One of the most significant challenges for signal processing in data-based structural health monitoring (SHM) is a lack of comprehensive data; in particular, recording labels to describe what each of the measured signals represent. For example, consider an offshore wind-turbine, monitored by an SHM strategy. It is infeasible to artificially damage such a high-value asset to collect signals that might relate to the damaged structure in situ; additionally, signals that correspond to abnormal wave-loading, or unusually low-temperatures, could take several years to be recorded. Regular inspections of the turbine in operation, to describe (and label) what measured data represent, would also prove impracticable -- conventionally, it is only possible to check various components (such as the turbine blades) following manual inspection; this involves travelling to a remote, offshore location, which is a high-cost procedure.
Therefore, the collection of labelled data is generally limited by some expense incurred when investigating the signals; this might include direct costs, or loss of income due to down-time. Conventionally, incomplete label information forces a dependence on unsupervised machine learning, limiting SHM strategies to damage (i.e. novelty) detection. However, while comprehensive and fully labelled data can be rare, it is often possible to provide labels for a limited subset of data, given a label budget. In this scenario, partially-supervised machine learning should become relevant. The associated algorithms offer an alternative approach to monitor measured data, as they can utilise both labelled and unlabelled signals, within a unifying training scheme.
In consequence, this work introduces (and adapts) partially-supervised algorithms for SHM; specifically, semi-supervised and active learning methods. Through applications to experimental data, semi-supervised learning is shown to utilise information in the unlabelled signals, alongside a limited set of labelled data, to further update a predictive-model. On the other hand, active learning improves the predictive performance by querying specific signals to investigate, which are assumed the most informative. Both discriminative and generative methods are investigated, leading towards a novel, probabilistic framework, to classify, investigate, and label signals for online SHM. The findings indicate that, through partially-supervised learning, the cost associated with labelling data can be managed, as the information in a selected subset of labelled signals can be combined with larger sets of unlabelled data -- increasing the potential scope and predictive performance for data-driven SHM.
