Neural prediction of buckling loads of cylindrical shells with geometrical imperfections

Abstract

The paper is a continuation of Waszczyszyn et al. (in: R. Tadeusiewicz, L. Rutkowski, J. Chojcan (Eds.), Proceedings of the Third Neural Networks and their Applications, TU Czestochowa, Poland, 1997, p. 14), where back-propagation neural networks (BPNNs) were used for predicting the buckling loads for axially compressed cylindrical shells with initial geometrical (manufacturing) imperfections. The paper was based on the measured imperfections and tests on laboratory shell specimens, gathered in the Imperfection Data Bank at the Delft University of Technology (P. II, Report LR-559, TU Delft, Faculty of Aerospace Engineering, Delft, 1988). In the presented paper the idea of data compression is explored. The application of BPNN replicators enables us to significantly reduce the number of inputs so the master BPNN formulated for the buckling load prediction is considerably smaller than BPNNs discussed in Waszczyszyn et al. (1997). It is proved that the compression of imperfection parameters seems to be a new efficient tool for the analysis of experimental data of the problem considered.