Fatigue life prediction considering mean stress effect based on random forests and kernel extreme learning machine

• Fatigue life mapping in presence of mean stresses is explored based on RF and KELM . • Genetic algorithm and Grid search method are integrated to optimize hyper-parameters. • A rich experimental database is established for model training and evaluation. • The proposed models are demonstrated to be superior to semiempirical models. • The RF- and KELM-based models are favorable for different application scenarios. The mean stress effect plays a vital role in fatigue life analysis, affecting both macro-mechanical response and micro-crack evolution of materials. Even though semiempirical models are widely used in practice because of their simplicity, the mean stress effect for a broad range of materials and loading conditions may not be uniformly reflected due to the disorganized description of fatigue damage. To overcome such deficiency, two machine learning (ML)-based models, as useful alternatives to semiempirical models, are proposed to predict the fatigue life in presence of mean stresses using random forests and kernel extreme learning machine respectively. In the models, the monotonic, cyclic and fatigue properties as well as the cyclic stress–strain responses of materials are employed to map the fatigue life with mean stress effect. Also, hyperparameters are automatically optimized by the genetic algorithm/grid search method to avoid arbitrary procedures. A total of 354 experimental results generated for various materials and mean stress levels are collected to evaluate the prediction accuracy, performance stability and generalization ability of the proposed models. It is shown that these two proposed models, based on different prediction mechanisms, can both exhibit superior prediction performance over semiempirical models, and hold distinct prediction characteristics favorable for different application scenarios.