Integrating DL-based surrogate within an Interacting Particle Ensemble Kalman Filtering framework for computationally efficient condition monitoring of FOWT moorings

Abstract

Traditional model-driven strategies for monitoring the condition of moorings in Floating Offshore Wind Turbines (FOWTs) depend on resource-intensive simulations of intricate high-fidelity models. Conversely, data-driven methods offer unbiased and prompt results but necessitate comprehensive labeled datasets that cover various damage scenarios, which are limited for FOWTs. Furthermore, the significant uncertainties associated with FOWTs’ operational environments complicate matters, making probabilistic estimation essential. To tackle this, filtering-based probabilistic methods are pertinent, although they typically require repetitive simulations of high-fidelity models.

An innovative solution involves replacing the high-fidelity model with a Deep Learning (DL)-based surrogate acting as a “data-based predictor” (DBP) to simulate FOWT dynamics. Utilizing a Fully Connected Neural Network (FCNN) architecture, the DBP is trained on synthetic time series data from a calibrated OpenFAST model, enabling it to learn FOWT dynamics and make one-step-ahead predictions based on current responses and health states, thus facilitating real-time monitoring. This surrogate model is integrated into a stochastic inverse estimation framework using the Interacting Particle Ensemble Kalman Filter (IPEnKF) to assess mooring health. Comprehensive tests on the NREL 5-MW wind turbine model mounted on an OC4 semi-submersible platform under varying noise levels, damage, and sea conditions demonstrate the method’s accuracy, precision, detection promptness, and reliability.