Comparative estimation of models with different structures for estimation of mechanical properties of light steel profiles
The paper presents linear, quadratic, signomial and radial-based neural networks for the estimation of the mechanical properties of steel profiles for construction obtained from the chemical composition of the batches, the cross-section of the profile to be laminated, for the lamination workshops taken as case studies. As primary information, a database with the batches produced in the Antillana de Acero rolling mills is used for more than ten years. The results obtained show that the radial base neural networks applying Landweber's iterative regularization method to network training provide the highest precision. The signomial, quadratic and linear models reach similar values of precision taking as a criterion of comparison the standard deviation of the estimate with respect to the results of the passive experiments obtained from the quality control of the production. The modeling work is done for the case studies of the laminating workshops 250 and 300 of the steel company Antillana de Acero.