Machine Learning Framework for Scour Detection Across Multiple Offshore Wind Farms

Abstract. This paper proposes a global scour detection framework for monopile foundations across offshore wind farms, based on data from a single accelerometer installed on the turbine tower. The framework is designed to address the real-world complexity and heterogeneity of offshore wind systems installed across multiple offshore sites. To achieve this, numerically generated acceleration data are obtained for various offshore wind turbines (OWTs) across multiple wind farms using a coupled OpenFAST and bespoke soil-structure interaction (SSI) model. The simulation accounts for a wide array of offshore conditions, from sea states and soil properties to structural characteristics and site-specific scour. For each OWT, acceleration data are generated using foundation stiffness derived from the SSI model, reflecting the site conditions and turbine characteristics. A multi-source domain generalisation (DG) strategy is then employed, in which a model trained on a combined dataset containing one turbine per farm (referred to as source turbines) is used to detect scour around the remaining, previously unseen turbines (referred to as target turbines) across geographically-disparate wind farms. The results demonstrate that the proposed method can identify the scour state for multiple target turbines across multiple wind farms with acceptable accuracy. In addition, the choice of source turbine significantly impacts model performance, with shallow water, low-stiffness turbine foundations providing the most reliable training base. Furthermore, obtaining sensor data from the tower base significantly improves scour detection, while increasing the number of source turbines in the training dataset enhances prediction accuracy, specifically at lower scour states.