G-Twin: Graph neural network-based digital twin for real-time and high-fidelity structural health monitoring for offshore wind turbines

Abstract

The development of digital twin (DT) of real-time and high-fidelity structural health monitoring (SHM) is critical for ensuring the structural safety of an offshore wind turbine (OWT) during its service life. However, reconstruction of high-fidelity stress field in SHM faces great challenges because the monitoring stress data from sensors is normally sparse and limited. In this study, a novel graph neural network (GNN)-based DT, named herein G-Twin, is proposed to reconstruct the high-fidelity stress field in real time using sparse monitoring data. In G-Twin, structures of an OWT are represented as graphs, with nodes and edges capturing the structural geometry in a non-Euclidean space. Graph features are designed as the sparse monitoring data and these features are iteratively aggregated and updated through a message-passing mechanism in terms of the local topology of the graph and the high-fidelity stress field is then achieved. Moreover, an enhanced Mixup technique is developed for data augmentation to minimize the prediction errors when the OWT is subjected to the extreme loading. A series of numerical experiments are conducted and the results show that the G-Twin can accurately predict the high-fidelity stress distribution of an OWT in terms of sparse sensor data in real time (the inference time for the G-Twin on a consumer-grade GPU is approximately 0.013 s on average). The proposed G-Twin has demonstrated its great capability and feasibility for DT of real-time and high-fidelity SHM for OWTs.