Deep learning-based procedure for structural design of cold-formed steel channel sections with edge-stiffened and un-stiffened holes under axial compression

Abstract

This paper proposes a framework of deep belief network (DBN) for studying the structural performance of cold-formed steel (CFS) channel sections with edge-stiffened/un-stiffened web holes, under axial compression. A total of 50,000 data points for training the DBN are generated from elasto plastic finite element analysis, which incorporates both initial geometric imperfections and residual stresses. A comparison against 23 experimental results was conducted, and it was found that the DBN predictions were conservative by 3% for columns with un-stiffened web holes, and 8% for columns with edge-stiffened web holes. When compared with Backpropagation Neural Network (a typical shallow artificial neural network) and linear regression model based on PaddlePaddle, it was found that the proposed DBN outperformed better than both the methods, using the same big training data used in this paper. When the same comparison was made for Effective Width Method and Direct Strength Method, the results from them were conservative by 5% and 12% against the experimental results, respectively for columns with un-stiffened web holes. Hole effects on the structural performance of channel sections under axial compression were also investigated. Based on the DBN output data, design recommendations of axial capacity enhancement/reduction factors were given for columns (stub, intermediate and slender) with edge-stiffened/un-stiffened web holes. Based on DBN prediction data, a comprehensive reliability analysis was conducted, which shows the proposed equations can predict the enhanced and reduced axial capacity of CFS channel sections with edge-stiffened/un-stiffened web holes accurately.