A neural network model with finite element method for steel beams design

Resumo

This work aims to create an artificial neural network model to assist steel beam design either beam profile
and its connection with the column. The input data for the model are the beam length and the applied distributed
load. The output data are the beam profile, the connection angle profile, and the number and diameter of connection
bolts. We selected for training 10 W-type profiles, 3 angle-type profiles, and 2 bolt diameters. The data for training
the neural network has been acquired from the design results according to the Brazilian standard NBR-8800:2008
criteria. The bending moment and shear forces have been calculated from beam internal stresses. The internal
stress diagrams have been obtained from the finite element method (FEM) results. The beam analysis with the
FEM has been carried out with a flat shell quadrilateral element (plane stress plus Kirchhoff-Love plate effects)
for profiles and a three-dimensional frame element for bolts. All nodes at the hole edge and the bolt element in the
same plane are coupled. The modeled neural network has been evaluated with its confusion matrix and its accuracy
in indicating configurations that attend the design criteria. The results show a good prediction performance and
errors obtained are acceptable when compared to the level of safety factors of structural engineering.