Advanced Rotordynamics Analysis through CFD-Informed Machine Learning Mixing Prediction

Abstract

The Mixing Coefficient (MC) is an uncertain, assumed parameter for predicting the leading-edge temperature of a pad film domain. The conventional approach, employing an assumed MC has shown limitations for accurately predicting rotor-bearing system performance. This leads to a temperature discontinuity problem between pads due to the complete mixing assumption. Also, investigation has shown that the classical Finite-Element-Method (FEM) causes an irregular temperature discontinuity problem near the shaft surface. These problems diminish the accuracy of predictions for rotordynamic response. Therefore, this research proposes (1) full CFD model and (2) a FVM-based 3D Hybrid Between Pad (HBP) model, utilizing CFD databased machine learning ML, to accurately model radial and axial temperature distributions at the leading-edge of a pad film domain. This research provides the specific modeling methodology and simulation results for static characteristics and dynamic coefficient predictions.

The modeling approach is also applied to the synchronous instability Morton Effect (ME) analysis. The Morton Effect (ME) is a thermally induced, rotordynamic instability phenomenon. It occurs when the journal in a rotor-bearing system experiences asymmetric heating due to synchronous vibration, resulting in thermal bowing of the rotating assembly. The bow may increase vibration and the asymmetric heating of the journal, which in turn further increases the bow. This positive feedback loop ultimately terminates with the vibration exceeding allowable limits which trips the unit to zero speed. The multiphysics nature of the ME, involving conjugate heat transfer, thermal deformation, fluid flow and vibrations, with broadly separated time constants and stationary and rotating reference frames, presents a simulation modeling challenge. The assumptions and methods employed in prior models yield highly approximate, asymmetric journal temperature predictions, which is a key component of ME prediction. These approaches lead to temperature discontinuity problems between pads and near the journal surface. The novel approach presented here corrects these weaknesses. The developed model shows notable improvements in terms of computation speed, removal of energy conservation violations and agreement with available test data, including the data conducted by the author.