Application of Neural Network to Width Estimation in Flat Rolling

Back-propagation (BP) neural networks have been applied to estimate exit width for flat rolling. The data from real rolling process are used to train and test the networks. It is found that the BP neural networks are reasonably accurate in estimating the actual width. Furthermore, it is found that the estimation accuracy can be improved when some input processing elements of the BP neural networks are modified according to rolling theory. In order to obtain more accurate estimation results, integrated functional-linked (IFL) neural networks are introduced, in which a functional-linked network is integrated to BP networks. Different linked functions produce different results, the application indicates that integration of some functions to BP networks can successfully capture the inherent feature of width deformation. It is found that the accuracy of width estimation by the IFL networks is satisfactory.