Keywords: structural re-analysis, artificial neural networks, multilayer perceptrons.

In the structural design, mainly when dealing with structural optimization, sometimes it is necessary to accomplish several structural re-analyses due to significant changes in dimensions, materials and boundary conditions such as loads and restraints. The classical methods for structural re-analysis may be classified on exact and iterative methods. A good review of such methods can be found in Abu Kassim and Topping [1], where the main conclusions are that the correct choice of the method to use depends on the problems circumstances and the desired structural modifications. They conclude that there is no method which could be used with same efficiency on a wide variety of problems.

This paper implements the main ideas presented by Jenkins [2], using a multilayer perceptron network for the structural re-analysis. Jenkins [2] proposes a structural re-analysis technique using an ANN which has the main characteristics that each layer of the networks retains the previously acquired knowledge and only readapts its architecture for unknown situations. This implicates that there will exist a learning process in the network which will supply the knowledge. Afterwards, the networks will be tested to verify the learning.

Literature has shown that the advantages of this method appear for problems with several re-analyses since the processing time using the ANN is extremely fast. Another advantage in using an ANN is that this method does not depend on matrix inversion nor factorizations nor global stiffness assembling. It works with a local stiffness matrix, at the element level. Since this paper represents a first step towards the creation of a re-analysis tool using an ANN, it is analyzed only simple planar trusses. The structural re-analysis happens due to changes in loading conditions as well as changes in structural properties.

Four example planar trusses were analysed. Results for nodal displacements and nodal reactions were compared. The results show a good agreement with the finite element analysis. In terms of time consumption, the ANN spends most of the time learning the structural behaviour, however for the re-analysis it is faster than a finite element program, indicating that there will exist a critical number of samples to be re-analyzed were the ANN will be advantageous. The main results were suitable; however several improvements should be accomplished for the satisfactory use of the tool in complex structures. Nevertheless, the proposed code is able to deal efficiently with structural trusses with a large number of degrees of freedom. It was noticed that the use of the proposed algorithm has benefits when handling with structures with several degrees of freedom and several structural runs (for example in reliability analysis or Monte Carlo simulations) since the classical finite element method is faster than ANNs for each structural re-analysis. The advantages in time-consumption will be noticed once the ANN has been trained and validated.

1

A.M. Kassim, B.H.V. Topping, "Static reanalysis: a review", Proceedings American Society of Civil Engineers, Journal Structures, V.113, pp.1029-45, 1987. doi:10.1061/(ASCE)0733-9445(1987)113:5(1029)

2

W.M. Jenkins, "A neural network for structural re-analysis", Computers and Structures, V.72, 687-698, 1999. doi:10.1016/S0045-7949(98)00311-3

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)