Verification and validation of physics-based models for structural health monitoring

Structural health monitoring (SHM) refers to a group of methods through which engineers aim to infer the health state of a piece of engineering infrastructure given some measured data from that structure. These inferences are then used to inform ongoing maintenance strategies to ensure that the structure remains operating in its optimal condition, without compromising performance or safety. Given the age of much of the existing infrastructure around the world, life-extending technology such as SHM is clearly of significant potential benefit, and the use of SHM when developing new infrastructure offers the potential for maximising the return on the investment in that infrastructure.

SHM strategies can generally be divided into two types. Data-driven SHM uses data taken from the structure of interest to train a statistical model, such that the trained model will be able to label new data from the structure as being indicative of a particular health state. Model-driven SHM traditionally uses a physics-based model that can be adjusted to fit its predictions to live data taken from the structure, such that the adjustments to the model inputs give some indication of the structure's health state. Data-driven SHM methods are reliant on large sets of training data on which to develop statistical models; this is often difficult to acquire in practice, particularly when data are required from a structure in its damaged states. This issue can be mitigated by using a physics-based model to simulate the training data for the statistical model; however, in order for the physics-based model predictions to be considered trustworthy, the model must be validated against experimental data. This means that damage-state data are still required for the implementation of physics-based models in SHM in order to ensure the accuracy of their predictions. The acquisition of damage-state data from structures for this purpose therefore presents a significant set of problems for SHM methods. It is possible to reduce the difficulties associated with model validation for SHM by carrying out validation on models of the subassemblies and components that make up a larger assembly structure. The data required for this hierarchical validation task should be easier to acquire cheaply compared to validation data drawn from the full structure. If the submodels representing these individual substructures can each be validated then it should be possible to recover an assembly-level model which, given the validation tasks carried out, could be used to make confident predictions regarding the structure in a range of health states.

A framework is presented in this thesis which summarises the activities required to carry out this hierarchical validation strategy, which would then enable a model to be developed with demonstrable accuracy and quantifiable uncertainty in its predictions. This framework is applied to a target structure { a truss bridge { and the model is used to carry out a series of SHM tasks on test data drawn from the structure. These tasks are carried out using the validated model in a forward manner to generate training data for statistical damage recognition models, which are then compared { in terms of performance { to traditional data-driven methods.

Based on the research presented in this thesis, it is shown that model uncertainty can be accurately quantified through the hierarchical validation process by comparing the model predictions to the experimental test data before and after the validation process. After this it is shown that it is possible to develop accurate damage classification algorithms using validated model predictions, with the SHM methods developed via the hierarchical validation framework performing favourably compared to traditional data-driven methods when exposed to the test data. Further research areas that would advance the methodologies presented in the thesis are then outlined following discussions of the results.