Abstract:

During the setting of rolling factors with complicated correlation. It force in continuous hot strip rolling, force signals are influenced by is, therefore, difficult to establish an accurate model to describe the various rolling mechanism. In order to solve this problem, a multi-RBF neural network model is proposed. In this new model, the multi-resolution wavelet analysis method is employed to separate the rolling force signal into several sub-signals corresponding to different factors, and several RBF neural networks are established, each for a certain sub-signal. All the outputs of the sub-networks are integrated into a rolling force signal, and both the input and the output of each network relate to the affecting factors of the corresponding sub-signal. Thus, the sub-networks can well reflect the variation mechanism of the rolling force. Simulated results show that the proposed model decreases the number of system dimensions, improves the learning ability of the network, and reduces the error rate of rolling force setting from the original 10% in BP neural network model to 5%.

Key words: continuous hot rolling, rolling force, wavelet analysis, multi-RBF neural network