Keywords: block shear, steel structures, neural networks, back-propagation, parametric analysis, bolted joints.

High strength bolts are extensively used to connect steel structural elements in a great variety of applications. The joint resistance is generally achieved, due to the bolt high strength, with a reduced number of bolts, located in a small plate area. This fact induces the bolted connections to be associated to a failure mode known as block shear where a block area of the connected element is separated from the remaining parts.

Many models have been proposed to avoid shear block failure in steel design. These models incorporate a combination of the various hypotheses that control the structural rupture. In the majority of cases it involves a combination of rupture and yielding in the tension and shear planes of the designed joint. Various design formula for this structural engineering problem were proposed all over the years but a closed solution has not yet been reached, due to the influence of several independent parameters. Many studies [1,2,3,4], have already been carried out with the available experimental data, however, non negligible errors are still present in the current design formulae.

The main objective of the present paper is to present a back propagation neural network evaluation of the block shear phenomenon. The main neural network input parameters were the geometrical and material variables that control the block shear design. The neural network output parameter was the block shear capacity.

The neural network training was based on experimental results present in literature and adopted cross validation techniques to avoid the neural network overfitting. The present investigation was performed with the aid of the MATLAB software [5]. Four Matlab toolbox training functions were used to determine the neural network with the best performance. The study was centred on the determination of best performance for the neural network with ten inputs, five neurons on the hidden layer and one output.

The neural network results were associated to reduced associated errors and confirmed the possibility of using this methodology to generate trustworthy data. The new data, coupled with the existing experiments, can help in the production of more accurate design formulae. Future steps will use Bayesian neural networks to simulate the block shear capacity to help improve its performance.

1

Orbison J.G., Wagner M.E., Fritz W.P., "Tension plane behavior in single-row bolted connections subject to block shear", Journal of Constructional Steel Research, 49(3), 225-239, 1999. doi:10.1016/S0143-974X(98)90172-9

2

Franchuk C.R., Driver R.G., Grondin G.Y., "Experimental investigation of block shear failure in coped steel beams", Canadian Journal of Civil Engineering, 30(5), 871-881, 2003. doi:10.1139/l03-044

3

Topkaya C., "A finite element parametric study on block shear failure of steel tension members", Journal of Constructional Steel Research, 60(11), 1615-1635, 2004. doi:10.1016/j.jcsr.2004.03.006

4

Driver R.G., Grondin G.Y., Kulak G.L., "Unified block shear equation for achieving consistent reliability", Journal of Constructional Steel Research, 62(3), 210-222, 2006. doi:10.1016/j.jcsr.2005.06.002

5

MATLAB (c) The Mathworks Inc, version 7.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £62 +P&P)