On Population-Based Structural Health Monitoring for Offshore Wind Farms

A lack of data, spanning the full range of possible damage states and operational conditions for a given structure, is a significant challenge for data-driven structural health monitoring (SHM). This problem can hinder the insight that data-driven SHM provides to decision-makers, limiting its value. Population-based structural health monitoring (PBSHM) aims to address this challenge by sharing information between multiple structures (a population), such that more complete datasets are obtained, with improved model inference.
Offshore wind (OW) farms, which can be considered as populations of identical wind-turbines, are an intuitive application for PBSHM; the installation of these farms is accelerating rapidly worldwide, however the cost of operating and maintaining them remains high. By increasing the insight available to decision-makers, PBSHM methods have the potential to reduce costs and risks, strengthening the economic case for the OW sector, and ultimately contribute to mitigating climate change.
This thesis builds upon, and applies two key methods identified as being suitable for PBSHM, namely hierarchical Bayesian models (HBMs), and domain adaptation (DA), which both belong to the broader machine-learning (ML) category of transfer learning (TL). In the first example, it is shown how HBMs can more robustly estimate the foundation stiffness of wind-turbine-like structures, using observations from a population, compared to models trained on individual structures. It is also demonstrated how this approach could support the detection of scour, a critical form of damage to OW turbine foundations that is otherwise costly to monitor. In the second example, HBMs are used in a novel way which leverages spatial correlations in turbine-specific model parameters, supporting accurate and probabilistic predictions of turbine output power, even for turbines with no data. Finally, DA is used to transfer information between two separate wind-farms, allowing the prediction of turbine alarms on an unlabelled target wind-farm, with promising results.