We show how to reconstruct the vortex-induced vibrations of a riser from experimental strain measurements using a machine learning framework. We employ a modal decomposition technique followed by inference of the expansion modes using a two-stage optimization routine. A stochastic mode search algorithm is developed and its capabilities and limitations are demonstrated using the MIAMI II riser field experiments, conducted in the Gulf Stream off the coast of Miami, FL using a densely instrumented riser model. Validation is done according to a k-fold cross-validation scheme, and a VIV physics-based examination is presented. The reconstruction framework's complexity in terms of data required for successful training is finally evaluated.
Vortex induced vibrations (VIV) are driven by the periodic shedding of vortices formed in the wake behind bluff bodies placed within currents (Triantafyllou et al., 2016). The vibration amplitude does not typically exceed one to two body diameters (Bernitsas et al., 2019). Rigid cylinder VIV have become the canonical problem for study of the phenomenon (Williamson and Roshko, 1988; Wu et al., 2014; Zdravkovich, 1996). Flexible body VIV are similar to rigid body vibrations as they are driven by vortex sheeding, but with the added complexity that the loading is non-uniform along the span as the flexible body undergoes spatially traveling and/or standing waves.
Riser motion reconstruction has been done by leveraging the physics-based modal expansion technique (Mukundan, 2008; Triantafyllou et al., 1999; Wang et al., 2019) developed to model vibrations of continuous flexible bodies, such as beams (Rao, 1995). In this work, the modal expansion approach is employed, followed by a data-informed selection of the expansion modes to restrict the model's complexity while satisfying the motion constraints imposed by VIV physics. The challenge is that a large number of parameters are involved in riser modeling, a problem which is common within the field of regression and has led to a variety of techniques for variable subset selection (Guyon and Elisseeff, 2003). This framework is used to satisfy physics-based VIV motion constraints (for example, amplitude restriction), while still utilizing the modal decomposition model.
