Development of a machine learning model for predicting liner wear in SAG mills

Abstract

This thesis investigates liner wear in Semi-Autogenous Grinding (SAG) mills and develops machine learning models for predicting wear patterns. The study leverages data from ten industrial SAG mills of varying sizes and configurations to create both individual and generic predictive models. The research begins by analyzing the current limitations in liner wear prediction methods and the economic implications of wear management. Traditional approaches based on direct tonnage measurements or simple regression models often fail to capture the complex relationships between operational parameters and wear patterns. Additionally, while Discrete Element Method (DEM) simulations provide valuable insights, they require significant computational resources and site-specific calibration. A comprehensive methodology was developed that combines operational data analysis with machine learning techniques. The study utilized hourly plant information data including throughput, mill speed, feed percent solids, and power draw, alongside periodic liner measurements from mill inspections. Various machine learning models were evaluated, including Multiple Linear Regression (MLR), Decision Trees, Random Forests, XGBoost, and Multilayer Perceptron (MLP). To understand the theoretical foundations of liner wear, a simulation tool was developed based on Powell's First Principle Model and Morrell's C-Model. This tool enables visualization of charge motion and ball trajectory changes as liners wear, providing insights into the mechanical aspects of the wear process. The results demonstrate that XGBoost outperformed other methods in developing a generic prediction model, achieving Mean Absolute Percentage Error (MAPE) values as low as 3.86% for lifter height and 6.68% for plate thickness predictions under optimal conditions. The model's performance varied based on mill specifications, performing better when predicting wear patterns for mills similar to those in the training dataset. SHAP (SHapley Additive exPlanations) analysis revealed that operational conditions predominantly influence wear predictions, while design parameters showed more complex, interrelated effects. This finding suggests the need for more comprehensive data collection, particularly regarding mill filling metrics and material properties. This research contributes to the field by demonstrating the feasibility of developing generic machine learning models for SAG mill liner wear prediction, while also highlighting the current limitations and areas for future improvement. The developed models and insights can help optimize liner maintenance schedules and mill performance across various operational contexts.