Monitoring a population of structures from a spatiotemporal perspective: an application to offshore wind farms

Population-based structural health monitoring (PBSHM) is an emerging field that is motivated by the lack of labelled data necessary for developing diagnostic tools that assess the type and severity of damage in structures. A key theme in PBSHM is to share data across multiple structures within a population whenever possible. Recent advances in the field focus on methods of knowledge transfer through abstract spaces. This thesis concerns PBSHM for a population of structures whose behaviours are spatiotemporally correlated, for example, an offshore wind farm. A main aim of the work is to utilise the spatiotemporal correlation across a population for the development of physics-inspired, population-level methods for structural condition assessment. Within the scope of the thesis, the focus is to explore how the knowledge of spatiotemporal correlations may contribute to the task of wind farm monitoring.

A key contribution of the thesis is the development of a population-level anomaly detector that accounts for the spatiotemporal correlation between structures, known as the mapping method. The mapping method consists of a spatiotemporal model, which predicts the normal undamaged condition of a population, and a detection criterion, based on which error maps are created to visualise the detection of anomalies.

With regard to the modelling techniques, two approaches are developed to capture the spatiotemporality across a wind farm, both of which are based on Gaussian process (GP) regression. The first approach applies GP regression directly to model the turbine performance across the entire wind farm, capturing the spatiotemporal variation through its covariance function. The second approach models the wind farm performance as a spatial autoregressive process, which leads to the development of the Gaussian process spatial autoregressive model with exogenous inputs (GP-SPARX). The model structure of GP-SPARX is motivated and optimised by considering the physics of wind farms; in this sense, the GP-SPARX is a physics-inspired model.

The developed methods are demonstrated with simulated examples and the data from an operating offshore wind farm. The results show that population-level anomaly detection is viable using a spatiotemporal approach, which is worth investigating into as part of the ongoing research in PBSHM.