Deep learning-based axial capacity prediction for cold-formed steel channel sections using Deep Belief Network. (October 2021)

• Record Type: Journal Article
• Authors: Fang, Zhiyuan; Roy, Krishanu; Mares, Jiri; Sham, Chiu-Wing; Chen, Boshan; Lim, James B.P.
• Is Part Of: Structures. Volume 33(2021)
• Page Start: 2792
• Page End: 2802
• Publication Date: 2021-10
• Subjects: Cold-formed steel -- Axial capacity -- Deep learning -- Finite element analysis; Structural engineering -- Periodicals; 624.1
• DOI: 10.1016/j.istruc.2021.05.096
• Languages: English
• ISSNs: 2352-0124

Abstract

Abstract: In this study, a deep learning-based axial capacity prediction for cold-formed steel channel sections is developed using Deep Belief Network (DBN). A total of 10, 500 data points for training the DBN are generated from non-linear elasto plastic finite element analysis, which incorporated both initial imperfections, as recommended by the Australian/New Zealand Standard (AS/NZS 4600:2018) and residual stresses as recommended by Moen et al. A comparison against experimental results found in the literature was conducted. It was found that the DBN was conservative by 9%, 6% and 8% for stub columns, intermediate columns, and slender columns, respectively. When compared against a typical shallow artificial neural network (Backpropagation Neural Network) and a typical machine learning model (Linear regression model based on PaddlePaddle), it was shown that DBN performed around 2% better than both with the same training data. When the same comparison was made for both the Effective Width Method and the Direct Strength Method, it was found that they were conservative by 15%, 13%, and 15%, respectively. Based on the DBN output data, new and improved design equations for AS/NZS 4600:2018 were proposed.