Applying Online Ore Hardness Estimation to Sag Operation and Optimisation

"Comminution represents a significant part of mineral processing plant operating costs. Yet, autogenous (AG) or semi-autogenous grinding mills (SAG) are difficult to control because of disturbances like feed size distribution and rock hardness. AMIRA, within project GeMIII, developed a comminution index from laboratory crusher data. The index correlated with the Bond ball mill work index and with the A×b parameter. The present work extends the use of crusher data as a grindability indicator to real-time industrial applications. The aim is to improve comminution operation using online hardness estimation. Because the operational work index is proportional to the ore work index, rock hardness variations can be estimated from the feed and product size distributions, the throughput, and the power draw of a crusher. Real-time estimation of rock hardness can help understand the dynamic behavior of AG and SAG mills. Multivariate time series were analyzed to understand the impact of hardness on a SAG operation. Results showed that the estimated hardness influences the SAG operation and its throughput. Moreover, non-linear dimension reduction was used to ease the process visualisation and search for potential control strategies. In the future, online hardness measurements should also impact the understanding of comminution circuits and improve their operation.INTRODUCTIONThe variability in hardness, and the feed size distribution, are two of the main disturbances affecting comminution circuits (Sbarbaro & del Vilar, 2010). Size distribution can be measured by imagebased granulometry systems such as WipFrag (Maerz, Palangio, & Franklin, 1996), or by 3D vision systems when the fines are difficult to discriminate by 2D segmentation (Faucher, Makni, Gagnon, Lavoie, & Roberge, 2015).There are few examples of online hardness measurements. One example was the work of Casali et al. (2001) who developed a hardness indicator. The ore grindability soft sensor used the fraction of each lithological class in the ore, as determined by image analysis, to predict the work index of the material. The model was developed using a moving window with a width of 12 months and trends were shown over many years with one data point per month."