Accurate prediction of cyclic hysteresis behaviour of RBS connections using Deep Learning Neural Networks. (15th November 2021)

• Record Type: Journal Article
• Authors: Horton, Thomas Alexander; Hajirasouliha, Iman; Davison, Buick; Ozdemir, Zuhal
• Is Part Of: Engineering structures. Volume 247(2021)
• Publication Date: 2021-11-15
• Subjects: Cyclic hysteresis -- Neural Networks -- Reduced Beam Sections -- Cascade forward-feed neural network -- Modified-Ibarra–Krawinkler models; Structural engineering -- Periodicals; Structural analysis (Engineering) -- Periodicals; Construction, Technique de la -- Périodiques; Génie parasismique -- Périodiques; Pression du vent -- Périodiques; Earthquake engineering; Structural engineering; Wind-pressure; Periodicals; 624.105
• DOI: 10.1016/j.engstruct.2021.113156
• Languages: English
• ISSNs: 0141-0296

Abstract

Abstract: Reduced Beam Section (RBS) connections have been widely used to improve the seismic performance of steel framed buildings by providing a seismic fuse mechanism. While in general the modified-Ibarra–Krawinkler (mIK) model can reliably capture the cyclic hysteresis behaviour of fully welded RBS connections, there is currently no reliable or accurate method available to predict the parameters which define this model without appropriate experimental or finite element (FE) tests of full scale models of the RBS connection for calibration purposes. This paper presents, for the first time, an accurate method of predicting the mIK parameters using a number of different deep learning Neural Networks through a logical process to provide accurate predictions of the required parameters based on the geometrical dimensions of the steel beam and RBS. The proposed networks are trained based on the database of 1480 experimentally validated FE models. First, the cyclic moment–rotation-hysteresis results from the selected database are calibrated with simple equivalent beam representations in the OpenSees software. Then a number of different deep learning Neural Networks are developed to predict each of the seven key parameters which define the mIK spring model. Finally, a Matlab toolbox is developed to predict the mIK parameters for any steel beam and RBS geometry without the need for complex and time consuming full scale cyclic experimental tests or FE analyses. It is shown that the developed tool box provides over 96% accuracy in predicting the key design parameters and therefore should prove useful in the preliminary design and assessment of steel RBS frames. Graphical abstract: Highlights: Deep learning framework developed to simulate the cyclic response of RBS connections. Comprehensive database of 1480 calibrated cyclic mIK models used for training purposes. A logical process proposed to determine key parameters given the geometrical input data. The ensemble-bootstrap-aggregating algorithm provided 96% accuracy to predict buckling. A set of cascade forward-feed NN showed 98% accuracy to estimate key mIK parameters.