머신러닝을 활용한 고속용 베어링 케이지의 마찰손실 예측

Abstract

This study details the development of an artificial neural network (ANN) model trained with data gen erated from computational fluid dynamics (CFD) simulations. To construct a reliable training dataset, the study strategically utilizes periodic CFD results. The CFD model, built in ANSYS Fluent, utilizes a single-phase steady-state configuration with three ball-boundary conditions, referred to as "periodic CFD." The training data are generated to systematically assess the effects of key factors influencing cage frictional losses, including oil viscosity, oil volume fraction, operating speed, and cage design parameters. A parametric study is performed to identify the optimal ANN architecture, including the number of hidden layers and nodes, to maximize prediction performance. In addition, a supervised constraint enforcement method is applied to mitigate outlier predictions and improve the model's stability and accuracy. The model is validated through a case study involving a reduced cage height, where the ANN predictions exhibit trends consistent with the original CFD simulations.