Characterizing Damage in Wind Turbine Mooring Using a Data-Driven Predictor Model within a Particle Filtering Estimation Framework

Abstract

Floating Offshore Wind Turbines (FOWT) represent a promising solution to renewable energy challenges, yet effective maintenance remains critical for cost management. Traditional machine learning (ML) approaches for detecting FOWT damage often rely on extensive real-world data, which can be impractical and economically unfeasible. Alternatively, stochastic filtering-based time-domain approaches leverage physical understanding through dynamic models, typically finite element models. However, these methods are hindered by excessive simulation calls within the recursive filtering frameworks. This study proposes a novel filtering-based approach that replaces the computationally intensive process model with a Deep Neural Network (DNN) surrogate, addressing the aforementioned limitations. The proposed approach utilizes synthetic data generated from the high-fidelity calibrated OpenFAST model of FOWT dynamics to train a DNN toward learning the dynamic evolution of the FOWT conditioned on the current health state. By offering a computationally efficient representation of system dynamics conditioned on health state, this approach allows for real-time damage detection and interpretable information on damage severity within a stochastic inverse estimation framework, specifically employing Particle Filtering in this study. This approach contrasts with traditional black-box ML-based methods, which typically struggle to provide interpretable information on damage characteristics. Extensive numerical investigations on damaged FOWT mooring lines demonstrate this approach's practical applicability and superiority over traditional ML-based methods. Eventually, integrating explainable ML models within the filtering framework induces promptness in detection without sacrificing transparency.