A novel generalization ability-enhanced approach for corrosion fatigue life prediction of marine welded structures

Abstract

A generalization ability-enhanced approach for corrosion fatigue life prediction was proposed. The data augmentation and analysis of influence factors were conducted via Borderline-SMOTE and XGBoost algorithms to improve the prediction accuracy. The results of weight analysis were integrated into DCNN model via Attention mechanism. The average error and error standard deviation of the proposed model were more than 89.8% and 87% smaller than that of exiting models, respectively. The average gradient-signal-to-noise-ratio was introduced to evaluate the generalization ability of the proposed model, which was significantly higher than that of the linear regression. The predicted results under different stress ratios, frequencies and environmental conditions emphasized the good prediction accuracy and generalization ability of the proposed model. Consequently, the proposed method could reduce the consumption of fatigue test and evaluate the reliability of engineering components, which could provide a technique support for the intellectualization and digitization of fatigue behavior of offshore platforms welded joints.