Keywords: neural network, back error propagation, non-linear behaviour, dynamics concrete towers.

Recently, the non-parametric identification models such as neural networks have evolved into an important research area in system identification approach. The neural networks as efficient tools offer exciting attributes as the ability to adapt with highly nonlinear problems, fast parallel architecture and the capability of working as efficient subsystems in large nonlinear super-systems. The most widely used type of neural networks in the field of engineering are the general feed-forward back error propagating perceptrons.

In this study an artificial neural network simulator (ANNS), the general back error propagating perceptron was used to model the nonlinear behavior of hollow cylindrical concrete towers when subjected to base excitation shock. A nonlinear finite element concrete model which, is capable of simulating crack and crash in concrete, was implemented to create the input data for supervised learning procedure. The analyses were performed using the ANSYS general-purpose finite element program.

The network input units were three geometric parameters including height, outside diameter and thickness and the output units of the network were the nonlinear pick response of the tower and phase delay. The adaptive ascending algorithm or the dynamic node creation scheme was developed and employed in a computer program to construct the best network architecture. In order to pass by the problem of insufficient input data, a multiple random data generation were also used to create more training epochs. The sigmoid and Tanh activation functions were used and the results such as the learning rates were compared.

Parametric studies were performed for different network parameters including number of hidden layers, number of hidden neurons and the activation functions' parameters. It was also concluded that biases do not improve the network performance as we expect. Using various test data, the domain of application of this network or the extrapolation capability was determined. Base on various trials, some key tips in learning procedure of these networks were also concluded. From the results presented in this study it was finally concluded that neural networks show great promise in becoming useful tools in practical structural analysis applications especially for problems with severe nonlinearities in which the analysis time and cost are important factors.

1

Bishop C.M, "Neural networks for pattern recognition", Oxford University Press, Oxford, U.K., 1995.

2

Masri S.F. et al, "Application of neural networks for detection of changes in nonlinear systems", Int. Jour. of Engng. Mechanics, ASCE, pp 666-676, 2000. doi:10.1061/(ASCE)0733-9399(2000)126:7(666)

3

Joghataie A., Ghaboussi J.,Wu X, "Learning architecture determination through automatic node generation", Proc. ANNIE95, ASME, New York, 45-50., 1995.

4

Nikzad K., Ghaboussi J., "Neural network modeling of structural response under resistant external excitation", Proc. of SEE2 conf., Tehran, Iran, 1995.

5

ANSYS, General purpose finite element program, ANSYS Co., USA, 1997.

6

Ash T., "Dynamic node creation in back-propagation networks", ICS report 8901, UCSD Inst. for cognitive sciences, 1989.

7

Haykin S., "Neural networks; a comprehensive foundation", 2nd edition, Prentice Hall Inc., 1999.

8

Masri S.F., Smyth A.W., Chassiakos A.G., Caughey T.K. & Hunter N.F., "Application of neural networks for detection of changes in nonlinear systems", Int. Jour. of Engng. Mechanics ASCE, Vol. 126, No. 7, 2000. doi:10.1061/(ASCE)0733-9399(2000)126:7(666)

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 £78 +P&P)