Keywords: neural networks, finte element analysis, 3D-parametric modelling, timber structural design.

This paper represents the efficiency of trained Neural Networks in producing useful solutions of structural design assignments and potential of their application. It is clear that we could build necessary engineering knowledge into the NN although the possible ways for doing it might be various [5]. The prerequisite for using NN in engineering environment is a developed description of structures and a model which can be suitably translated into simply recognisable symbols sufficiently significant for NN [1]. There is not a single "best" solution to describe a design problem [3]. Classifying structures into typically recognisable groups, susceptible to the same type of analysis and repetitive design procedure, might be an appropriate one [2]. Typical structural forms provide the same kind of safety, serviceability and stability checks, issued and clearly defined in accordance with valid design codes. These facts give sense to the parametric (3D) modelling of typical structural forms, as the most suitable manner to describe them [4].

An acceptable example of Neural Networks application is in optimisation and design of all structural elements of timber structure, where glued laminated main- girders belong to a various groups of special geometrical shapes (symmetrical and non-symmetrical tapered forms, curved and pitched-cambered shapes with the constant depth of non-curvature zone). Purlines are glued laminated beams, and the bracing system has got steel diagonal elements. The whole roof structure is laid on concrete substructure. Parametric 3D timber structure modelling in FE COSMOS/M program makes all necessary changes of the span possible, as well as the distance between main-girders and their relevant dimensions (width and characteristic cross- sections depths, all in accordance with their geometry and shape). For an each type of main-girder two useful models of bracing (with four and six fields) were prepared in correlation with the structure span and main-girder distance, and with a number of purlines and roof's area as well. Steel element profiles and purline?s cross-sections were also subject of parametrically defined values. The results of FE static and buckling analysis are relevant for evaluation of the states of safety, serviceability, and stability (buckling mode) of the whole structure and its corresponding element. The most important fact is that the input parameters and final products of FEA were recognisable for NN that has been trained on those. NN models (using Ward's NShell2 (4)) generated separately for an each type of main-girder had been trained on the set pattern of an average number of ninety different cases. The tested pattern contained set of rotationally chosen dates in 15% percentage size of the whole training spreadsheet. The BackPropagation algorithm and supervised learning had been used in each generated model of 3-slabs NN, but the experiments had been done with two different configurations (Standard and Jump Connection), and also, with the number of data inputs and outputs. The most successfully NN model contained five significant input variables (span, number of bracing fields, distance between main-girders, and widths of purlines and main-girders), i.e. six inputs for the main-girders with curvature intrados which had got the radius of curvature zone as an additional input data. The number of output variables varied between 18-19. They have been separated as significant values for an engineering estimation of design acceptability. Outputs are referred as necessary geometrical properties (of critical cross-sections of all structural elements), support reactions, stresses values significant for states of safety description (of all structural elements), deformations influenced by vertical and lateral loads, and at last, buckling mode, as the global safety factor of the whole structure. The learning was interrupted after approximately three to four hours, when the average error of test patterns reached value less than 0.4%. The successful completion of an experiment was achieved when trained NN models had been claimed to produce results of their own on the basis of, until then, unseen inputs. Maximum deviation between produced (by NN) and expected values (by FEA) did not exceed 2.5-3.0%, and average values were less than the mentioned one, whereas Jump Connection configuration had been better equalised in comparison with Standard configuration of NN. The application of Neural Networks in structural engineering could be accepted as a very powerful source of new possibilities in representing and modelling of various problems. Finally, we come to the following conclusion; engineering practice produced (and still uses) a great amount of tables, and future might confirm that NN would obtain a role of "intelligent tables" on at least preliminary design level of typical structures.

1

J.H.Jr. Garret, "Artificial NN", ISMES Conference, Bergamo, Italy, 1995.

2

T.R. Tyson, "Effective Automation for Structural Design", ASCE Journal of Computing in Civil Engineering, 5(2), 132-140, 1991. doi:10.1061/(ASCE)0887-3801(1991)5:2(132)

3

D. Clarke, "Computer Aided Structural Design", J. Wiley & Sons, Ltd., 1978.

4

Z. Zagar, "Use of NN in the Design of Wooden structures", Gradevinar, 54(10), 577-583, 2002.

5

W.M. Jenkins, "NN Based Approximations for Structural Analysis", Development in NN and Evolutionary Computing, Civil Comp. Press, 23-35, 1995.

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 £82 +P&P)