A Unified Multiaxial Fatigue Life Prediction Method Based on Transformer Sequence‐to‐Sequence Translation
TL;DR
A transformer‐based multiaxial fatigue life prediction method, innovatively formulating the task as a sequence‐to‐sequence transformation problem, which consistently outperforms traditional LSTM‐MLP models in both prediction accuracy and generalization capability under current test conditions.
Abstract
This paper proposes a transformer‐based multiaxial fatigue life prediction method, innovatively formulating the task as a sequence‐to‐sequence transformation problem. The model takes fused features (comprising strain histories and material properties) as input, and employs a multi‐head self‐attention mechanism in the encoder to capture the synergistic effect of loading sequences and material properties on fatigue life. The decoder performs stepwise prediction by combining the modeled fatigue life increment sequence with a masked self‐attention mechanism and an autoregressive mechanism. The approach automatically learns long‐range dependencies and cross‐channel contextual features within loading histories, eliminating reliance on manual feature engineering or empirical cycle counting methods. Experiments on three representative cases (multiaxial fatigue for multiple materials under multiple loading paths, thermo‐mechanical fatigue and variable‐amplitude fatigue) demonstrate that the proposed model consistently outperforms traditional LSTM‐MLP models in both prediction accuracy and generalization capability under current test conditions.