Scholarlink

<p>This research presents a data-driven approach for predicting and reducing particle size distribution (PSD) variability at the feed of a semi-autogenous grinding (SAG) mill in a copper mining operation. Using two years of high-frequency industrial data from the stockpile and feeder system, a hybrid modeling framework was developed that integrates unsupervised clustering (K-Means) with cluster-specific Random Forest regression to capture regime-dependent PSD behaviors. The models predict PSD outcomes using feeder speed setpoints, stockpile levels, and throughput, and can be deployed in real-time applications for operational decision support. In addition to achieving high predictive accuracy (average R<sup>2</sup> &gt; 0.90), the framework incorporates a univariate sensitivity analysis to quantify the influence of feeder setpoints on PSD targets and a variability assessment demonstrating reduced dispersion when model-guided inputs are applied. Because the workflow relies solely on operational data rather than mineralogical characterization, it can be adapted to other sites with minor adjustments. These results highlight the potential of machine learning to support future control strategies aimed at stabilizing SAG mill feed characteristics.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Cluster-Based Machine Learning Modeling for Particle Size Variability in SAG Mill Feed

  • 期刊论文

  • 发表日期：

    2026年4月22日

查看全文
  • Matias Saavedra,
  • Nathalie Risso,
  • Moe Momayez,
  • Ricardo Nunes,
  • Victor Tenorio

摘要

This research presents a data-driven approach for predicting and reducing particle size distribution (PSD) variability at the feed of a semi-autogenous grinding (SAG) mill in a copper mining operation. Using two years of high-frequency industrial data from the stockpile and feeder system, a hybrid modeling framework was developed that integrates unsupervised clustering (K-Means) with cluster-specific Random Forest regression to capture regime-dependent PSD behaviors. The models predict PSD outcomes using feeder speed setpoints, stockpile levels, and throughput, and can be deployed in real-time applications for operational decision support. In addition to achieving high predictive accuracy (average R2 > 0.90), the framework incorporates a univariate sensitivity analysis to quantify the influence of feeder setpoints on PSD targets and a variability assessment demonstrating reduced dispersion when model-guided inputs are applied. Because the workflow relies solely on operational data rather than mineralogical characterization, it can be adapted to other sites with minor adjustments. These results highlight the potential of machine learning to support future control strategies aimed at stabilizing SAG mill feed characteristics.