Evaluation of damage tolerance is significant in aerospace structures, while predicting the fatigue life accurately and promptly is greatly in need for complex practiced structures. Although traditional experiment and analysis based on fracture theory have been well established and have accumulated considerable data, time consumption and diseconomy are still encountered. To address these challenges, a new data-driven approach based on machine learning is presented. The back propagation neural network is trained and adopted to predict the basic curve da/dN-ΔKeff, i.e., fatigue crack growth rate versus the crack closure-based effective stress intensity factor range, and the fatigue life of surface cracked specimens under constant amplitude cycle loading. The influence of different number of neural units and layers on effectiveness of the trained neural network are investigated, Moreover, the neural network model is compared with other regression models. The test result shows that the trained neural network model is superior to other models and can obtain material data and fatigue life accurately and rapidly. Finally, through the active learning algorithm of query by committee algorithm, we select a small number of samples for training to obtain a model that meets the accuracy. This model can directly bridge the gap between the preliminary detailed stress analysis and the final fatigue life, and greatly reduce the cost of obtaining data sets. This provides a new way to solve engineering problems.
Read full abstract