A TCN-based feature fusion framework for multiaxial fatigue life prediction: Bridging loading dynamics and material characteristics

Abstract

Multiaxial fatigue represents one of the most prevalent and critical fatigue issues in engineering applications, yet its life prediction remains challenging due to the combined effects of material characteristics and dynamic loading paths. This study proposes a cross-material adaptive framework that innovatively fuses dynamic loading sequence features with material properties through an attention-based mechanism. The framework employs lightweight Temporal Convolutional Networks (TCN) for temporal feature extraction while considering frequency-domain features, and incorporates both static and cyclic material properties as structured inputs. The attention-based fusion mechanism enables adaptive integration of temporal and material features, enhancing the model’s ability to capture complex interactions between loading conditions and material characteristics. To validate this approach, a multiaxial fatigue dataset comprising 499 data points was established across six major categories of metallic materials. The model’s performance underwent rigorous evaluation using hierarchical nested cross-validation, while key features were identified through SHAP analysis and recursive feature elimination. Results demonstrate that the proposed unified model exhibits robust predictive performance, with TCN showing superior stability and efficiency compared to other temporal feature extraction methods. Furthermore, the study explores cross-material prediction, revealing that fine-tuning with a small proportion of new material data can significantly enhance prediction accuracy. This research provides a novel approach to multiaxial fatigue life prediction while laying the groundwork for future cross-material prediction capabilities.