On risk-based decision-making for structural health monitoring

Structural health monitoring (SHM) technologies seek to detect, localise, and characterise damage present within structures and infrastructure. Arguably, the foremost incentive for developing and implementing SHM systems is to improve the quality of operation and maintenance (O&M) strategies for structures, such that safety can be enhanced, or greater economic benefits can be realised. Given this motivation, SHM systems can be considered primarily as decision-support tools. Although much research has been conducted into damage identification and characterisation approaches, there has been relatively little that has explicitly considered the decision-making applications of SHM systems. In light of this fact, the current thesis seeks to consider decision-making for SHM with respect to risk. Risk, defined as a product of probability and cost, can be interpreted as an expected utility. The keystone of the current thesis is a general framework for conducting risk-based, SHM generated by combining aspects of probabilistic risk assessment (PRA) with the existing statistical pattern recognition paradigm for SHM. The framework, founded on probabilistic graphical models (PGMs), utilises Bayesian network representations of fault-trees to facilitate the flow of information between observations of discriminative features to failure states of structures of interest. Using estimations of failure probabilities in conjunction with utility functions that capture the severity of consequences enables risk assessments -- these risks can be minimised with respect to candidate maintenance actions to determine optimal strategies. Key elements of the decision framework are examined; in particular, a physics-based methodology for initialising a structural degradation model defining health-state transition probabilities is presented. The risk-based framework allows aspects of SHM systems to be developed with explicit consideration for the decision-support applications. In relation to this aim, the current thesis proposes a novel approach to learn statistical classification models within an online SHM system. The approach adopts an active learning framework in which descriptive labels, corresponding to salient health states of a structure, are obtained via structural inspections. To account for the decision processes associated with SHM, structural inspections are mandated according to the expected value of information for data-labels. The resulting risk-based active learning algorithm is shown to yield cost-effective improvements in the performance of decision-making agents, in addition to reducing the number of manual inspections made over the course of a monitoring campaign. Characteristics of the risk-based active learning algorithm are further investigated, with particular focus on the effects of \sampling bias. Sampling bias is known to degrade decision-making performance over time, thus engineers have a vested interest in mitigating its negative effects. On this theme, two approaches are considered for improving risk-based active learning; semi-supervised learning, and discriminative classification models. Semi-supervised learning yielded mixed results, with performance being highly dependent on base distributions being representative of the underlying data. On the other hand, discriminative classifiers performed strongly across the board. It is shown that by mitigating the negative effects of sampling bias via classifier and algorithm design, decision-support systems can be enhanced, resulting in more cost-effective O&M strategies. Finally, the future of risk-based decision-making is considered. Particular attention is given to population-based structural health monitoring (PBSHM), and the management of fleets of assets. The hierarchical representation of structures used to develop the risk-based SHM framework is extended to populations of structures. Initial research into PBSHM shows promising results with respect to the transfer of information between individual structures comprising a population. The significance of these results in the context of decision-making is discussed. To summarise, by framing SHM systems as decision-support tools, risk-informed O&M strategies can be developed for structures and infrastructure such that safety is improved and costs are reduced.