Neural Network Modeling of Surface Roughness and Residual Stress Induced by Ball Burnishing

Abstract

Under the influence of high temperature and high pressure, aero-engine parts are prone to fatigue fracture. In order to improve their fatigue performance, the surface of parts should be further strengthened. Ball burnishing can not only improve the surface roughness of parts but also produce compressive residual stress and hardening layer on the surface layer. In addition, ball burnishing can also greatly reduce production costs. Due to its advantages, ball burnishing has been more and more widely used in the aerospace field to improve the surface properties of parts. In this article, to clarify the relationship between the surface roughness and residual stress and the ball burnishing parameters, the three-level four-factor ball burnishing experiments were performed using response surface methodology with Box-Behnken design (BBD) considering burnishing pressure, burnishing speed, feed rate and number of passes. An artificial neural network predictive model of surface roughness and residual stress was established based on the selected 19 experimental data. The samples are divided into the training samples, validation samples and test samples. The number of nodes of the input layer, the output layer and the hidden layer are 4, 2 and 7, respectively. The results show that the maximum errors between the predicted values and experimental values are 5.8% and 9.1%. The correlation coefficients (RC) are 0.99316, 0.96187, 0.97534 and 0.94563 for the training samples, validation samples, test samples and all samples respectively, which shows that the ANN model can effectively predict the surface roughness and residual stress. The surface roughness and surface residual stress were optimized by the non-dominated sorting genetic algorithm-II (NSGA-II) method. The optimal ball burnishing parameters combination are determined, which is 18MPa, 0.1mm/r, 600r/min and 1 for burnishing pressure, burnishing feed, burnishing speed and number of passes respectively. Finally, the determined optimal result was verified by the experiment. The errors between experimental values and predicted values of surface roughness and residual stress are 7.5% and 2.1%, respectively.