Development of a Predictive Wear Model Based on Test Data Using Machine Learning Algorithms

This study explores the application of machine learning algorithms to predict abrasive wear based on the chemical composition and hardness of materials, validated through tests using ASTM G-65 standards. Three models— Artificial Neural Networks (ANN), Extreme Learning Machines (ELM), and k-Nearest Neighbors (KNN)—were evaluated for their accuracy and computational efficiency. ANN emerged as the most accurate model, achieving 85.75% accuracy with an absolute mean error of 0.228 g and a root mean square error of 0.294 g, demonstrating effective learning and error reduction. ELM showed comparable performance with less parameter tuning but required significant computational resources at higher neuron counts. In contrast, KNN, while computationally simpler, was unable to model complex relationships, converging to average wear values. The findings validate the potential of machine learning in the prediction of wear for component recovery industries. Future research should expand datasets and explore diverse wear conditions to enhance model reliability and industrial applicability.