Artificial neural networks (ANNs) are employed as an alternative to physical modeling for calculation of the relations between the production path process parameters (melting of scrap steel and alloying, continuous casting, hydrogen removal, reheating, rolling, and cooling on a cooling bed) and the final product mechanical properties (elongation, tensile strength, yield stress, hardness after rolling, necking) of steel semi products. They provide a much faster technique of response evaluation complementary to physical modeling. The Štore Steel company process path for production of steel bars is used as an example for demonstrating the new approach. The applied ANN is of a multilayer feedforward type with sigmoid activation function and supervised learning. The entire set of 123 process parameters has been reduced to 34 influential ones and 1879 data sets from the production line have been used for learning. The results of parametric studies performed on the ANN based model seem consistent with the expectations based on industrial experiences. However, further improvements in data acquisition and analytical procedures are foreseen in order to obtain a methodology, reliable enough for use in the everyday industrial practice. The methodology seems to be for the first time applied in the through process modeling of steel production.
COBISS.SI-ID: 2601467
ŠTORE STEEL ltd faces a problem of production of a huge amount (approximately 1400) of different steel compositions in a relatively small quantities (approximately 15 tons). This production is performed in batches of predetermined quantities (50-53 tons). The purpose of this paper is to present the methodology for optimizing the production of predetermined steel grades in predetermined quantities before a customer’s set deadline in such a way as to reduce the non-planned and ordered quantities with the date before the deadline and minimize the number of batches. The particle swarm method was used for the optimization. The results of the research have been used in practice since 2006 with reducing the non-planned and ordered quantities from 17.17% up to 10.12% since then.
COBISS.SI-ID: 2410747
We describe the original development of meshless methods for calculation of diffusion problems on extremely non-uniform distribution of nodes, used in extremly non-uniform node distributions. A least squares method is used instead of collocation in such situations. The described research and experiences gained will be incorporated in multiscale (micro/macro) simulation systems for casting, rolling and heat treatment of steels.
COBISS.SI-ID: 1998331