Performance analysis of tilting pad journal bearing using COMSOL Multiphysics and Neural Networks

The objective of this paper is to study the effect of geometrical parameters on the performance of tilting pad journal bearing. COMSOL Multiphysics software is used to simulate the tilting pad journal bearing at different eccentricity ratios and pad clearances. The effect of changing pad numbers (from 3 to 6 pads) and the pads clearance angles (from 2° to 6°) on the performance parameters such as the load carrying capacity, frictional torque and attitude angle was analyzed. It was found that as the number of pads increases the load carrying capacity and the friction coefficient increase, while the attitude angle decreases. Decreasing the clearance between pads leads to an increase in the load carrying capacity of the bearing while it has minor effect on the other performance parameters. The characteristic data obtained from COMSOL program is used to train three suggested neural networks. The first is a feed-forward Neural Network which consists of three layers with {80 sigmoid, 5 sigmoid and 1 linear} neurons. The second is a radial bases Neural Network and the third is a generalized regression Neural Network. Applying the three trained Neural Networks (Feed forward, Radial basis, and Generalized regression) to predict the performance of tilting pad bearing at values not used in the training process. The results show a high accuracy to predict the attitude angle, the coefficient of friction and the load carrying capacity. The percentage relative errors between the predicted values and that obtained by COMSOL are between (0 and 0.98) %. The neural networks model show high capacity in predicting the output variables correctly for both the training data and that which was not included.