible nonparametric Bayesian approach to regression, which is extended to handle dynamic models within the Gaussian Process NARX framework. The challenge in training Gaussian Process models is seldom discussed and the work shown here aims to offer a quantitative assessment of different learning techniques including discussions on the choice of cost function for optimisation of hyperparameters and the choice of the optimisation algorithm itself. Although rarely considered, the effects of these choices are demonstrated to be important and to inform the use of a Gaussian Process NARX model for wave load identification on offshore structures. The work is not restricted to only Gaussian Process models, but Bayesian state-space models are also used. The novel use of Particle Gibbs for identification of nonlinear oscillators is shown and modifications to this algorithm are applied to handle its specific use in Structural Health Monitoring. Alongside this, the Bayesian state-space model is used to perform joint input-state-parameter inference for Operational Modal Analysis where the use of priors over the parameters and the forcing function (in the form of a Gaussian Process transformed into a state-space representation) provides a methodology for this output-only identification under parameter uncertainty. Interestingly, this method is shown to recover the parameter distributions of the model without compromising the recovery of the loading time-series signal when compared to the case where the parameters are known. Finally, a novel use of an online Bayesian clustering method is presented for performing Structural Health Monitoring in the absence of any available training data. This online method does not require a pre-collected training dataset, nor a model of the structure, and is capable of detecting and classifying a range of operational and damage conditions while in service. This leaves the reader with a toolbox of methods which can be applied, where appropriate, to identification of dynamic systems with a view to Structural Health Monitoring problems within the offshore industry and across engineering.
Supervisor: Cross, Elizabeth J. ; Manson, Graeme ; Worden, Keith Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available