A neural network approach to describing the scatter of S–N curves

Abstract For service life prediction of a structural component, the probability distribution of material fatigue resistance should be determined, given that the scatter of loading spectra is known and a proper damage cumulating model is chosen. In the randomness of fatigue resistance of a material, constant amplitude fatigue test results show that at any stress level the fatigue life is a random variable. Fatigue life in this instance is affected by a variety of influential factors, such as stress amplitude, mean stress, notch factor, temperature, etc. The scope of the paper is to prove that the statistical scatter of the fatigue life Nf at various factors' levels of constant values can be described by the Weibull or lognormal conditional probability density function, which is modelled with a multilayer perceptron. In order to estimate the unknown parameters of the conditional distribution, generally composed of an arbitrary but finite number of lognormal or Weibull component distributions, we introduced an algorithm based on neural network modelling. To support the main idea, two examples are presented. It can be concluded that the suggested neural computing method is suitable for describing the fatigue data trends and the statistical scatter of fatigue life under constant loading conditions for an arbitrary set of influential factors, once the optimal neural network is designed and trained.