The ever-growing need for renewable energy has driven the development of increasingly large offshore wind turbines. Alongside improved design codes and changing control strategies, this has led to fatigue becoming an operational concern. Farm operators require information about the impact of their decisions (e.g. curtailment) on the structural reserve of each turbine, which necessitates a tool that can predict both quickly and for yet unseen situations. Current approaches rely either on numerical simulations, which are too slow, or data-driven methodologies, which often suffer from data sparsity and limited generalization capabilities. Therefore, our goal is to develop a surrogate model that enables real-time control through rapid inference while improving extrapolation capabilities. To this end, we propose the use of a Graph Neural Network (GNN).
GNNs are ideally suited to interpret and learn from interdependent non-Euclidean systems such as wind farms. They are able to process many different farm lay-outs, meaning they can learn to capture the underlying relations between the variables and the positions of the turbines. When the model is pretrained on data from one kind of farm and later fine-tuned on data from another, valuable knowledge can be transferred between both cases in a process called Transfer Learning. This has been successfully implemented on wind farms in previous work, but for simulated data only. In this contribution, we will extend this approach to real-world data.
First, the network is trained on a large collection of generated farm layouts with global wind inflow conditions (wind speed, wind direction and turbulence intensity) drawn from their respective distributions while simulated predictions are obtained using PyWake, an open-source wind farm simulation tool capable of computing wake effects and power production of individual turbines. The goal is to predict local (i.e per-turbine) wind conditions and power as well as fore-aft (FA) and side-side (SS) damage equivalent moment (DEM). Afterwards, the model will be fine-tuned on real-world data from instrumented turbines in a North Sea wind farm. Farm-wide DEMs (our ground truth) are obtained as machine learning model predictions using accelerometer data. Multiple transfer learning techniques will be compared, including layer freezing and low rank adaptation (LoRA), while it will be investigated how to minimize negative transfer. Predictions from models obtained with TL will be contrasted with those obtained from models trained either from scratch or without TL to expose its benefits. Additionally, predictions on unseen data will further determine whether effective generalization has taken place, indicating the underlying fundamental physics has been captured and been successfully transferred.

Passive impact localization using piezoelectric sensor (PZT) networks and machine learning is an established approach for structural health monitoring of composite aerospace panels. A key practical limitation is that models trained on one sensor layout fail when deployed on a structurally similar panel with a different instrumentation arrangement, and full retraining is costly. This paper investigates transfer learning (TL) as a strategy to adapt a multilayer perceptron (MLP) trained on a stiffened AS4/8552 CFRP panel to ten alternative sensor layouts, simulated via controlled sensor-index permutations grouped into three families of increasing severity: intra-column swaps, intra-row swaps, and full-plate reflections. Six TL strategies, from output-head-only retraining to full fine-tuning, are benchmarked at five adaptation data fractions against a from-scratch baseline. Results show that full fine-tuning is the only strategy that consistently recovers baseline localization accuracy (25.7~mm mean Euclidean error) across all permutation families, requiring as little as 5\% of the adaptation dataset for mild and moderate shifts and 25\% for severe full-plate reflections. Partial fine-tuning strategies fail to recover acceptable accuracy at any data fraction, due to the globally distributed nature of sensor-layout encoding in hand-feature MLPs. These findings provide practical guidelines for redeploying impact localization models across panels with varying sensor configurations.

In structural health monitoring (SHM), supervised damage localization relies on knowledge of the damage, which can either be obtained from direct measurements or inferred through a structural model due to the scarcity of complete damage datasets. Model-based approaches employ these structural models for damage localization and can be enhanced by incorporating a relative least-squares metric, improving robustness against model inaccuracies. Recently, hybrid approaches have become increasingly popular which combine data-driven techniques with structural information via transfer learning. Here, the structural models are used to generate source domain-labeled data. However, like purely model-based approaches, hybrid approaches are challenged by incomplete or imprecise structural model information and the resulting model inaccuracies. Using a cantilever beam with varying mass scenarios, this study investigates how the precision of structural models affects the damage localization result of both the purely model-based and the hybrid approaches. The presented results provide guidance for SHM decision makers who have to deal with the limitations of precise structural modeling when only limited data and information are available.