A frequency-domain enhanced multi-view network for metal fatigue life prediction

Abstract

Fatigue damages and failure widely exist in engineering structures. However, predicting fatigue life for various structural materials subjected to multiaxial loading paths remains a challenging problem. A novel multi-view deep learning model incorporating frequency-domain analysis for fatigue life prediction is proposed. The model consists of two main analytical components: one for analyzing multiaxial fatigue loading paths and the other for examining the mechanical properties of materials and specimen geometrical characteristics. In the module analyzing multiaxial fatigue loading paths, convolutional neural network (CNN), long short-term memory network (LSTM), and FNet are connected in parallel to extract features individually. Features of materials and specimens are extracted through fully connected neural networks (FCNNs). Subsequently, the features from these two parts are thoroughly integrated based on attention mechanisms, and connected to multiple FCNNs to accomplish fatigue life prediction. A fatigue experimental database comprising 557 samples, spanning 46 multiaxial loading paths and 19 metal materials, has been established for model training and testing. Additionally, 6 materials were respectively used as test sets to evaluate the extrapolation ability of the model. The results suggest that the proposed model exhibits robust predictive performance and extrapolation capabilities. We anticipate that the multi-view approach, along with its accuracy and applicability, can provide an unparalleled alternative for researchers in the field of engineering fatigue and beyond.