The Research Based On CAMC Neural Networks Of Cold Rolling Mill Flatness Prediction Model

Plate and strip production is an important foundation material in nation economy. Theflatness is an important quality indicator of plate and strip production. With theincreasingly highly demand of modern industry on the accuracy of flatness, the quality ofplate and strip production becomes a problem that people need to improve. In the flatnesscontrol system, either the control characteristics of the adjusting structure or the onlinereal-time controlling requires the precise flatness prediction. How to establish a preciseflatness prediction, has become a hot topic between national researchers. A comprehensiveanalysis on flatness prediction has been made on the current situation both in home andabroad, the problems of the previous methods has been shown in this paper and a deepresearch made on the model of flatness prediction was proposed in the paper.Firstly, an improvement on quantizing structure of CMAC neural network ispresented. In order to solve the problem of its dramatic increase of weights storage spacewhen input is high-dimensional, this paper presents an improved CMAC neural network.It uses canopy-k-means algorithm to cluster data samples. The clustering results make thedata with similar output assign to a same cluster. The width of each quantizing interval issame to the interval width of each cluster. According the final number of clusters, thestructure allocates the corresponding storage space. This will save a lot of storage spacewhen the input is high-dimensional data, and realize the nonlinear quantizing structure ofCMAC neural network.Secondly, the flatness prediction based on improved CMAC neural network isestablished. By analyzing the influences of variety of means on flatness controllingprocess, the predicting model inputs are coming from three main areas: rolling parameters,the adjustment amount of the flatness adjusting means and the basic parameters of rollingand roll. The model outputs must reflect the basic model of the flatness. We select theLegendre polynomials to represent the corresponding residual stress distribution of thethree pairs of two opposite flatness basic pattern.Thirdly, the flatness prediction based on distributed CMAC neural network isestablished. First, by using the basic parameters of flatness, the model can cluster the basic information. According the sub learning data, the model can establish a corresponding subCMAC neural network to learn the data. Finally, the model can establish a generaldistributed CMAC neural network. In the forecasting process, we can use the sub CMACneural network to predict the corresponding data.At last, the simulation experiments of the flatness prediction based on distributedimproved CMAC neural network are realized by using MATLAB software. Thesimulation data are actual producing data and taken from the rolling process of a1220four-roller five-stand cold strip mill.