An Study On Model Of Neural Network Identification Based On Rough Sets Theory In Sheetmetal Forming

The intellectualization of sheet metal forming is the process in which we use intelligent ways ,for example ,using neural network to predict material property parameters(elastic modulus, coefficient of hardening , coefficient of intensity , coefficient of anisotropy )by easy-monitored parameters, combined with best force-travel curve to form automatabilely . The identification of material property parameters is a great problem in material processing intellectualization. The excellent ability of data reduction of Rough set provide for neural network more proper structure and more proper sample data. Rosetta is a special software of data processing based on Rough set theory and has some powerful functions, for example, data discretization and data reduction.The relations necessary absolutely and relations not necessary absolutely can be discovered based on the result of Rosetta .Neural networks reduced by rough sets have fewer inputs and outputs and are more efficient than original neural networks. Rosetta software's analysis is confirmed by the excellent convergence of neural network experiments. And for deep drawing and rectangular box ,the relative prediction error precisions of non-sample data are all below 5%, even as low as 0.1586%.That shows rough neural network is better than original BP network and fuzzy neural network.In this paper parallel micro neural network is presented for the first time to replace original neural network. Using the excellent ability of classification of rough sets, combined with the nonlinear approach ability of neural network, parameters identification in bending intelligent control is handled by two micro neural networks. Every micro neural network reduced by rough sets has fewer inputs and outputs and is more efficient than original neural network. Satisfactory results are acquired compared with original neural network.. The relative prediction error precisions of non-sample data are all below 5%, even as low as 0.058%. The parallel micro neural network can be a new idea of handling those problems that have many inputs and outputs and problems thatcan not convergence.