Multi-Objective Optimization and Supervised Machine Learning for Synthesis of Rolling Element Bearing and Lubricant Selection

Abstract

This article introduces a novel approach for optimizing rolling element bearings design and lubricant selection utilizing elastohydrodynamic lubrication (EHL) theory. First, a Multi-Objective Optimization (MOO) is formulated, using a Non-dominated Sorting Genetic Algorithm (NSGA-II), to generate Pareto front optimal designs, capturing the trade-off between the two objective functions, namely maximizing minimum film thickness and minimizing the total friction torque. Second, a machine learning model is trained using the results obtained from the NSGA-II model. The machine learning algorithm used in this article is Random Forest Regression (RFR), a supervised ensemble learning method combining multiple decision trees to improve predictive accuracy. A case study from a rolling element bearings manufacturer is used in this article to validate the proposed method. Initially, several bearing configurations (inner diameter, outer diameter, bearing width, radial force, and rotation speed) of a certain series were used as input to the NSGA-II model. After that, the Pareto front for all bearing configurations was used as a training and test set for the machine learning model. Finally, a new rolling element bearing configuration (different from the bearing configurations used in the first step) is used to predict the internal dimensions and compare the results with previous literature and manufacturer catalogs. The results produced with RFR are also compared with another supervised machine learning algorithm called Support Vector Regression (SVR) to test the superiority of RFR. The effectiveness of the proposed approach is evident in computational time, as the machine learning model can predict the optimum design of rolling element bearings compared to applying traditional metaheuristic techniques.