Civil infrastructure are critical elements to a society’s welfare and economic thriving. Understanding their behaviour and monitoring their serviceability are relevant challenges of Structural Health Monitoring (SHM). Despite the impressive improvement of miniaturisation, standardisation and diversity of monitoring systems, the ability to interpret data has registered a much slower progression across years. The underlying causes for such disparity are the overall complexity of the proposed challenge, and the inherent errors and lack of information associated with it. Overall, it is necessary to appropriately quantify the uncertainties which undermine the SHM concept. This thesis proposes an enhanced modular Bayesian framework (MBA) for structural identification (st-id) and measurement system design (MSD). The framework is hybrid, in the sense that it uses a physics-based model, and Gaussian processes (mrGp) which are trained against data, for uncertainty quantification. The mrGp act as emulators of the model response surface and its model discrepancy, also quantifying observation error, parametric and interpolation uncertainty. Finally, this framework has been enhanced with the Metropolis–Hastings for multiple parameters st-id. In contrast to other probabilistic frameworks, the MBA allows to estimate structural parameters (which reflect a performance of interest) consistently with their physical interpretation, while highlighting patterns of a model’s discrepancy. The MBA performance can be substantially improved by considering multiple responses which are sensitive to the structural parameters. An extension of the MBA for MSD has been validated on a reduced-scale aluminium bridge subject to thermal expansion (supported at one end with springs and instrumented with strain gauges and thermocouples). A finite element (FE) model of the structure was used to obtain a semi-optimal sensor configuration for stid. Results indicate that 1) measuring responses which are sensitive to the structural parameters and are more directly related to model discrepancy, provide the best results for st-id; 2) prior knowledge of the model discrepancy is essential to capture the latter type of responses. Subsequently, an extension of the MBA for st-id was also applied for identification of the springs stiffness, and results indicate relative errors five times less than other state of the art Bayesian/deterministic methodologies. Finally, a first application to field data was performed, to calibrate a detailed FE model of the Tamar suspension bridge using long-term monitored data. Measurements of temperature, traffic, mid-span displacement and natural frequencies of the bridge, were used to identify the bridge’s main/stay cables initial strain and friction of its bearings. Validation of results suggests that the identified parameters agree more closely with the true structural behaviour of the bridge, with an error that is several orders of magnitude smaller than other probabilistic st-id approaches. Additionally, the MBA allowed to predicted model discrepancy functions to assess the predictive ability of the Tamar bridge FE model. It was found, that the model predicts more accurately the bridge mid-span displacements than its natural frequencies, and that the adopted traffic model is less able to simulate the bridge behaviour during periods of traffic jams. Future developments of the MBA framework include its extension and application for damage detection and MSD with multiple parameter identification.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:759687 |
| Date | January 2018 |
| Creators | Jesus, André H. |
| Publisher | University of Warwick |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://wrap.warwick.ac.uk/109522/ |
Page generated in 0.2249 seconds