The Research On Inverse Problems In Sheet Metal Forming Processes Based On Neural Network

Sheet forming is a very important technology in manufacturing system, which is widely applied in automobile, aviation, national deference industry and so on. Mould design is the key point in sheet forming technology. Now, the mould is designed empirically and sometimes the design project is verified using computer aided engineering (CAE) technology in the engineering domain. In this article, the researches on the methods of solving the inverse problems in sheet forming processes have been performed, regarding the mould design as inverse problems, such as the identification of the mould parameters and blank design from known part size and expected forming effect. The neural network (NN), a commonly used parameter identification method, was studied deeply. In order to improve the generalization performance, three structure design methods of multilayer feedforward neural network (MFN) based on generalization performance were developed. These methods were applied to solve the typical inverse problems in sheet forming processes such as drawbead geometric parameter identification, blank design and variable blank holding force identification. The feasible or guide design project could be obtained quickly by inverse solution.The main innovative points of this paper are given as follows:(1) Several improving methods of generalization performance were introduced to improve the generalization performance of MFN and make MFN become more robust, which can be depicted as the compositive application of cross validation method, early stopping method, samples contaminated with noise and regularization method.(2) From the view point of global optimization, the research on the structure optimization of the commonly used MFN with the sigmoid function as the activated function of neurons in the hidden layers, using genetic algorithm (GA), was done. The structure optimization design method, SOMFNGA (Structure optimization of multilayer feedforward network using genetic algorithm), was developed and programed. Based on IP-μGA (a modified micro genetic algorithm with the strategy of Intergeneration Projection), SOMFNGA designed structure of MFN, corresponding to given training sample set and test sample set. During the structure optimization, the compositive application of several improving methods of generalization performance mentioned above, the determination method of learning ratio, momentum factor, jumping factor and regularization coefficient adopting the quick searching mechanism were employed to train the NN with the structure, the individual in solution space. The research results of...