Fatigue damage prediction model for broadband non-Gaussian random processes based on Bayesian optimized random forest

Abstract

Marine structures operating in harsh ocean environments are continuously exposed to cyclic loads that may induce fatigue damage. Accurate fatigue damage prediction is therefore essential for ensuring the structural integrity of marine structures. When the structural responses exhibit broadband non-Gaussian characteristics, the existing frequency-domain methods appears to have insufficient accuracy. Although the time-domain rainflow counting method can achieve higher accuracy, it incurs high computational costs. To address these challenges, this study develops a fatigue damage prediction model for broadband non-Gaussian random processes by combining the random forest (RF) algorithm with Bayesian optimization (BO) approach. The BO-RF model is trained by employing the database including the fatigue damage related to diverse power spectra with a broad range of bandwidth parameters and a variety of S-N curve slope, skewness and kurtosis. Extensive numerical simulations demonstrate that the developed BO-RF model is superior to the tradition frequency-domain methods in terms of accuracy and robustness. Furthermore, comparative analysis against an artificial neural network (ANN) model further reveals the advantages of BO-RF model in both training efficiency and generalization capability. This study demonstrates the proposed BO-RF model can provide a feasible solution for accurate and efficient fatigue damage prediction of marine structures under broadband non-Gaussian random loading.