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

Area of Science:

• Materials Science
• Mechanical Engineering
• Computational Science

Background:

• Metal fatigue is a critical failure mechanism in engineering components.
• Predicting fatigue life under uncertainty remains a significant challenge.
• Stress-strain hysteresis loops contain valuable information about material behavior during fatigue.

Purpose of the Study:

• To develop machine learning models for predicting metal fatigue life under uncertainty.
• To extract and utilize stress-strain data from hysteresis loops for fatigue analysis.
• To quantify uncertainty in fatigue life predictions.

Main Methods:

• Analysis of stress-strain hysteresis loops from Q235B under strain-controlled loading.
• Extraction of key stress-strain data points and transformation into polar coordinates.
• Quantification of uncertainty by extending parameters to intervals and generating random data.
• Construction and optimization of three machine learning models: back-propagation neural network, support vector regression, and random forest.

Main Results:

• The developed machine learning models provide point and interval predictions for low-cycle fatigue life.
• The models demonstrate feasibility and advantages in predicting fatigue life under uncertainty.
• Leave-one-out cross-validation was used to optimize the back-propagation neural network parameters.

Conclusions:

• Combining machine learning models with stress-strain hysteresis loop analysis offers a powerful approach to understanding material fatigue behavior.
• The proposed method effectively addresses the challenge of small fatigue test datasets.
• This study provides insights into predicting metal fatigue life with quantified uncertainty.