Prediction of fatigue crack propagation lives based on machine learning and data-driven approach

Numerous influence factors will lead to the inaccurate prediction of fatigue crack propagation (FCP) life of the metal structure based on the existing FCP model, while the prediction method based on machine learning (ML) and data-driven approach can provide a new idea for accurately predicting the FCP life of the metal structure. In response to the inconvenience of the online prediction method and the inaccuracy of the offline prediction method, an improved offline prediction method based on data feedback is presented in this paper. FCP tests of reduced scale models of balcony opening corners in a cruise ship are conducted to obtain experimental data with respect to the a - N curves. The crack length corresponding to the cycle is trained using a support vector regression (SVR) and back propagation neural network (BP NN) algorithms. FCP prediction lives of test specimens are performed according to the online, offline, and improved offline prediction methods. Effects of the number of feedback data, the sequence length (SL) in the input set, and the cycle interval on prediction accuracy are discussed. The generalization ability of the proposed method is validated by comparing the prediction results with the experimental data in the literature. The larger the number of feedback data, the higher the prediction accuracy. The results show that 1/5 and 1/2 feedback data are needed in the SVR and BP NN algorithm with SL is 5, respectively. Furthermore, the SVR algorithm and SL = 5 are recommended for FCP life prediction using the improved offline prediction method.