Performance of fatigue life models based on reliable estimators and an artificial neural network for materials

• Ano de defesa: 2019
• Autor(a) principal: Barbosa, Joelton Fonseca
• Orientador(a): Freire Júnior, Raimundo Carlos Silvério
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Programa de Pós-Graduação: PROGRAMA DE PÓS-GRADUAÇÃO EM ENGENHARIA MECÂNICA
• País: Brasil
• Palavras-chave em Português: Fatigue; Mean stress; High cycle fatigue; Neural network; Probability
• Área do conhecimento CNPq: CNPQ::ENGENHARIAS::ENGENHARIA MECANICA
• Link de acesso: https://repositorio.ufrn.br/jspui/handle/123456789/28092

Mechanical failures of equipment and structural components cause loss of performance of the required function and unexpected shutdowns, leading to an increased need for corrective maintenance, which increases maintenance costs and decreases the reliability of mechanical systems. The mean stress effect plays an important role in the fatigue life prediction, its influence significantly changes high-cycle fatigue behaviour (HCF), directly decreasing the fatigue limit value with increasing mean stress. Geometric discontinuities - such as cross-section shifting, holes, notches, keyways, among others - cause a considerable increase in the value of nominal stress acting in the adjacent vicinity of the stress concentrator. This enhances the positive mean stress effects on damage over the life cycle of the material, directly influencing the design fatigue strength reduction factor (Kf). Numerous empirical models, such as Gerber, Goodman, Soderberg, and Morrow, have been developed to correct the mean stress effect, but despite advances, there is no unified model in the literature that considers the stochastic behaviour of fatigue failure for prediction of the maximum means stresses supported in the high cycle region for the structural details. Thus, the purpose of this work is to develop a new probabilistic constant life diagram model based on an artificial neural network applied to metallic materials and structural details, capable of estimating the fatigue resistance reduction factor for different mean stresses. The results show that trained neural network was able to determine regions of material operation reliability under the aspects of mean stress, stress amplitude and stochastic behaviour of a number of cycles to failure. In addition, it was possible to estimate the values of the fatigue strength reduction factor corresponding to the strength limit using a small amount of experimental data.