Bayesian-based probabilistic fatigue crack growth evaluation combined with machine-learning-assisted GPR

• Published in: Engineering fracture mechanics 2020-04, Vol.229, p.106933, Article 106933
• Main Authors: Hu, Dianyin, Su, Xiao, Liu, Xi, Mao, Jianxing, Shan, Xiaoming, Wang, Rongqiao
• Format: Article
• Language: English
• Subjects: Algorithms; Bayesian analysis; Bayesian method; Computer simulation; Crack propagation; Dimensional analysis; Fatigue crack growth; Fatigue failure; Fatigue tests; Fracture mechanics; Full-scale experiment; Gaussian process; Gaussian process regression; Life assessment; Life prediction; Low cycle fatigue; Machine learning; Markov chains; Mathematical analysis; Model accuracy; Probabilistic methods; Regression analysis; Spin tests; Statistical analysis; Turbine disks; Uncertainty; Uncertainty quantification
• ISSN: 0013-7944; 1873-7315
• DOI: 10.1016/j.engfracmech.2020.106933

•Propose a probabilistic FCG life prediction framework with uncertainty quantification.•Apply GPR machine learning method for uncertainty propagation.•Establish a stress altering factor to assist uncertainty quantification with complex geometry.•Employ the Bayesian method to quantify coupled uncertainties in FCG model. This paper presents a Bayesian-based calibration method that simultaneously improves the model accuracy and the computational efficiency for fatigue crack growth (FCG) life prediction on turbine discs. Uncertainties derived from geometry, material and models are elaborately quantified based on the data from measurements and experiments. A Bayesian approach is used for uncertainty quantification, where Markov Chain Monte Carlo algorithm is employed to estimate posterior distributions. Gaussian process regression (GPR) is introduced to describe the propagation of uncertainties and improve the efficiency in high-dimensional analysis. With the integrated methodology, uncertainties are embodied in life prediction results of a whole turbine disc. A full-scale spin test is carried out under low cycle fatigue loading, where three effective crack samples are generated and monitored. Compared with the experimental results, the mean values of the predictions are bounded within a factor of ±2.0, validating the potential usage of the proposed method in the probabilistic FCG life assessment.