Lye Tee Siang, Adolphus 
ORCID: 0000-0002-1803-8344
(2023)
Robust and Efficient Probabilistic Approaches towards Parameter Identification and Model Updating.
Doctor of Philosophy thesis, University of Liverpool.
|
Text
201367015_May2023.pdf - Author Accepted Manuscript Download (25MB) | Preview |
Abstract
In engineering, the virtual behaviour of structures under operational and extreme conditions are investigated using mathematical or physics-based models. To obtain numerical responses that best reflect the structure under investigation, the physical input parameters describing the geometric, material, and damping properties of these models need to be identified or inferred. However, the presence of uncertainty poses significant challenges in parameter identification. Often, these uncertainties would stem from the following: 1) the aleatory uncertainty due the variations in the response measurements of nominal identical structures under same loading conditions due to manufacturing and material variability, thus, leading to the parameter not having a single "true" parameter value representation; 2) the epistemic uncertainty associated with the "fuzziness" to the knowledge of the parameter(s) as a result of the experimental data/measurements being usually affected by "noise"; and 3) the model uncertainty due to the modelling errors associated with the failure of the model in capturing the physics of the problem. This presents the need to not only perform an inference on the parameter(s), but also quantify the uncertainty associated with the estimates. An approach towards this would be Bayesian model updating, which serves as the context of this dissertation. The dissertation provides details to the efficient and robust approaches towards probabilistic parameter identification and model updating via the aforementioned approach. To realize this, an extensive literature review on Bayesian inference and the existing sampling tools is provided. This is done to identify the key research gaps, as well as limitations to the current sampling algorithms. From there, the Transitional Ensemble Markov Chain Monte Carlo sampler is proposed to which its strengths include its robustness in sampling from skewed distributions, quicker computational time, and the removal of any need for tuning by the users. To demonstrate this, the algorithm has been implemented on both numerical and real-world examples. The latter involves a structural health monitoring problem and the recent NASA-Langley Uncertainty Quantification challenge. Following which, the analysis is extended towards inferring time-varying parameter(s) via on-line Bayesian inference. This motivated the development of the Sequential Ensemble Monte Carlo sampler to which its strengths include its robustness in identifying the most probable Markov kernel under uncertainty. Such strengths are demonstrated through the experimental example involving a single-storey structure subjected to a time-varying Coulomb friction. Finally, the dissertation presents an approach to merge Artificial Intelligence tools with Bayesian statistics towards the probabilistic prediction of material properties for Nuclear power plant structures. Such development seeks to enable the Artificial Intelligence models to provide a more robust probabilistic prediction on the material properties under very limited data and model uncertainty. For the interest of the relevant practitioners, the algorithms to the proposed methods presented in the dissertation are made accessible on OpenCOSSAN, an open-source software for uncertainty quantification, as well as GitHub.
| Item Type: | Thesis (Doctor of Philosophy) |
|---|---|
| Divisions: | Faculty of Science and Engineering > School of Engineering |
| Depositing User: | Symplectic Admin |
| Date Deposited: | 02 Jun 2023 13:11 |
| Last Modified: | 02 Jun 2023 13:12 |
| DOI: | 10.17638/03170546 |
| Supervisors: |
|
| URI: | https://livrepository.liverpool.ac.uk/id/eprint/3170546 |
Dimensions
Dimensions
