Estimation of fatigue life of welded structures incorporating importance analysis of influence factors: A data-driven approach

Abstract

The fatigue life prediction of welded joints with different specifications under different conditions was a challenging issue due to the quite complex influence. Specifically, the current fatigue life prediction methods lacked comprehensive analysis of multiple influence factors and reasonable incorporation of physical models. So, the analysis of factors influencing the fatigue life of welded joints and the fatigue life prediction were critical to the safety and reliability of engineering structures. In this study, a prediction approach for fatigue performance was proposed based on influence factor analysis using data-driven methods. The fatigue performance dataset was processed via physical models to realize the physics-based analysis of the factors affecting fatigue performance. The weights of the physics-informed influence factors on the fatigue life were analyzed using the extreme gradient boosting (XGBoost) algorithm and verified by cross-validation. The prediction results extracted from the SN curves predicted by the deep convolutional neural network (DCNN) model incorporating the weight analysis of influence factors exhibited better accuracy and stability than direct prediction and other existing prediction models. Because of the advantages of DCNN in avoiding over fitting and local optimization, the proposed approach can better describe and estimate the fatigue performance of welded structures.