A Flexible Methodology for Uncertainty-Quantified Monitoring of Abrasive Wear in Heavy Machinery Using Neural Networks and Phenomenology-Based Feature Engineering

Abstract

This paper introduces a cutting-edge methodology for the monitoring of abrasive wear, particularly focusing on SAG (Semi-Autogenous Grinding) mills liners. The lack of a regular inspection regime has historically led to opportunistic and thus, irregular wear measurements that are challenging to integrate into machine learning algorithms for condition-based maintenance. The study unveils a virtual sensor designed to estimate the mill liner's remaining thickness, aiming to offer daily updates and assist the maintenance team in determining the optimal timing for liner replacements without the need for halting operations. This approach is positioned as a strategic response to the critical need for efficient maintenance strategies, addressing the inherent challenges in real-world industrial settings where data quality may be poor and operational realities dominate. A significant aspect of this methodology is its emphasis on uncertainty quantification, vital for informed maintenance decision-making. This novel approach has been successfully applied to SAG mills at Minera Los Pelambres, showcasing its potential for broader applications across scenarios characterized by sporadic data collection. The results showcase an error of ±7.4254 mm of remaining thickness on the validation set, demonstrating the effectiveness of the methodology. The key contributions of this work lie in its ability to utilize low-quality data effectively and its low complexity, reducing barriers to implementing predictive health monitoring (PHM) algorithms. The successful implementation highlights the methodology's adaptability and flexibility, marking a significant advancement in the domain of maintenance strategy for the mining industry.