Optimizing neural network prediction of composite fatigue life under variable amplitude loading using bayesian regularization

M. I.P. Hidayat, P. S.M.M. Yusoff

Research output: Chapter in Book/Report/Conference proceedingChapterpeer-review

3 Citations (Scopus)


Neural networks (NN) found its application in fatigue field, especially in fatigue life assessment of composite materials in recent years. The use of NN in the field of application also implies the necessity for optimizing the NN prediction of the composite fatigue life with respect to the presence of limited fatigue data available and fatigue condition of varying stress amplitudes. In the present chapter, optimizing NN prediction of fatigue life under variable amplitude loading (various stress ratio conditions) in relation with the availability of limited fatigue data for polymeric-based composites is presented. Multilayer perceptrons-based NN model is developed, and the training algorithm of Levenberg-Marquardt incorporating adaptive Bayesian regularization is used in the present study. From the simulation results obtained, it can be shown that training the developed network with fatigue data of only two stress ratios, which represent limited fatigue data, gave reasonably accurate fatigue life prediction under wide range of stress ratio values. The reliability and accuracy of the NN prediction were quantified by small mean square error (MSE) values. Finally, when using much less training fatigue data (22% from the total fatigue data), the network can still produce significant coefficient of determination between the prediction results and those obtained by the experiment.

Original languageEnglish
Title of host publicationComposite Materials Technology
Subtitle of host publicationNeural Network Applications
PublisherCRC Press
Number of pages30
ISBN (Electronic)9781420093339
ISBN (Print)9781420093322
Publication statusPublished - 1 Jan 2009


Dive into the research topics of 'Optimizing neural network prediction of composite fatigue life under variable amplitude loading using bayesian regularization'. Together they form a unique fingerprint.

Cite this