A deep learning approach to predict fretting fatigue crack initiation location

Abstract

Fretting fatigue is a destructive failure that usually occurs in the parts or components in the aviation field. One of the important research topics in fretting fatigue is the prediction of crack initiation location. This paper proposed a new prediction methodology of crack initiation location, which is a deep learning model based on the statistical bootstrapping technique and finite element method (FEM). In this work, two kinds of deep learning models, namely Model-I and Model-II, are established based on different input features, and with the crack initiation location as output feature. The input features of Model-I are the contact load, the tangential load, the cyclic axial stress, and the radius of the fretting pad. In Model-II, one additional feature, namely the half contact-width, is considered. Firstly, three different normalization methods (Max-min, Zero-mean, and Nonlinear) are compared. It is found that the prediction of deep learning models based on the Zero-mean is the most accurate and stable. Then, through the study of sample size reliability, it is found that the prediction accuracy gradually increases initially then becomes stable with the increase of the sampling number. The estimated intervals also gradually become narrowed and then stable. Further, another highlight from the comparison between the performance of Model-I and Model-II is that with the physical-mechanical reasoning factor as additional input features, a more stable and accurate estimated crack initiation location can be obtained from deep learning models. The range of estimated intervals is also effectively narrowed. As an alternative to FEM, the computational efficiency of deep learning is at least 30% higher. The range of estimated intervals for crack initiation location are limited to 250 μm and the errors between predictions and target values are lower than 5%. Therefore, the deep learning method proposed in this paper can effectively and efficiently predict fretting fatigue crack initiation location.