A novel deep learning approach of multiaxial fatigue life-prediction with a self-attention mechanism characterizing the effects of loading history and varying temperature. (September 2022)

• Record Type: Journal Article
• Authors: Yang, Jingye; Kang, Guozheng; Kan, Qianhua
• Is Part Of: International journal of fatigue. Volume 162(2022)
• Publication Date: 2022-09
• Subjects: Multiaxial life prediction -- Deep learning -- Self-attention mechanism -- Loading history -- Varying temperature; Materials -- Fatigue -- Periodicals; Materials -- Fatigue; Periodicals; 620.1122
• DOI: 10.1016/j.ijfatigue.2022.106851
• Languages: English
• ISSNs: 0142-1123

Abstract

Highlights: A deep-learning based novel approach is established for multiaxial fatigue life prediction. A advanced deep learning mechanism called as self-attention is incorporated to characterize the effects of complex loading history and varying temperature. The capability of the established approach is verified by dealing with the proportional and non-proportional multiaxial loading paths. The effect of random multiaxial loading with complex history on the fatigue life is successfully described. For a thermo-mechanical case, the synergistic effect of multiaxial loading and varying temperature on the fatigue life is also well characterized. Abstract: A novel deep learning approach is established in this work to directly model the highly nonlinear mapping between the complex loading conditions (input) and the multiaxial fatigue life (output). An advanced deep learning mechanism, named as self-attention mechanism, is incorporated in this approach to characterize the effects of complex loading history and varying temperature on the fatigue life. Three typical examples are performed to verify the capability of proposed approach to achieve the mechanical and thermo-mechanical multiaxial fatigue life-predictions. The results demonstrate that both the effects of loading history and varying temperature on the multiaxial fatigue life are reasonably captured, and the predicted lives by the proposed approach are almost located within the scatter band of 1.5 times.