This paper proposes a novel data-driven framework for scour detection around offshore wind turbines (OWTs), where damage features are derived from wind and wave-induced acceleration signals collected along the tower. A numerical model of the NREL 5 MW wind turbine, which considers aerodynamic and hydrodynamic loading with soil-structure interaction (SSI) and servodynamics, is developed. The model is used to simulate the acceleration responses along the tower for a healthy structure, and a structure affected by progressive scour. A data segmentation process is initially performed on the collected data, which is followed by a feature selection scheme based on the analysis-of-variance (ANOVA) algorithm, to eliminate irrelevant characteristics from the time domain feature set of responses. The proposed framework consists of two main components: (a) offline training, and (b) real-time classification. The acceleration responses collected from the healthy structure and the structure subjected to three different damage scenarios (different scour depths) and under various load conditions, are used in the offline training mode. The selected feature vector from the feature extraction process is used as input to a Naive Bayes classifier (NBC) algorithm to train the model. In the real-time classification, a prediction of the scour depth affecting the structure is performed using a new dataset simulated from unseen load cases and scour conditions of the OWT. The results show that the model trained in the offline stage can predict the scour depth in the real-time monitoring stage with performance measures over approximately 94%.

Development of digital twins is emerging rapidly in geotechnical engineering, and it often requires real-time updating of numerical models (e.g., finite element model) using multiple sources of monitoring data (e.g., settlement and pore water pressure data). Conventional model updating, or calibration, often involves repeated executions of the numerical model, using monitoring data from a specific source or at limited spatial locations only. This leads to a critical research need of real-time model updating and predictions using a numerical model improved continuously by multi-source monitoring data. To address this need, a physics-informed machine learning method called multi-source sparse dictionary learning (MS-SDL) is proposed in this study. Originated from signal decomposition and compression, MS-SDL utilizes results from a suite of numerical models as basis functions, or dictionary atoms, and employs multi-source monitoring data to select a limited number of important atoms for predicting multiple, spatiotemporally varying geotechnical responses. As monitoring data are collected sequentially, no repeated evaluations of computational numerical models are needed, and an automatic and real-time model calibration is achieved for continuously improving model predictions. A real project in Hong Kong is presented to illustrate the proposed approach. Effect of monitoring data from different sources is also investigated.