Predicting wear damage in moving mechanical contacts: comparative analysis of regression algorithms and feature selection techniques

Abstract

Accurate wear prediction is essential for industries such as manufacturing, transportation, and power generation, as it helps reduce operational risks, minimise downtime, and extend the lifespan of critical components. This study presents a machine learning-based predictive model for estimating wear volume in pin-on-disc systems. The methodology comprises four key stages: feature selection, sample size determination, regression model selection, and model evaluation. The experimental data include parameters such as friction coefficient, tangential force, penetration depth, sliding distance, sound pressure, and load. Feature selection is employed to identify the most relevant parameters for wear prediction, utilising two methods —wrapping and embedding —to refine the feature subset and enhance accuracy. To optimise model performance, the sample size is determined to balance underfitting and overfitting. Initially, linear regression is applied, followed by adjustments to the sample size. Where necessary, more complex algorithms, such as support vector machines (SVMs) and random forests (RFs), are explored to enhance accuracy. Model evaluation employs metrics including mean absolute error (MAE), mean bias error (MBE), root mean square error (RMSE), and the coefficient of determination (R²) to assess predictive performance. This research offers a systematic approach to wear volume estimation and presents a comparative analysis of regression algorithms, providing valuable insights for researchers and practitioners in wear prediction applications.