Wind farm-level scour detection around offshore monopile foundations using unsupervised domain adaptation

This study proposes a novel Unsupervised Domain Adaptation (UDA) approach for scour detection around offshore wind turbines (OWT) at the wind farm level. In this approach, acceleration data collected from a single reference OWT (so-called source turbine) are used to detect scour presence around other turbines (so-called target turbines), located within the same wind farm. Acceleration data are collected along the tower height under varying environmental conditions and are processed to extract the tower bending modal frequencies, which serve as structural health features. To address the distributional shift in the data across OWTs within a wind farm - arising from changes in structural configurations, environmental loading conditions, and soil properties - four UDA techniques, Geodesic Flow Kernel (GFK), Unsupervised Domain-Invariant Component Analysis (UDICA), Subspace Alignment (SA), and Correlation Alignment (CORAL), are employed. By using these UDA methods, the healthy feature distributions from the source OWT are aligned with those of the target turbines and subsequently used to train an Autoencoder (AE) deep learning algorithm for scour detection. The study evaluates the limitations of traditional anomaly detection methods for scour in terms of their ability to be generalised across turbines, and demonstrates how UDA techniques enable model generalisation at the wind farm level. The results highlight the effectiveness of the GFK in mitigating feature distribution discrepancies and enabling an accurate detection of scour at multiple OWTs within the same wind farm, followed by SA and UDICA. However, CORAL demonstrated a lower capacity for reducing the distribution shift across different OWTs, making it less feasible for a scour detection application.