A process-data-driven BP neural network model for predicting interval-valued fatigue life of metals

The experimental observations of fatigue life of metals always exhibit uncertainty even under the same settings. How to effectively capture the uncertainty when predicting fatigue life of metals? This paper proposes an interval-valued fatigue life prediction model for coping with the challenge of uncertainty. First, a novel back-propagation (BP) neural network model is established, where the data extracted from fatigue failure process of metals is used. The fatigue life of metals is predicted as an interval by considering the measurement errors of mechanical loadings and geometry of materials. Second, the low cycle fatigue life of un-notched/notched Q235 steel is predicted using the proposed model. The real- and interval-valued predictions of fatigue life are reported and compared with the experimental observations. It is found that the prediction accuracy is enhanced by extracting the data from fatigue failure process. Third, the performance of the proposed method is analyzed under different data sets and numbers of neurons together with network layers. The appropriate numbers of neurons, network layers and cycles are reported for predicting fatigue life with enough accuracy. Comparisons with finite element method (FEM) and critical distance method (CDM) reveal the advantages of the proposed model in predicting fatigue life of metals.