Investigation of Prediction Method of Low Cycle Fatigue Life using Machine Learning and Numerical Model

Abstract

In this study, a method for predicting low-cycle fatigue life using machine learning and a numerical model was investigated. The relationship between the remaining fatigue life and the distribution of fatigue cracks was calculated under various crack initiation parameters using the numerical model that simulates small crack initiation, growth and coalescence under low cycle condition. A prediction model for the remaining fatigue life was constructed using machine learning. Simulation results were used as training data for the machine learning model. The fatigue life from actual low-cycle fatigue tests was also predicted using the machine learning model. The remaining fatigue life predicted by the machine learning model was closer to the actual values than that predicted by fracture mechanics based on a single crack. In addition, the effect of parameters used for generating the training data on the prediction accuracy was investigated. The results showed that the number of training samples had a significant impact on fatigue life prediction accuracy. In particular, the number of simulation conditions had the largest effect. All parameters used in this study (number of cycles, number of cracks, surface crack length, and the sum of surface crack lengths) contributed to reducing the prediction error of remaining fatigue life.