Deep Neural Network Approach for the Prediction of Journal Bearing Static Performance Characteristics

Abstract

Deep neural network approach is an excellent way of performance predictions of mechanical systems due to the advancement in computational technologies. In the present document, the predictions of static performance characteristics are made for the hole-entry hybrid journal bearing. Maximum pressure and minimum fluid film thickness values are obtained using FEM and used as target output for feedforward backpropagation neural network model. In this model, hidden layers and number of neurons in these layers are decided heuristically. Logistic activation function is used for hidden and output layer neurons. Using the developed model, predictions for journal bearing performance are made within and out of the prescribed range of input parameters. The percentage error obtained for neural network training, testing and predictions is very small (−1.0% < error < 1.0%). It is concluded that a lot of time is saved in predictions using deep neural network approach compared to the mathematical analysis the journal bearing performance. The use of multiple hidden layers for journal bearing performance predictions and multiple data sets for input and output neurons is the novelty of the present work.