Forecasting Of The Ultimate Resistance Of Steel Beams Subjected To Concentrated Loads Using Data Mining Techniques

Abstract

In the conception and design of civil engineering structures several factors should be considered: aesthetics, functionality, deformability, durability, resistance and cost. In general, that exercise is conditioned to the search of the safest solution at low cost. This concern, associated to the evolution of the materials' properties and of the computational tools, has been leading to the use of more and more refined design methods. In the special case of steel structures with very slender sections, the error presented by the current design formulas to forecasting the ultimate resistance of steel beams subjected to concentrated loads is significant, due to: the influence of several independent parameters in the behaviour; the insufficient number of experimental data that allows a parametric analysis; and the calibration of simplified models. While parametric analysis is intensive and hard work, the construction of models using Data Mining (DM) techniques in a Knowledge Discovery from Databases (KDD) process could be used to induce predictive models in a more flexible and efficient fashion. This paper exploits the above described approach to predict the ultimate resistance of steel beams subjected to concentrated loads. The work involved a macro-analysis phase to discover clusters based on Kohonen Networks, followed by a micro-analysis phase based on the C5.0 algorithm and Artificial Neural Networks (ANN), to induce models better suited to each cluster. The accuracy attained by the models is indubitably better than the current design formulas, presenting a mean error lower than 10%.