Rolling force prediction based on wavelet transform and RBF neural network

Rolling force prediction is very important in hot strip rolling process, and neural network is an effective tool for it. As the rolling force can be decomposed into several components, a rolling force predictor consisting of three radial basis function neural networks is built. Each of the networks predicts one component. An improved wavelet transform algorithm is first applied to decompose the historical rolling force signal, and then the sub-components are reconstructed as the training data of the networks. To eliminate the frequency aliases inherent in the Mallat algorithm, the Fast Fourier Transform and Inverse Fast Fourier Transform are combined with the Mallat algorithm. This anti-aliasing algorithm guarantees that the reconstructed sub-components reflect the real situations. The synthesis of the wavelet algorithm and the implementation of the predictor are described in detail. Experimental examination shows that the proposed predictor achieves better performance than ordinary single network predictor, decreasing the prediction error rate from 10% to less than 5%.