Fatigue Life Assessment of Polymer-Matrix Composites Under Variable Amplitude Loading Using Neural Networks

In the current study, Multi Layer Perceptron (MLP) based neural networks (NN) model with one hidden layer was utilized to predict fatigue lives of polymer-matrix composites under wide spectrum of fatigue stress ratios. E-glass/epoxy (Material I), E-glass/polyester (Material II) and AS4/PEEK (Material III) composites with, respectively, six, nine and six stress ratio values from the corresponding fatigue database were employed. Two stress ratios of each material were selected as the training set which represented a limited set of stress ratios. The Levenberg-Marquardt training algorithm which implemented Bayesian regularization was utilized to deal with the limited training set. In addition, the quality of the fatigue lives predicted by the NN model compared to those obtained by the experiments or other method was measured using mean square error (MSE) value.

The prediction results showed that the NN model yield consistent and reasonably accurate fatigue life prediction under a wide range of stress ratio values for all the materials. Also, the fatigue life prediction was still consistent when the stress ratios of the training set were varied from Tension-Tension (T-T) to Compression-Compression (C-C) fatigue mode. Due to the simplicity of T-T fatigue testing with no buckling supports, the use of T-T training set may be preferred compared to T-C and C-C training sets. Moreover, the MSE values obtained in the current study were comparable with that of the previous work benchmarked. The optimum number of hidden nodes for Materials I, II and III was respectively 15, 6, and 30, with the corresponding optimum MSE values were 0.108, 0.185 and 0.117. Related to the optimum MSE values, it is observed that the best prediction results were achieved when using training set with R values far separated or symmetrical position in the CLD region. The strategic position of the stress ratios provided the best distribution of the fatigue information which in turn resulted in the best prediction results among the other training sets.