A novel method of multiaxial fatigue life prediction based on deep learning

Abstract

It is well-known that conventional multiaxial fatigue life prediction models are generally limited to specific materials and loading conditions. To remove this limitation, a novel attempt is proposed in this work based on the deep learning (i.e., an improvement of artificial neural network in machine learning approaches, which is powerful to learn representations of data with multiple levels of abstraction). To comprehensively evaluate the prediction capability of proposed deep learning-based method, six series of existing fatigue data of different materials are, respectively, analyzed, in which the main loading conditions concerned in the low-cycle and high-cycle fatigue researches are included, such as loading modes (stress-controlled/strain-controlled modes), loading levels (stress/strain amplitude and mean stress/strain), and loading paths (uniaxial/multiaxial and proportional/non-proportional paths), as well as for low-cycle and high-cycle fatigue regimes. Comparison of the predicted and experimental results shows that: all the loading conditions mentioned above can be handled satisfactorily by the proposed deep learning-based method; excellent prediction accuracy is achieved, and the predicted lives in each study case fall almost within the scatter band of 1.5 times. In addition, four groups of specifically designed data are used to evaluate the extrapolation capability of the proposed method, and the results show that the extrapolation capability gets weaker if the distinctions between the loading paths involved in the training dataset and test one increase.