Combining experiments and explainable AI to identify key parameters influencing brake lining tribological performance

Given the numerous parameters influencing the tribological behavior of brake linings, including material properties and braking conditions, this study aims to identify the parameters with the most significant impact on this behavior. This information will allow for the selection of representative test conditions and a better understanding of the resulting tribological responses. Three gradient boosting models (CatBoost, LightGBM, and XGBoost) were used to predict the responses in terms of friction and wear. The SHAP algorithm was then applied to interpret the models and determine the parameters most influential on the coefficient of friction and wear. An experimental study was conducted on an industrial brake lining to validate the machine learning predictions. Among the models, CatBoost achieved the best prediction accuracy compared to experimental measurements, as confirmed by regression metrics. The SHAP analysis revealed that sliding speed is the predominant factor influencing wear, while material properties primarily determine the coefficient of friction. Experimental results confirm these observations, showing that an increase in sliding speed has a limited effect on the coefficient of friction and a significant effect on wear.