Stress-Strain Hysteresis Loop-Based Machine Learning Models for Predicting Metal Fatigue Life Under Uncertainty.

This paper reports machine learning models for predicting metal fatigue life under uncertainty by extracting stress-strain data from hysteresis loops. First, the hysteresis loops of Q235B under strain-controlled constant amplitude loading are analyzed. The values of stress and strain in six key points are extracted from each hysteresis loop at the earliest stages of the fatigue process, and transformed into polar coordinates. Second, the uncertainty is quantified by extending the applied strain amplitude and the selected stress-strain values to intervals. A great deal of data are generated randomly in each interval for coping with the challenge of a small fatigue test dataset. Third, three machine learning models are constructed, where the parameters of the back-propagation neural network model are optimized by using the leave-one-out cross-validation technique, and the models of support vector regression and random forest are selected carefully. The point and interval predictions of the low-cycle-fatigue life of Q235B are reported to reveal the feasibility and advantage of the proposed models. The results help to identify how to understand the fatigue behavior of materials by combining machine learning models and stress-strain hysteresis loops.