Abstract
Reynolds based thermo-elasto-hydrodynamic (TEHD) simulations of tilting pad journal bearings (TPJBs) generally provide accurate results; however, the uncertainty of the pad’s leading edge thermal boundary conditions causes uncertainty of the results. The highly complex thermal-flow mixing action between pads (BPs) results from the oil supply nozzle jets and geometric features. The conventional Reynolds approach employs mixing coefficients (MCs), estimated from experience, to approximate a uniform inlet temperature for each pad. Part I utilized complex computational fluid dynamics (CFD) flow modeling to illustrate that temperature distributions at the pad inlets may deviate strongly from being uniform. The present work retains the uniform MC model but obtains the MC from detailed three-dimensional CFD modeling and machine learning, which could be extended to the radially and axially varying MC case. The steps for implementing an artificial neural network (ANN) approach for MC regression are provided as follows: (1) utilize a design of experiment step for obtaining an adaptable training set, (2) conduct CFD simulations on the BP to obtain the outputs of the training set, (3) apply an ANN learning process by Levenverg–Mardquart backpropagation with the Bayesian regularization, and (4) couple the ANN MC results with conventional TEHD Reynolds models. An approximate log fitting method provides a simplified approach for MC regression. The effectiveness of the Reynolds TEHD TPJB model with ANN regression-based MC distributions is confirmed by comparison with CFD based TEHD TPJB model results. The method obtains an accuracy nearly the same as the complete CFD model, but with the computational economy of a Reynolds approach.
Combining Machine Learning with 3D-CFD Modeling for Optimizing a DISI Engine Performance during Cold-Start
