The use of neural networks for the prediction of fatigue lives of composite materials

Constant-stress fatigue data for five carbon-fibre-reinforced plastics and one glass-reinforced plastic laminate have been used to evaluate possible a rtificial neural network architectures for the prediction of fatigue lives and to develop network training methods. It has been found that artificial neural networks can be trained to model constant-stress fatigue behaviour a t least as well as other current life-prediction methods and can provide ac curate (and conservative) representations of the stress/R-ratio/median-life surfaces for carbon-fibre composites from quite small experimental data-ba ses. Although their predictive ability for minimum Life is less satisfactor y than that for median life, and is non-conservative, the procedures develo ped in this work could nevertheless be used in design with little further m odification. Some success has been achieved in modelling fatigue under bloc k-loading conditions, but this problem is more difficult and requires much more effort before ANNs could be used with confidence for variable-stress c onditions. (C) 1999 Elsevier Science Ltd. All rights reserved.