Predicting the Lifespan of Twisted String Actuators Using Empirical and Hybrid Machine Learning Approaches.

Area of Science:

• Robotics and Mechanical Engineering
• Materials Science and Engineering
• Computational Science

Background:

• Predicting the fatigue lifespan of Twisted String Actuators (TSAs) is vital for the reliability of robotic systems.
• Traditional empirical methods often fail to capture complex nonlinearities and stochastic factors in fatigue behavior.

Purpose of the Study:

• To compare the effectiveness of four machine learning models (Linear Regression, Random Forest, XGBoost, GPR) for predicting TSA fatigue lifespan.
• To develop and validate a hybrid physics-guided model integrating empirical equations with machine learning for enhanced life prediction.

Main Methods:

• Trained and validated four distinct machine learning models using experimental data from 144 TSA actuation tests.
• Developed a hybrid model combining an empirical fatigue life equation with an XGBoost residual-correction approach.
• Evaluated model performance using R-squared, RMSE, MAE, and cross-validation consistency.

Main Results:

• The hybrid physics-guided model achieved superior performance with R² = 0.9856, RMSE = 5299.47, and MAE = 3329.67 cycles.
• Achieved high cross-validation consistency (CV R² = 0.9752), demonstrating robustness.
• Outperformed standalone machine learning models in prediction accuracy and generalization.

Conclusions:

• Physics-informed machine learning offers superior interpretability and generalization, especially with limited experimental data.
• Hybrid empirical-ML models show significant potential for accurate component life prediction in variable conditions.
• This approach enhances reliability predictions for robotic actuation systems with flexible transmissions.