Optimization and control of metal forming processes

Abstract

Inevitable variations in process and material properties limit the accuracy of metal forming processes. Robust optimization methods or control systems can be used to improve the production accuracy. Robust optimization methods are used to design production processes with low sensitivity to the disturbances in the process. Efficient robust optimization procedures have been developed to determine a robust design of the production process with limited computational resources. The procedure is based on the use of an approximation of a finite element (FE) model and sequential improvement of the approximate model through additional evaluations of the FE model. The second approach for improvement of production accuracy is the use of control systems. The resulting improvement of production accuracy depends on the rate of variation in the process disturbances. Slowly varying disturbances can be controlled with feedback control, whereas disturbances which vary from product to product require feedforward control to be eliminated. When using feedforward control, process measurements are used to estimate the effect of product-to-product variations on the final product. In this work, it is studied whether force measurements can be used in a process estimator for feedforward control. Such a process estimator may either be built based on historical data from the process or may be determined with an FE model of the process. The effectiveness of control of metal forming processes is studied based on data from a demonstrator process with multiple forming stages. Several large datasets are used to investigate the feasibility of feedforward control for the demonstrator process. A significant part of the variations in the process can be predicted based on force measurements using LASSO regression. Another approach for building a process estimator is using an FE model of the process. The FE model is used to identify the causes of small variations in the force measurements and to predict their effect on the final properties of the product. The proposed procedure for real-time parameter estimation involves proper orthogonal decomposition, model interpolation and Bayesian estimation.