Digital-Twin Approach Predicts Fatigue Damage of Marine Risers

Abstract

This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper OTC 30985, “From Data to Assessment Models, Demonstrated Through a Digital Twin of Marine Risers,” by Ehsan Kharazmi and Zhicheng Wang, Brown University, and Dixia Fan, SPE, Massachusetts Institute of Technology, et al., prepared for the 2021 Offshore Technology Conference, Houston, 16–19 August. The paper has not been peer reviewed. Copyright 2021 Offshore Technology Conference. Reproduced by permission. Assessing fatigue damage in marine risers caused by vortex-induced vibrations (VIV) serves as a comprehensive example of using machine-learning methods to derive assessment models of complex systems. A complete characterization of the response of such complex systems usually is unavailable despite massive experimental data and computation results. These algorithms can use multifidelity data sets from multiple sources. In the complete paper, the authors develop a three-pronged approach to demonstrate how tools in machine learning are used to develop data-driven models that can be used for accurate and efficient fatigue-damage predictions for marine risers subject to VIV. Introduction In this study, machine-learning tools are developed to construct a digital twin of a marine riser. The digital twin uses various sources of training data, including field data, experimental data, computational-fluid-dynamics simulations, extracted databases, semiempirical codes, and existing knowledge of underlying physical models. The authors also show that a well-trained digital twin can use the streaming data from a few field sensors efficiently to provide an accurate reconstruction of motion and to provide fatigue-damage prediction. Several machine-learning algorithms have been developed in the literature to predict the life span of the structure through the changes in parameters. To the best of the authors’ knowledge, most existing methods are developed as black boxes that return parameters by only feeding experimental data and therefore are ignorant of the underlying physics. In the first of three approaches, the authors enhance the capabilities of semiempirical codes by developing efficient databases through active learning. In the second approach, the LSTM-ModNet framework is applied to reconstruct and analyze the entire motion of a riser in deep water from sensor measurements through modal decomposition in space and the sequence-learning capability of recurrent neural networks in time. The formulation described in the paper provides a tool that efficiently combines different types of sensor measurements, such as strain and acceleration. In the third approach, a higher level of abstraction is introduced and the nonlinear operator that maps the inflow current velocity to the root-mean-square function of the riser response is approximated. In particular, the newly developed neural network DeepONet is used as a black box to learn the mapping between the input parameters (the inflow velocity, riser bending stiffness, and tension as a function of water depth) to the output parameters (strain, amplitude, and exciting frequencies as a function of water depth). In these approaches, data from the high-mode VIV test is used to train the networks.