Deep graph learning approach-based damage detection using vibration responses of numerical model: Verification on an experimental offshore monopile model

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Monopiles are the most commonly used foundation type in offshore wind turbine structures and are constantly subjected to environmental forces. Wind, waves, and earthquakes are among the environmental factors that can cause serious and irreversible damage to these structures. Among the various types of defects, scouring is the most prominent, as waves consistently intensify this defect in the structures. Therefore, structural health monitoring of these systems is particularly important. In this study, a digital twin (DT) is proposed as a solution for condition monitoring. The core of the DT is finite element modeling, which is used to simulate and obtain dynamic responses. In this process, modeling a damage-sensitive model is prioritized. To achieve this, the model parameters are updated using the proposed optimization method and the results of experimental modal analysis (EMA). For EMA and obtaining dynamic responses under scour defect conditions, a monopile with specific dimensions is prepared and then excited using a defined signal to acquire the dynamic responses. Based on the proposed graph neural network (GNN; GAT_GATv2), damage-sensitive features are learned from the simulation data under different scouring depths to train the model, and the data obtained from the experimental models are used for testing to classify different damage scenarios. The GAT_GATv2 model utilizes the local stability of the GATConv layer and the enhanced dynamic response of the GATv2Conv layer to achieve graph-level classification. This method is superior to the traditional GNN in classification accuracy.