Abstract: Based on the outstanding characteristic of nonlinear convergence of neural network, an improved BP neural network was established. Orthogonal experiments were carried out to provide batch training samples for the network. And the LevenbergMarquardt algorithm was used to minimize the errors of the network. In addition, the Bayesian regularization was employed to optimize the combination of squared errors, weights and the sum of squared threshold. The experimental results show that the improved BP neural network has fast rate of convergence, strong generalization capability and good stability, which can simulate and predict the maximum temperature in sphere grinding with high accuracy.
 

Key words: neural network, sphere grinding, rthogonal experiment, grinding temperature, Levenberg-Marquardt algorithm