The Research Of Improved CMAC Network Flatnese Forecast Model Based On Rough Set

Steel is an important foundation for developing the heave industry of a country, asone of the main products of steel, flatness strip quality has become an important symbol tomeasure the competitiveness of the steel industry of a country. With the development ofmodern industry, the requirement of flatness accuracy is more and more high, and theaccuracy requirement of automatic flatness control is also getting higher and higher. Theprecise flatness prediction model as the foundation and the premise of flatness control cannot only provide accurate and reliable the flatness information for automatic flatnesscontrol system, but also it is an indispensable part of the control experiment of the virtualrolling mill. So it has become increasingly urgent need to solve the problem for improvingthe prediction accuracy of flatness prediction model, in order to improve the flatnesscontrol precision. Based on the complexity of the rolling environment and analyzing thecurrent research situation at both home and abroad, to shape prediction model are studied,the paper makes a study of flatness forecast model.Firstly, multilayer feed forward neural network prediction model has the weakness incomplicated network structure and slow convergent speed. The CMAC NN has theadvantages of a local update neural network and simple structure, therefore it can rapidlearn any nonlinear function and its approximate ability is insensitive to the order oftraining data a CMAC neural network forecast model has been established.Secondly, according to the multidimensional of flatness influencing factors andstructure complexity of high dimensional CMAC network structure complexity, usingrough set theory attribute reduction method to extract the main flatness factors, the inputdimension of CMAC network is reduced and simplify the network structure, thus themodel forecast performance is improved.Again, according to the CMAC network learning algorithm with error meandistribution and fixed learning rate, making the network convergence speed slow and notsearching the optimal solution of the problem through learning, the error credit allocationand dynamic learning rate are introduced into CMAC flatness prediction model. The improved forecasting model avoids the corrosion on the historical study results in thecourse of weights adjusted and improves convergence speed of the flatness prediction.Meanwhile shape prediction learning speed is speeded up by dynamic learning rate.Finally, simulation experiment is carried out about CMAC network flatness forecastmodel that are put forward by Matlab, and proves the effectiveness of the method.